Over and Evans

This Element is on new developments in the psychology of

reasoning that raise or address philosophical questions. In
traditional studies in the psychology of reasoning, the focus
was on inference from arbitrary assumptions and not at all from
beliefs, and classical binary logic was presupposed as the only
standard for human reasoning. But recently a new Bayesian Philosophy
paradigm has emerged in the discipline. This views ordinary
human reasoning as mostly inferring probabilistic conclusions of Mind
from degrees of beliefs, or from hypothetical premises relevant
to a purpose at hand, and as often about revising or updating
degrees of belief. This Element also covers new formulations
of dual-process theories of the mind, stating that there are two

Human Reasoning
Human
types of mental processing, one rapid and intuitive and shared
with other animals, and the other slow and reflective and more
characteristic of human beings. Finally, we discuss implications

Reasoning
for human rationality.

About the Series Series Editor

This series provides concise, authoritative Keith Frankish
introductions to contemporary work The University
in philosophy of mind, written by of Sheffield
leading researchers and including both
established and emerging topics. It David E. Over and
provides an entry point to the primary
Jonathan St B. T. Evans

https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

literature and will be the standard
resource for researchers, students, and
anyone wanting a firm grounding in this
fascinating field.

Cover image: Baac3nes / Moment /

Getty Images ISSN 2633-9080 (online)
ISSN 2633-9072 (print)
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press
 Elements in Philosophy of Mind
edited by
Keith Frankish
The University of Sheffield

HUMAN REASONING

David E. Over
University of Durham
Jonathan St B. T. Evans
University of Plymouth
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press
 Shaftesbury Road, Cambridge CB2 8EA, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre,
New Delhi – 110025, India
103 Penang Road, #05–06/07, Visioncrest Commercial, Singapore 238467

Cambridge University Press is part of Cambridge University Press & Assessment,

a department of the University of Cambridge.
We share the University’s mission to contribute to society through the pursuit of
education, learning and research at the highest international levels of excellence.

www.cambridge.org
Information on this title: www.cambridge.org/9781009495363
DOI: 10.1017/9781009495349
© David E. Over and Jonathan St B. T. Evans 2024
This publication is in copyright. Subject to statutory exception and to the provisions
of relevant collective licensing agreements, no reproduction of any part may take
place without the written permission of Cambridge University Press & Assessment.
When citing this work, please include a reference to the DOI 10.1017/9781009495349
First published 2024
A catalogue record for this publication is available from the British Library.
ISBN 978-1-009-49536-3 Hardback
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

ISBN 978-1-009-49531-8 Paperback

ISSN 2633-9080 (online)
ISSN 2633-9072 (print)
Cambridge University Press & Assessment has no responsibility for the persistence
or accuracy of URLs for external or third-party internet websites referred to in this
publication and does not guarantee that any content on such websites is, or will
remain, accurate or appropriate.
 Human Reasoning

Elements in Philosophy of Mind

DOI: 10.1017/9781009495349
First published online: May 2024

David E. Over
University of Durham
Jonathan St B. T. Evans
University of Plymouth
Author for correspondence: David E. Over, [email protected]

Abstract: This Element is on new developments in the psychology of

reasoning that raise or address philosophical questions. In traditional
studies in the psychology of reasoning, the focus was on inference from
arbitrary assumptions and not at all from beliefs, and classical binary
logic was presupposed as the only standard for human reasoning. But
recently a new Bayesian paradigm has emerged in the discipline. This
views ordinary human reasoning as mostly inferring probabilistic
conclusions from degrees of beliefs, or from hypothetical premises
relevant to a purpose at hand, and as often about revising or updating
degrees of belief. This Element also covers new formulations of
dual-process theories of the mind, stating that there are two types of
mental processing, one rapid and intuitive and shared with other animals,
and the other slow and reflective and more characteristic of human
beings. Finally, we discuss implications for human rationality.

Keywords: new paradigm, dual processes, Bayesian reasoning, probability

https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

judgement, human rationality.

© David E. Over and Jonathan St B. T. Evans 2024

ISBNs: 9781009495363 (HB), 9781009495318 (PB), 9781009495349 (OC)
ISSNs: 2633-9080 (online), 2633-9072 (print)
 Contents

1 Introduction 1

2 Deductive Reasoning 2

3 Reasoning and Probability 13

4 Non-deductive Reasoning 23

5 Conditionals 31

6 Dual Processes in Reasoning 41

7 Rationality and Reasoning 52

8 Overall Conclusions 60

References 62
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press
 Human Reasoning 1

1 Introduction
Reasoning is the inferring of conclusions from premises, and this Element is
an account of what psychologists have discovered about human reasoning.
We have not found it hard to select topics for it. Two broad contemporary
themes in the psychology of reasoning have been influenced by logicians and
philosophers and clearly raise philosophical questions.
The first of these themes is the emergence of a new paradigm in the
psychology of reasoning, replacing with Bayesian subjective probability
theory the traditional presupposition that classical logic and its conditional
set the normative standard for human reasoning. This theme is primarily found
in Sections 2–5. Section 5 has at its core the psychological findings on the
hypothesis that the probability of the natural language conditional, P(if p then q),
is the conditional probability of q given p, P(q|p). This hypothesis was first stated
by philosophical logicians and has long been of great interest to them, and
its confirmation in psychological experiments, for many conditionals, in turn
supports the new Bayesian paradigm. More generally, logicians and philosophers,
going back at least to de Finetti and Ramsey, have substantially influenced the
new paradigm, as will become clear in our references.
The second theme is research in psychology on dual-process theories of the
mind. These state that there are two types of mental processes, one rapid and
intuitive and shared with other animals, and the other slow and reflective and
more characteristic of human beings. Section 6 is on this topic, explaining how
the psychological understanding of it has been deepened and refined over time.
Philosophers have helped in this process, and no doubt will continue to do so.
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

Here, as elsewhere, we cite relevant philosophical as well as psychological work.

Section 7 concerns what the new paradigm and new forms of dual-process
theory tell us about human rationality, epistemic and instrumental. This section
rapidly leads to normative and philosophical questions that have to be left in
philosophical hands.
The psychology of reasoning, as we define it, is the science in which
hypotheses and theories about human reasoning are proposed and tested in
experiments. Human participants are asked to solve reasoning problems under
controlled conditions, with features of the problems manipulated to address
theoretical questions (Evans, 2005). Most of these experiments have had literate
participants, but there have also been studies of participants who do not have or
use a written language (Boissin et al., 2024). The answers people give, and
sometimes the speed with which they do so, inform us about the nature of the
mental processes they use. This field has always paid attention to the writing of
philosophers about how they believe people should reason. The results have
 2 Philosophy of Mind

informed both philosophy and psychology alike. In this Element, we summarise

the essential psychological findings and their implications for philosophers.

2 Deductive Reasoning
The importance of classical logic (Shapiro & Kissel, 2022) is widely recog-
nised. It is binary in having truth and falsity as its only truth values, and the
conclusions of its valid inferences are true assuming that the premises are true.
Psychologists working on human thinking and reasoning in the mid twentieth
century presupposed that classical validity was the only logical standard for
human rationality, and this had a profound influence on the psychology of
reasoning. It led to hundreds of published experiments using the deduction
paradigm (Evans, 2002) based on classical logic. To contrast it with later
developments, we will sometimes call this the traditional approach or para-
digm. This field of work was directly or indirectly inspired by philosophical
writing, and it is of importance that philosophers be aware of the findings that
were reported. A detailed review of this period of work is in Evans et al. (1993)
and a brief summary follows in this section. This traditional approach domin-
ated the field until around 1990, when a number of authors started to call for
a different approach, now known as the ‘new paradigm psychology of reason-
ing’ (Oaksford & Chater, 2020; Over, 2009, 2020). This new approach made
degrees of belief and so subjective probability judgements central to the study of
human reasoning. The relevant normative standards for it are probability theory
and probability logic (Demey et al., 2023).
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

2.1 The Deduction Paradigm

One huge influence on the psychology of reasoning was the work of the Swiss
psychologist Jean Piaget who argued that children’s thinking develops through
a series of stages which conclude with that of formal operational thought. This
meant that adults were supposed to be able to reason abstractly and hypothetic-
ally (Inhelder & Piaget, 1958). Error and bias had been reported in very early
psychological studies of reasoning (Wilkins, 1928; Woodworth & Sells, 1935),
but some psychologists argued that these studies were misleading. People’s
underlying reasoning could be logical but obscured by a failure to apply
reasoning to the problem set, or by misinterpreting the given information, so
that they were in effect reasoning from different premises than those assumed
by the experimenter (Henle, 1962; Smedslund, 1970). However, the idea that
human reasoning was basically logical was also subjected to strong empirical
challenge, for example in the early studies of the British psychologist Peter
Wason and his colleagues (Wason & Johnson-Laird, 1972).
 Human Reasoning 3

In the traditional paradigm, experimenters asked people to evaluate or gener-

ate logical arguments, with their answers compared with classical logical norms
to assess their correctness. Psychologists took the view that people ought to be
innately logical and that they should therefore not have had formal logical
instruction. Indeed, it is standard practice with this method to exclude people
who have received any formal training in logic. In more detail, the method
consisted of asking participants in an experiment to assume the truth of given
premises, which were sometimes abstract, arbitrary, or unbelievable. They were
then asked whether a given conclusion necessarily followed, or to infer
a conclusion which necessarily followed. The term ‘necessarily’ in these
instructions was rarely defined for the participants (but see Lassiter &
Goodman, 2015, on how they might understand it).
The other key component in the deduction paradigm was the use of
classical logic as the normative theory for deciding whether the deductions
made are correct or incorrect. This is noteworthy because outside of the study
of reasoning and decision-making, normative theory is rarely used in cogni-
tive psychology. Only in the study of reasoning and decision-making do
psychologists and philosophers debate rationality. For example, no one is
described as irrational for failing to remember a long list of words or being
unable to read tiny print, but people have been accused of irrationality for
reasoning illogically, or for violating the axioms of decision theory. It is also
striking that the majority (but by no means all) of the psychologists working
on reasoning have had until recently only an elementary understanding of
logic, confined mostly to classical logic. The problem of alternative norma-
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

tive accounts (see Section 7) was not addressed by psychologists until most of
the deduction studies had been published (but see Elqayam & Evans, 2011,
Oaksford & Chater, 2007, and Stanovich, 1999). The great bulk of studies in
the deduction paradigm have used Aristotelian syllogisms or conditional
inferences, and we will briefly summarise this work and its main findings in
Sections 2.2–2.6.

2.2 Classical Syllogisms

Psychologists have extensively used the syllogisms first described by Aristotle
(Kneale & Kneale, 1962) to study deductive reasoning. Each syllogism consists
of two assumed premises and an inferred conclusion. The three terms related in
a syllogism are often described as S and P, representing the subject and predicate
of the conclusions, and M, a middle term which links them, as in

Some M are P, All S are M

Therefore, Some S are P
 4 Philosophy of Mind

A realistic version of the above might be

Some scientists are astronomers

All psychologists are scientists
Therefore, Some psychologists are astronomers

This argument is clearly invalid. It is possible that some psychologists are also
astronomers but there is no necessary reason why they should be, given the
premises. However, participants are much more likely to say this argument is
valid than invalid when it is given in an abstract form, such as Some A are B, All
C are A, therefore Some C are B.
The premises and conclusions may each take one of four different forms,
classically known as A (All A are B), E (No A are B), I (Some A are B) and
O (Some A are not B). The mood of the syllogisms is determined by these forms.
For example, the above argument has the mood IAI. Syllogisms can also have
figures, depending on how the terms are arranged. The above syllogism is in figure
one: M-P, S-M, S-P. Assuming that the conclusion is always of the form S-P, there
are three other ways to arrange the terms, for example, P-M, S-M, S-P is in figure
two. With 64 possible moods and four possible figures, there are 256 distinct
syllogisms. Of these a mere 24 are valid in Aristotelian logic, even allowing weak
conclusions, such as Some S are P when All S are P could have been concluded.
Some S are P does not validly follow from All S are P in contemporary classical
logic (Shapiro & Kissel, 2022), which implies that All S are P is true when S is an
empty term. But this difference will not affect our psychological points here. From
a psychological point of view, we can double the number of syllogisms by
switching the positions of the first and second premises. Though this does not
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

affect the logic, it can and does affect the inferences ordinary people draw.
The psychological studies divide into those using abstract terms (typically the
letters A, B and C) and those using realistic content. Both have provided
interesting findings, and we consider the abstract reasoning first. The standout
finding of the abstract studies is that people endorse the conclusions of most
syllogisms, even though the great bulk of them are invalid. To our knowledge,
only one published study (Evans et al., 1999) presented every possible combin-
ation of moods and figures, with participants asked in one group to judge
whether the conclusion was necessary (equivalent to valid), and in another
whether it was possible, given the premises. (The authors provided an appendix
showing the percentage saying ‘Yes’ in each case.) Asked to judge the necessity
(validity) of conclusions given the premises, participants accepted about
72 per cent that were valid conclusions, around 45 per cent that were invalid
but possible conclusions, but only 10 per cent that were logically impossible
given the premises. So clearly there was some understanding of the logic being
 Human Reasoning 5

shown, but with very high error rates. When asked if the same conclusions were
possible, these numbers increased in all cases, so that ‘possible’ seemed to
participants simply to be a cautious form of ‘necessary’. The effect was greatest
for syllogisms with possible conclusions, but still present for necessary and
impossible ones. Basically, and contrary to both logic and the instructions,
participants like to endorse conclusions unless there is a clear reason not to do
so. The overwhelming finding in this research is the very high endorsement rate
of fallacies. Abstract syllogistic reasoning is clearly very difficult for (typically)
university student participants when not trained in logic.
Both the mood and the figure of the syllogisms also bias the judgement of
validity. It was claimed in the earliest study (Woodworth & Sells, 1935) that
people like to endorse conclusions which match the mood of the premises, the
so-called atmosphere effect. While this finding has been broadly replicated in
later studies (Evans et al., 1993), more nuanced psychological accounts of the
precise nature of the bias have been developed (e.g., Chater & Oaksford, 1999).
The figure of the syllogisms also biases reasoning, in that people generally
prefer conclusions in which the order of the terms corresponds with their order
of presentation in the premises. This has been observed both when people
generate their own conclusions (Johnson-Laird & Bara, 1984) and when con-
clusions are given for evaluation (Morley et al., 2004). In brief, abstract
syllogistic reasoning is very difficult for ordinary people, who make many
logical mistakes, frequently endorse fallacious conclusions as necessary and
are systematically biased by logically irrelevant variants of the presentation (see
Oaksford & Chater, 2020, for new paradigm studies of syllogisms).
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

2.3 The Belief Bias Effect

One of the most important findings in the study of syllogistic reasoning arises
when realistic content is introduced. As Wilkins (1928) observed almost
a century ago, people’s judgements of validity are influenced by whether or
not they believe the conclusion. After a number of early reports based on
questionable methodology, the basic phenomena of belief bias were established
by Evans et al. (1983) followed by a flurry of research interest which persists to
the current day in the psychology of reasoning (see Evans et al., 2022, for
a recent review). Evans et al. (1983) devised syllogisms which fitted into four
categories, depending on the validity of the arguments and the believability of
the conclusions. Hence, conclusions could be Valid-Believable (VB), Valid-
Unbelievable (VU), Invalid-Believable (IB), and Invalid-Unbelievable (IU).
Comparing performance on these four types lead to three clear findings, all
highly statistically significant, which have been replicated many times since:
 6 Philosophy of Mind

1. Validity effect. People endorse more valid than invalid arguments, that is
(VB+VU) > (IU+IB)
2. Belief effect. People endorse more arguments with believable than
unbelievable conclusions, that is (VB+IB) > (VU+IU)
3. Interaction effect. Belief bias is larger for invalid arguments, that is
(IB-IU) > (VB-VU).

All three effects are about equally large and highly reliable. Figure 1 shows the
endorsement rates of the combined experiments of Evans et al., 1983. Most
replication studies show a statistically significant belief bias for valid as well as
invalid arguments, but of smaller size, as Figure 1 clearly illustrates.
To illustrate the findings with some examples, the participants in the Evans
et al. (1983) study found the following argument quite compelling, with
71 per cent declaring it valid:

No addictive things are inexpensive

Some cigarettes are inexpensive
Therefore, some addictive things are not cigarettes

By contrast they were very unimpressed with this argument; only 10 per cent
thought it valid:

No millionaires are hard workers

Some rich people are hard workers
Therefore, some millionaires are not rich people

Endorsment rates (%), Evans et al. 1983

https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

100
90
80
70
60
50
40
30
20
10
0
VB VU IB IU

Figure 1 Percentage of conclusions endorsed as valid in the study of

Evans et al. (1983) for valid-believable (VB), valid-unbelievable (VU),
invalid-believable (IB), invalid-unbelievable (IU) syllogisms
 Human Reasoning 7

The two arguments have precisely the same logical form and neither is valid. The
difference is that people believe the first conclusion and disbelieve the second.
However, the belief bias effect in syllogistic reasoning (demonstrated with
other reasoning tasks) does not mean that participants ignore the logic, when
clearly instructed to assess the necessity of the conclusion given the premises.
Rather it suggests that there is some kind of conflict between the logical task and
influence of belief. Evans et al. (1983) showed that this was not due, as one
might suppose, to some participants using logic and others belief, but rather that
the conflict was within individual reasoners. That is, people would sometimes
go with logic but other times with belief. This was one of the key findings that
lent support to dual-process theories of reasoning discussed in Section 6. The
interaction of belief with validity may relate to the general finding that people
are all too willing to endorse fallacies generally. One reason to withhold an
inference might be perception of its invalidity, but another could be the
unbelievability of its conclusion. Various authors have suggested that people
might be more motivated to disapprove of an argument and look for
a counterexample case when they disbelieve its conclusion (Evans, 2007a).
De Neys (2012) surprisingly found that belief–logic conflict may be detected
rapidly and preconsciously, reflected in response times, confidence levels and
autonomic responses, leading to an argument that there may be ‘logical intu-
itions’ (but see Section 6.3). Moreover, the logic of the problem can interfere
when people are instructed to respond on the basis of belief (Handley et al.,
2011). These findings seem very odd. How could one know whether an argu-
ment was valid or not, without first engaging in a slow process of reasoning?
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

However, such studies have generally used very much simpler logical argu-
ments than classical syllogisms (Section 6.3), which may not engage working
memory (Evans, 2019). What all these studies demonstrate beyond doubt,
however, is that most people find it difficult or impossible to disregard their
prior beliefs completely when engaged in a deductive reasoning task.
Traditional studies of syllogistic reasoning did point the way to the new para-
digm by showing clearly that this reasoning could not be fully explained without
taking account of people’s beliefs. Sections 3–5 will be on the new paradigm topics
of how belief should influence reasoning and how it actually does.

2.4 Conditional Syllogisms

Another major paradigm in the psychology of deductive reasoning, the condi-
tional inference task, presented its participants with conditional syllogisms for
evaluation, again with two assumed premises and an inferred conclusion. As
with classical syllogisms, these problems can be given with abstract or thematic
 8 Philosophy of Mind

content. They are somewhat simpler, however, since they relate only two rather
than three terms. Participants are usually tasked with the following four argu-
ment types to evaluate:

Modus Ponens (MP)

If the letter is A then the number is 4; the letter is A, therefore the number is 4.
Denial of the Antecedent (DA)
If the letter is A then the number is 4; the letter is not-A, therefore the number is
not-4.
Affirmation of the Consequent (AC)
If the letter is A then the number is 4; the number is 4, therefore the letter is A.
Modus Tollens (MT)
If the letter is A then the number is 4; the number is not-4, therefore the letter is
not-A.

Of these inferences, MP and MT are valid, and DA and AC are invalid, in both
classical logic and probability logic (Evans & Over, 2004). While frequencies of
endorsement vary somewhat between different studies in the literature, there are
some very clear trends when materials are abstract, as in the above examples.
MP is endorsed very highly, with a number of studies reporting 100 per cent
acceptance rates. MT, the other valid inference, is endorsed far less frequently,
about 60 per cent of the time. The two fallacies, AC and DA, are frequently
endorsed but the extent varies considerably across studies. The latter finding
could be accounted for by assuming that people adopt a biconditional interpret-
ation or accept invited pragmatic inferences (Evans & Over, 2004). This cannot
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

explain the difficulty with MT inferences, which is valid on all plausible

pragmatic interpretations, but see Section 3.4 on these inferences.
Generally speaking, individual differences in general intelligence correlate
highly with finding the normatively correct solution to reasoning and decision
problems (Stanovich, 2011). In line with this, there is evidence that participants of
higher general intelligence are better able to suppress the fallacies AC and DA in
abstract conditional inference (Evans, Handley et al., 2007; Newstead et al., 2004).
However, the same studies show, surprisingly, that high-ability participants draw
fewer valid MT inferences. One possible explanation is that MT is hard for people to
draw when a genuine attempt at deductive reasoning is being made (but again see
Section 3.4 on MT). Lower-ability participants might also engage in simple equiva-
lence reasoning. Given if p then q, they may expect p and q to go together, although
this cannot explain the full pattern of findings (Evans et al., 2007).
The difficulty of MT became a focus for an argument between theorists
favouring a mental models account of deductive reasoning (Johnson-Laird,
1983; Johnson-Laird & Byrne, 1991) and those offering a mental logic account
 Human Reasoning 9

(Braine & O’Brien, 1991; Rips, 1994). Mental logic theorists argued that people
use an innate ‘natural logic’ comprised of natural deduction inference rules,
some of which take a simple, direct form. An example of such a rule is and-
elimination, inferring p from p & q (see also Section 3.2). Such inferences are
said to be made automatically and without effort, drawing on implicit rules that
are effectively part of the language processing system. Other inferences can
only be made by high effort, and presumably conscious, indirect reasoning
strategies. For example, as in classical natural deduction theories proposed by
logicians (Shapiro & Kissel, 2022), there is no direct rule for MT in mental
logic. It can only be derived indirectly using reductio ad absurdum reasoning. In
the above example, one would suppose that there was an A, and if there were an
A then there would be a 4. Since there is not a 4, the supposition of that there is
an A must be false by a reductio. However, this indirect reasoning is hypothe-
sised to be slow and error prone compared with the direct MP. The latter follows
immediately from an inference rule built into the natural language meaning of if
in the mental logic account.
Mental model theory, by contrast, denies that there are any inference rules in
the head, and proposes instead that deduction is based on a simple semantic
principle: an inference is valid if there is no counterexample to it. In this account,
people construct and use mental models in reasoning to represent possible states
of the world and search for counterexample models to try to infer or test the
validity of inferences. Presumably this is at least in part a conscious process
because core to the theory and its predictions is the claim that mental models load
working memory, increasingly so when they are more complex or more than one
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

needs to be considered. MP is simple and immediate because the p & q case is the
default mental model that is used to represent if p then q, along with initially
unspecified other possibilities. When presented with the premise not-q, no infer-
ence will follow unless the individual tries to ‘flesh out’ other states of the world
(models) that could be consistent with the conditional statement. These additional
models are: not-p & q and not p & not q, with only p & not-q being excluded. On
this basis, only the second of these models is consistent with a conditional
statement in which not-q also holds, and it yields the not-p conclusion. But
again, this reasoning process is supposed to be indirect, difficult, and prone to
error (see Section 3.1 for more on the conditional in mental model theory).
Conditional inferences have also been studied with a variety of realistic
content, which can affect the rates with which the conclusions are inferred in
the four inferences (Byrne, 1989; Stevenson & Over, 1995). Of particular import-
ance are the real-world beliefs that participants hold about the relationship
between p and q, for a conditional statement if p then q. Studies have manipulated
whether the occurrence of p appears to be either a sufficient or necessary
 10 Philosophy of Mind

condition for q with particular contents. It is not perceived as necessary when

people can easily think of alternative antecedents which would lead to q, and it is
not perceived as sufficient when they can think of disabling conditions which
would prevent p leading to q. The basic finding is that the valid inferences MP and
MT, are suppressed when perceived sufficiency is low, implying that P(q|p) is
low, and the fallacies, AC and DA, are suppressed when perceived necessity is
low, implying P(p|q) is low. Classical logic only guarantees that MP and MT lead
to true conclusions from true premises. If the premises are in some doubt, say
because of low sufficiency, the conclusion can rightly be doubted as well.
However, the suppression findings are retained when the participants are told to
assume that the premises are true (Thompson, 1994). Under these instructions,
some participants violate classical logic. They may find it hard to assume what
they do not believe, and finding the premises uncertain, they may take the
conclusion to be uncertain as well (see Sections 3.2 and 5.2). It could therefore
be argued that confidence in the conclusion is what is sometimes ‘suppressed’ and
not the inference itself (Over & Cruz, 2018).
A later study demonstrated belief bias more directly in conditional reasoning by
showing that all four inferences, both valid and invalid, could be suppressed when
participants had low belief in the conditional statement (Evans et al., 2010). This
effect, however, depended both upon the instructions given and the cognitive
ability of the participants. One group were given deductive reasoning instructions,
being told to assume the premises were true and asked to decide if the conclusion
necessarily followed. Another group were given pragmatic reasoning instructions
that made no reference to assuming the premises and were asked to judge
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

their degree of belief in the conclusion. All participants were given a test of
general intelligence and split into subgroups high and low on this measure.
Under pragmatic reasoning instructions, participants of all abilities showed
a similar belief effect, in that they assigned lower ratings to conclusions
drawn from unbelievable conditional statements. Under deductive reasoning
instructions, however, the belief bias effect was restricted to lower-ability
participants only. This finding is important as few studies of deductive rea-
soning have included intelligence measures. It indicates that people of higher
cognitive ability are better able to comply with instructions to assume the truth
of the premises. Participants of lower cognitive ability are less able to do this
and tend to display belief bias.
In a new paradigm approach (Section 2.6), we would ask how far being less
able to follow instructions to assume premises is a serious bias that limits
rationality (Section 7). But these findings are consistent with the broader
individual differences research programme of Keith Stanovich, who argues
that higher intelligence facilitates the ability to ‘decouple’ beliefs and reason
 Human Reasoning 11

hypothetically (Stanovich, 2011). We will discuss his programme of work in

detail in Section 6.

2.5 Summary of Findings with the Deduction Paradigm

Had space permitted, we might also have discussed here the extensive research
carried out with the Wason selection task (Wason, 1966). The selection task has
traditionally been seen as a deductive reasoning problem and studied by the same
researchers as well as others not normally working in the psychology of reasoning.
This has produced a rich set of findings (see Evans, 2022, for a recent review). The
task itself will be covered in Section 6.1, but suffice it to say for now that work on
the selection task supports the same general conclusions which we can draw from
the deduction studies briefly reviewed in this section. These are as follows:

1. Untrained participants show some degree of deductive competence: their

responses are generally influenced, at least in part, by the logical validity of
the arguments presented to them.
2. There is a strong general tendency to endorse invalid arguments whose
conclusions could be true but are not necessitated by the premises.
3. Participants are subject to a number of cognitive biases in reasoning, making
systematic as well as random errors.
4. When realistic content is used, people are almost always influenced by their
prior knowledge and belief about the content and context.
5. People higher in general intelligence generally are better able to reason
logically and ignore prior belief, but only if they are given strict deductive
reasoning instructions.
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

Taken as a whole, the above findings suggest that either much real-world reasoning
is irrational, or else that the yardstick for rationality should not simply be classically
valid inferences from assumptions. Of course, being unable to suspend one’s beliefs
when instructed to do so by a psychologist in an experiment cannot, in itself, be
irrational. As a result, this experimental work, along with that in other fields of
reasoning and decision-making, has initiated a great rationality debate (see Evans,
2021, and Section 7). Dissatisfaction with the traditional approach and its associated
deductive paradigm has also led to the new paradigm in the psychology of reasoning,
which we continue to describe in the next and later sections.

2.6 New Paradigm Psychology of Reasoning

What is often referred to as the ‘new paradigm’ is not very precisely defined
other than negatively. It is the psychological study of human reasoning that has
moved away from the definition of normativity based on classical binary logic,
 12 Philosophy of Mind

in which all propositions are assumed to be true or false, with no account taken
of their subjective probabilities. The original use of ‘new paradigm psychology
of reasoning’ was to refer to studies that placed subjective probability judge-
ments at the centre of the field (Over, 2009). But for some psychologists, the
problem is not just a demand to reason from assumptions in classical logic, but
the use of a strong normative theory to assess human reasoning, obscuring the
psychological study. It has been argued that the psychology of reasoning should
fall in line with the rest of cognitive psychology by focussing on what people do,
rather than recording errors and biases (Elqayam & Evans, 2011; Evans, 2002).
This change encourages a much deeper focus on pragmatics in reasoning
(Bonnefon, 2013).
However, there is also a strong tradition and contemporary practice of
proposing alternative normative models to classical logic for reasoning (Over,
2020). A sustained research programme of this type has been run by the
psychologists Mike Oaksford and Nick Chater, dating from their seminal
paper on the Wason selection task, in which they argued that the standard
‘erroneous’ responses could have a rational explanation from a different view-
point (Oaksford & Chater, 1994). Oaksford and Chater (2007) argued influen-
tially that Bayesian subjective probability theory could replace classical logic as
the norm for human reasoning. Philosophers contributed to the paradigm shift.
L. Jonathan Cohen’s (1981) critique of the psychology of reasoning and deci-
sion-making, which will be covered in Section 7, directly challenged psycholo-
gists’ use of standard normative theories. Also very influential was Edgington
(1995) on the natural language conditional if p then q (the first author learned
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

much from having her as his PhD supervisor). She explained the problems of
claiming that if p then q is logically equivalent to the classical material
conditional, itself logically equivalent to not-p or q. This claim was implicitly
assumed to be correct in much psychological work at that time (but see Braine &
O’Brien, 1991, and Rips, 1994, for dissenting voices).
Edgington (1995) argued strongly (see originally Adams, 1965) for what we
will call the Equation and, in experimental contexts, the conditional probability
hypothesis. The Equation/conditional probability hypothesis states that the
subjective probability of the natural language conditional P(if p then q) is the
conditional subjective probability of q given p:

P(if p then q) = P(q|p)

In the early twenty-first century, psychologists demonstrated that ordinary

people generally conform to this hypothesis. For the first experimental studies
of it, see Evans et al. (2003), Oberauer and Wilhelm (2003), and Over et al.
(2007). Section 5 covers later studies.
 Human Reasoning 13

A book published by the current authors (Evans & Over, 2004) drew together
philosophical and psychological work on conditionals and proposed
a suppositional theory of conditionals (see Section 5.1), linked to broader theory
of hypothetical thinking, in which P(if p then q) = P(q|p). Following this work,
a major field of study of conditionals within the new paradigm has developed,
with a range of logical ideas being taken from philosophy and applied in
a psychological context (see Sections 3, 5, and 6). The traditional deduction
paradigm has not been entirely abandoned, however. For example, it is still
relevant to the study of dual processes in reasoning (Section 6), as it requires
what is known as Type 2 processing – high effort, loading on working memory –
to comply with the instructions to assume the premises of inferences, disregard-
ing belief, and find conclusions that necessarily follow. Only some people with
special training, or relatively high cognitive ability, are able follow such
instructions fully and reliably.

3 Reasoning and Probability

The psychology of reasoning started out presupposing that ‘correct’ reasoning
is simply classically valid inference from premises assumed to be true. In
contrast, the new paradigm, as we interpret it, stresses that most reasoning, in
science and everyday affairs, takes place in a context of uncertainty. It is from
premises which are beliefs or hypotheses with reasonable probabilities, or it is
about possible actions in decision-making (Over, 2020). The difference
between the traditional paradigm and the new paradigm is particularly marked
in theories of the natural language conditional. In this section, we will explain
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

why the new paradigm rejected the traditional presupposition that the natural
language conditional is equivalent to the classical material conditional, and also
cover the logical and philosophical foundations of the new paradigm.

3.1 The Material Conditional in Psychology

The mental model theory of Johnson-Laird and Byrne (1991) is the best
example of the claim in traditional psychology of reasoning that a fully mod-
elled, or analysed, natural language conditional, if p then q, is equivalent to the
material conditional of classical logic, which is, in turn, logically equivalent to
not-p or q. One of their examples was:

(3.1) If Arthur is in Edinburgh (d), then Carol is in Glasgow (g).

Johnson-Laird and Byrne held that (3.1) is true when Arthur is in Edinburgh, and
Carol is in Glasgow, and false when Arthur is in Edinburgh, and Carol is not in
Glasgow. They then asked, ‘But, suppose its antecedent is false, i.e. Arthur is not
 14 Philosophy of Mind

in Edinburgh, is the conditional true or false?’, and answered, ‘It can hardly be
false, and so since the propositional calculus allows only truth or falsity, it must be
true’ (Johnson-Laird and Byrne, 1991, p. 7). The propositional calculus is the
branch of classical logic for reasoning with the connectives not, and, and or, and
so Johnson-Laird and Byrne were clearly presupposing here, without argument,
that the correct logic for the natural language conditional (3.1) is bivalent and
truth functional. They agreed with many other theorists that (3.1) is true when d is
true and g is true, and false when d is true and g is false. But then Johnson-Laird
and Byrne appealed to classical logic to infer that (3.1) is true when d is false and
g is true, and true when d is false and g is false. They are presupposing that (3.1) is
logically equivalent to not-d or g. This material conditional analysis of the natural
language conditional is summarised in Table 1.
Johnson-Laird and Byrne (1991, pp. 73–74) accepted the paradoxes of the
material conditional analysis. The first paradox is that this analysis implies the
logical validity of inferring if p then q from not-p, and the second paradox is that
it implies the logical validity of inferring if p then q from q. They rightly wrote
that we must decide ‘ . . . to abandon this analysis of the conditional . . ., or to
accept the validity of these apparently paradoxical inferences and to explain
why they seem improper’, and they added, ‘We shall embrace the second
alternative’. They went on to argue that these inferences appear paradoxical
because they ‘throw semantic informative away’. In other words, the conclusion
is less informative than the premise.
It is true that in cooperative communication we would not, generally, want to
deny our hearers information by making an assertion that we had inferred from
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

one with more semantic information (compare Grice, 1989, on these paradoxes).
But we do not have to worry about communicating well with other people when
we are making inferences from our own degrees of belief, for our own benefit. For
example, suppose we are wondering whether to invest in a cryptocurrency
scheme that promises a large profit, and consider:

Table 1 The material conditional

analysis of if p then q

pq if p then q
TT T
TF F
FT T
FF T

T = true, and F = false

 Human Reasoning 15

(3.2) If we invest in the scheme (i), then our money will be in honest hands (m).

If (3.2) were a material conditional, equivalent to not-i or m, its probability would

increase as the probability of its antecedent, P(i), decreased: P(not-i or m) goes up
when P(not-i) increases. Assume we become more and more convinced, on solid
grounds, that the scheme is a scam. That will make us less and less likely to invest in
it, and P(not-i) will get higher and higher. However, that will increase P(not-i or m),
the probability of the material conditional, so our belief in the conditional is
apparently strengthened by our reluctance to invest! Much more plausible
psychologically is that we refuse to invest in the scheme because the probability
of (3.2) gets lower and lower, as it will if we follow the Equation of 2.6, making
P(if i then m) = P(m|i). We will return to the Equation in Section 5 and explain how
it is generally supported by psychological experiments.
Problems with the material conditional analysis eventually brought about
a radical revision (beginning with Johnson-Laird et al., 2015) of mental model
theory, in which the paradoxes are no longer claimed to be logically valid (see
Over, 2023a, for critical comments on the revision). Williamson (2020) has
a much more sophisticated defence of the material conditional analysis. It has
also been criticised (Rothschild, 2023; van Rooij et al., 2023), but Williamson
stresses the role of suppositions in assessing conditionals, rightly in our view
given the psychological evidence. We introduced some of this evidence in
Section 2.6 and will come to more of it in Section 5.

3.2 Logically Valid Reasoning and Probability

Before the development of the new paradigm, psychologists did sometimes link
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

valid deductive reasoning and probability judgements. This link is most

strongly found in the famous Tversky and Kahneman (1983) article on the
conjunction fallacy, reporting that people sometimes judge the probability of
a conjunction, P(p & q), to be higher than the probability of one of its conjuncts,
P(p). As background for their experiments, Tversky and Kahneman pointed out
that P(p & q) cannot be greater than P(p) in normative probability theory.
If p & q could be more probable than p, then p could be false when p & q was
true, and that is logically impossible.
More generally and considering only probability assignments that are con-
sistent with probability theory, that have coherence, we can state that a one-
premise inference is logically valid if and only if the probability of its premise
cannot be greater than the probability of its conclusion by probability theory.
For example, the natural deduction inference rule of and-elimination,
p & q therefore p, is valid in probability logic as well as classical logic because
P(p & q) ≤ P(p). Probabilistic validity preserves probability, just as classical
 16 Philosophy of Mind

validity preserves truth. For inferences with more than one premise, we need to
define the uncertainty of a premise p as 1 – P(p) and of a premise if p then q as
1 – P(q|p), and then an inference is probabilistically valid, p-valid, if and only if
the uncertainty of its conclusion cannot exceed the sum of the uncertainties of
its premises (Adams, 1998). More informally, a valid inference cannot increase
our uncertainty: it cannot take us from less uncertainty to more uncertainty. An
example of a two-premise inference is and-introduction, inferring p & q from
p and q as separate premises. This inference is clearly classically valid, since its
conclusion cannot be false when its premises are true, and p-valid, because its
conclusion cannot be more uncertain than its premises. This definition of
probabilistic validity, p-validity, plays a fundamental role in probability logic
(Adams, 1998) and in new paradigm psychology of reasoning (Oaksford &
Charter, 2020; Over, 2020).
In the traditional paradigm, reviewed in Section 2, psychologists usually
asked their participants to assume the truth of the premises in a reasoning
experiment while setting aside any relevant beliefs they might have. These
premises could sometimes be highly abstract and were usually detached from
belief acquisition, practical decision-making, or scientific prediction, and could
be unbelievable when they were not abstract. Such premises are of little use in
everyday or scientific reasoning, where there is little or no point in arbitrarily
assuming what is irrelevant and unbelievable. Participants could find it hard to
make such assumptions and could automatically use what they judged to be
relevant beliefs for their inferences, with the result that they could be charged
with belief bias in a traditional approach. Using a reductio ad absurdum
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

argument, people do assume what they disbelieve, to try to derive an explicit

absurdity from it, but usually in ordinary reasoning, this inference form relies on
background beliefs and has the goal of refuting an opponent with the implicitly
absurd belief. The new paradigm psychology of reasoning is in the truly
psychological business of studying inferences from beliefs and relevant hypo-
thetical suppositions. The aim is to understand and explain people’s inferences,
from their mostly less than certain degrees of belief in the premises, to degrees
of belief in the conclusions (Cruz, 2020; Evans et al., 2015).
For instance, suppose we ask some friends of ours whether they will join us in
the investment referred to in (3.2): ‘If we invest in the scheme (i), then our
money will be in honest hands (m)’. They might reject this possibility out of
hand because P(m|i) is very low for them. This belief could be so strong in them
that they would resist a psychologist’s request to assume that (3.2) is true for the
purpose of discovering whether they endorse MP as a valid inference. But in the
new paradigm, we can test whether people comply with MP by asking for their
judgement about P(m|i). Suppose they judge that P(m|i) = .01, and that they are
 Human Reasoning 17

prepared, for the purpose of relevant decision-making, to assume that P(i) = 1.

Then as long as their judgement about the probability of the MP conclusion,
P(m), is not below .01, they have complied with the p-validity of MP, and there
can be no justification for charging them with a bias.
The new paradigm does not exonerate reasoners from all cognitive biases.
People can still have belief bias (Section 5.2) and other biases. They sometimes
judge P(p & q) > P(p) even when p is explicitly inferred from p & q (Cruz et al.,
2015). Evans et al. (2015) asked participants to assign probabilities to the
premises and conclusions of conditional arguments and found that they con-
formed to p-validity at above chance levels, but that they do so better when
premises and conclusions are grouped together as explicit inferences rather than
judged separately. We get a much fairer assessment of whether people conform
to logical and probabilistic principles, avoiding biases and fallacies, by studying
their degrees of belief in the new paradigm.

3.3 Interpretations of Probability

There are different interpretations of probability theory (Howson & Urbach,
2006). In subjective interpretations, probability judgements are the expres-
sion of subjective degrees of belief. In objective interpretations, probability
judgements describe a feature of the world – frequencies, proportions, or
propensities. The new paradigm in the psychology of reasoning is grounded
in the normative theories of de Finetti (1936/1995, 1937/1964) and Ramsey
(1926/1990. De Finetti had only a subjective interpretation of probability,
but Ramsey argued that both interpretations, subjective and objective, are
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

necessary to cover all the uses of probability theory. In this respect, most new
paradigm psychologists, including ourselves, side with Ramsey and not de
Finetti.
Both authors argued, however, that people’s degrees of belief can be meas-
ured by the bets they are willing to make, and both made use of the notion of
a conditional bet, for example, we might say that, if a certain coin is spun, we bet
it will come up heads. For this bet, suppose we observe that the coin has a worn
edge and decide that we will pay 60 Euro cents to make the bet under the
following conditions. We will win the bet and get 1 Euro when the outcome
turns out to be heads and will lose the bet and get nothing when the outcome
turns out to be tails. If the coin is not spun, the bet, being only conditional, is
‘void’, and we get our 60 cents back. The conditional subjective probability for
us is then .60 that a head will be the result given that the worn coin is spun, and
.60 is the expected value of the bet. Section 5.1 has more on the relation between
conditional bets and the use of natural language conditionals.
 18 Philosophy of Mind

We could use a heuristic, for example, the representativeness heuristic of

Tversky and Kahneman (1983), for making our bet in another example. The
person supplying and spinning the coin may appear to us to be representative,
that is, highly similar to our image, of a confidence trickster. He looks shady to
us and is called ‘Doc’ (our image being formed by films about confidence
tricksters). With this heuristic giving us the belief that the coin is biased,
some normative theorists might charge us with irrationality in this case. But
theorists with a subjective interpretation of probability would not care whether
we acquired our degrees of belief using heuristics. They would only ask if these
degrees conformed to probability theory. De Finetti (1936/1995) said that he
was specifying the logic of probability, and Ramsey’s (1926/1990) phrase was
the logic of partial belief. Hence, to have a subjective interpretation of prob-
ability is not to allow people to have any degrees of belief that they like. The
degrees are constrained by the logic of probability and partial belief, and
consistency with probability theory, coherence, is fundamental to it.
Why should we care about being coherent and conforming to probability
logic? The answer given by both de Finetti and Ramsey is that, if we violate the
principles of probability theory, a Dutch book can be made against us. This
a series of bets which, if we make them, will guarantee an overall loss (Howson
& Urbach, 2006). One could ask whether ordinary people are worried, impli-
citly, about Dutch books being made against them, but it seems highly likely that
people would revise their bets to escape a Dutch book if they became aware of
being caught in one. Gamblers play roulette, for example, when they know the
odds are against them, but playing when it becomes clear that the roulette wheel
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

is rigged, and they will necessarily lose, is quite another matter.

Implicitly using the representative heuristic, we might judge that people are
more likely to be confidence tricksters and called ‘Doc’ than they are to be
called ‘Doc’. Now this judgement does violate probability theory and its logic.
As a result, we have committed Tversky and Kahneman’s conjunction fallacy
and are incoherent, and it is possible to make a Dutch book against us (as Gilio
& Over, 2012, illustrate). The concept of coherence includes classical logical
consistency as a special case. It is inconsistent to imagine that p & q is true and
p is false, and it is incoherent to hold that P(p & q) = 1 and P(p) = 0. More
generally, and extending binary classical logic, it is p-valid to infer p from p &
q and logically incoherent to claim that P(p & q) > P(p).
We become vulnerable to Dutch books if we violate p-validity and conse-
quently are incoherent in our reasoning. But it is impossible for human beings to
certain that they are logically coherent. P(q|p) must be 1 by the logic of
probability when p logically implies q, and the question of whether or not
p logically implies q can be computationally too hard for human beings to
 Human Reasoning 19

answer in elementary logic, and it is not always a logically decidable question in

more advanced logic. Section 7 will focus on rationality and the relationship
that it has to coherence, defined as consistency with probability theory. But in
Section, 3.4, we will introduce a theorem of probability theory that is the
starting point of the Bayesian account of how to change degrees of belief.

3.4 Bayesian Reasoning

People use static reasoning to extend their degrees of belief beyond their
premises, but they also change their degrees of belief in a process of belief
revision or updating. Reasoning to change beliefs has been called dynamic
reasoning (Oaksford & Chater, 2020). Let us say that we have a degree of
belief at time 1 in a hypothesis h, P1(h), but are interested in running an
experiment to produce some evidence, e, at time 2 to enable us to update our
degree of belief in h to P2(h). To make this update as Bayesians, we need
a subjective conditional probability judgement about h given e, P1(h|e), at time
1, and we derive P1(h|e) in the following way. By probability theory, when P1(e)
is not 0, P1(h|e) is P1(e & h) divided by P1(e), formally P1(e & h)/P1(e). By
probability theory again, P1(e & h) = P1(h)P1(e|h), and so we have:

P1(h|e) = P1(e & h)/P1(e) = (P1(h)P1(e|h))/P1(e)

As e is logically equivalent to (h & e) or (not-h & e), we can derive in
probability theory the total probability theorem that

P1(e) = P1((h & e) or (not-h & e)) = P1(h)P1(e/h) + P1(not-h)P1(e/not-h)

https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

and the following form of Bayes’ theorem:

P1(h|e) = (P1(h)P1(e|h))/(P1(h)P1(e/h) + P1(not-h)P1(e/not-h))

The derivation of this theorem is thus an exercise in deductive reasoning from

the axioms of probability theory, and it is incoherent to violate it. But the
theorem is the basis for going beyond logic and deduction in a Bayesian account
of updating beliefs in a dynamic process. More formally, this updating is to go
from P1(h) to P2(h) in light of new evidence e that has been acquired.
A special case of Bayes’ theorem is when is it is logically valid to infer e from
h, implying that P1(e|h) = 1. In this case, if not-e is the result of an experiment, our
hypothesis has been falsified, and our degree of belief P1(h) ought to be revised to
become P2(h) = 0 at time 2. Bayesians can use this instance of the theorem to
interpret Popper’s (1959) method of falsifiability in science. For Popper the only
legitimate scientific method is to use classical logic to derive a prediction from
a hypothesis. If the prediction turns out to be false, that is decisive evidence
against the hypothesis. Popper’s philosophy of science had a far-reaching impact
 20 Philosophy of Mind

on traditional psychology of reasoning (Wason & Johnson-Laird, 1972), but there

are serious problems with its narrow focus on falsifying hypotheses, to the
exclusion of confirmation, in the philosophy of science and psychological
research (Evans, 2007a; Howson & Urbach, 2006). After long careers in the
experimental study of reasoning, we testify how difficult it can be to falsify
psychological theories definitively. These can be modified in the face of negative
results and yet, sometimes, claimed to be the same theory.
Bayesians use Bayes’ theorem in an account of belief change that is more
general than Popper’s. It allows for the confirmation, as well as disconfirmation
(itself wider than falsification), of hypotheses and general belief revision and
updating (Howson & Urbach, 2006; Sprenger & Hartmann, 2019). The theorem
can give us a value for P1(h|e), supposing we can make judgements about
P1(h), noting that P1(not-h) = 1 – P1(h), and the likelihoods P1(e|h) and
P1(e|not-h). Assume an experiment is conducted, or an observation is made,
and new information becomes available that e holds at time 2, P2(e) = 1. We can
then use Bayesian strict conditionalisation to update our prior degree of belief
in h, P1(h), to a new posterior degree of belief, P2(h) = P1(h|e), provided that
P1(h|e) is invariant: P1(h|e) = P2(h|e). Invariance holds when learning the new
information e does not also lead to a change in relevant conditional probability
judgements. As we have noted, the proof of Bayes’ theorem is a deduction from
the axioms of probability theory, and it would be logically incoherent to give
some other value to P1(h|e). However, a judgement that invariance holds,
P1(h|e) = P2(h|e), can go beyond deductive logic, and it is not necessarily
logically incoherent to infer that invariance does not hold in some contexts.
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

For a simple example of strict conditionalisation, suppose there are

two coins, one double-head and one fair. Let one of these coins be
selected at random and h be the hypothesis that the selected coin is the double-
headed one. Suppose the coin is spun and clearly comes up heads. As
Bayesians, we can reason in the following way. Before the spinning, we judge
P1(h) = P1(not-h) = .5. With he as the evidence the coin comes up heads,
P1(he|h) = 1 and P1(he|not-h) = .5. In this special case. Bayes’ theorem implies
that P1(h|he) = .5/.75 = .66. When the coin comes up heads, he, we can finally
conclude that P2(h) = .66 using strict Bayesian conditionalisation. This
dynamic reasoning process has given us a higher degree of belief that the
selected coin is the double-headed one.
Not even this simple example can be a purely deductive exercise about
coherence. We have inferred, at least implicitly, that the probability the coin
will end its spin on its edge, and not with a head or tail, is negligible, but that
may not be so for this particular coin. We have also effectively presupposed
invariance. If we learn, not only that he holds, but that a sleight of hand trick
 Human Reasoning 21

substituted a double-tailed coin for the fair one before the random selection,
then invariance will not hold, and we will conclude that P2(h) = 1. Non-
deductive inferences, the topic of Section 4, are present even in this rudimentary
example.
For a less artificial example, recall our earlier statement:

(3.1) If Arthur is in Edinburgh (d), then Carol is in Glasgow (g).

Assume we have a modest degree of belief that Arthur is in Edinburgh. We

could try to falsify it by going to Edinburgh to look for Arthur. The problem is
that we could search for him indefinitely if he is not there. But suppose we
have a high degree of confidence that (3.1) holds, making P1(g|d) high for us
via the Equation, and we have a low degree of belief that Carol is in Glasgow
given Arthur is not in Edinburgh, making P1(g|not-d) low for us. Though P1
(d) may be at a middling value for us, we can use Bayes’ theorem to infer that
P1(d|g) is quite high. At this point, if we discover that Carol is definitely in
Glasgow, P2(g) = 1, and we are quite confident that invariance holds, we can
use strict conditionalisation to increase our confidence that Arthur is in
Edinburgh, P2(d) > P1(d).
In strict conditionalisation, we are certain of the evidence, and it can be
thought of, given the Equation, as dynamic MP, inferring a new degree of belief
in q from a degree of belief in if p then q after learning that p certainly holds.
Other inferences, like MT, have dynamic, belief-changing forms as well. But
invariance can more easily fail for MT than MP. Consider:

(3.3) If Arthur is not in Edinburgh, then he is in Scotland.

https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

We might have a high degree of belief in (3.3) because we know that Arthur was
going to catch the London train to Edinburgh but, by the train times, might be just
short of Edinburgh. If we learn, however, that Arthur is not in Scotland as a result of
missing his train in London, we will not appeal to (3.3) and conclude, using MT and
double negation (inferring p from not-not-p), that he is in Edinburgh. It will be clear
to us that invariance has failed and that (3.3) has a probability of 0. We saw in
Section 2.4 that, although MP and MT are both valid inferences in classical and
probability logic, the endorsement frequency of MT is significantly lower than that
of MP. Whether this fact can be explained by some awareness of the greater
fragility of dynamic MT is yet to be explored in the psychology of reasoning.
Strict conditionalisation is called ‘strict’ because we use it when we are
certain of the evidence. But perhaps there is some uncertainty about whether
Carol is in Glasgow, and so we cannot increase P2(g) to 1. Let us suppose we can
increase P2(g), so that P2(g) > P1(g). In this case, we can use a more general
Bayesian procedure than strict conditionalisation to go from the prior to the
 22 Philosophy of Mind

posterior, P1(d) < P2(d). This procedure is Jeffrey conditionalisation (Jeffrey,

1983, Ch. 11):

P2 ðd Þ ¼ P2 ð gÞP1 ðdjgÞ þ P2 ðnot-gÞP1 ðdjnot-gÞ

Invariance, P2(d|g) = P1(d|g) and P2(d|not-g) = P1(d|not-g), is again assumed

here. In our example, the probability that Arthur is in Edinburgh is determined
by the probability that Carol is in Glasgow and what that tells us about the
probability that Arthur in Edinburgh plus the probability that Carol is not in
Glasgow and what that tells us about the probability that Arthur in Edinburgh
(for Arthur might be in Edinburgh even if Carol is not in Glasgow).
Jeffrey conditionalisation can be thought of as dynamic MP from an
uncertain minor premise, but some experimental results imply that people
do not always conform precisely to it (Hadjichristidis et al., 2014; Zhao &
Osherson, 2010). They may focus on only one conditional component,
P1(d|g) or P1(d|not-g). This is implied by the descriptive singularity principle,
the psychological hypothesis that people tend to simplify, as a default, their
reasoning by focusing on only one hypothetical possibility at a time. There is
much psychological evidence for this principle (see Evans, 2006, 2007a).
Such singular focus, however, does not necessarily imply that they are
incoherent. It could be coherent for them to presuppose, for example, that
P1(d|not-g) = 0 in some contexts.

3.5 Diagnosticity
In our example of Bayesian reasoning about Arthur and Carol, we took the
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

antecedent of (3.1), d, as the hypothesis, and its consequent, g, as the

evidence, and as we have just explained, the singularity principle implies
that people might tend to focus on P1(d|g) in Jeffrey conditionalisation and
to ignore P1(d|not-g). This principle implies as well that, in the use of Bayes’
theorem at the start of this reasoning process, people might tend to concen-
trate on P1(g|d) and to ignore P1(g|not-d), which could actually be higher
than P1(g|d). They might tend to rely only on the evidence given the
hypothesis, P1(e|h), when they should be conforming fully to Bayes’ theorem,
which requires them to take the evidence given the negation of the hypothesis,
P1(e|not-h), into account as well. There is, indeed, some support for this
conclusion in experimental studies.
By Bayes’ theorem, e has variable diagnosticity, with P1(h|e) increasing as
P1(e|h) increases and P1(e|not-h) decreases. The likelihood ratio compares
P1(e|h) with P1(e|not-h), and people should, at least implicitly, make use of it
to comply fully with the theorem. But beginning with Doherty et al. (1979),
 Human Reasoning 23

psychologists have found that participants in studies can prefer information

only relevant to assessing P1(e|h), while ignoring possible information relevant
to P1(e|not-h), or to a number of possible hypotheses that are reasonable
alternatives to h. Their participants tended to focus on a favoured hypothesis
and to seek only confirmation of it. In technical terms, they reasoned pseudo-
diagnostically and had a confirmation bias.
However, information on P1(e|h) can be more available than information
about P1(e|not-h). We may often have experienced a cough as a symptom of
a common cold but have no knowledge of its frequency when people have some
kind of new viral infection, and it could be hard to find that out. But in some
contexts, as we have pointed out, it is logically coherent to assume to that
P1(e|not-h) is 0. It could be coherent, with certain background beliefs, to infer
that there is 0 probability Carol is in Glasgow given Arthur is not in Edinburgh.
Research following Doherty et al. (1979) has presented a more complex, and
less negative, picture of people’s diagnostic reasoning, but those results con-
tinue to be consistent with the singularity principle that people tend as a default
to focus their reasoning on a single hypothesis (Evans, 2006, 2007a).

3.6 Conclusion
In recent years, most psychologists have abandoned the traditional paradigm
view that the conditional in natural language is a material conditional when
people fully model its meaning. Surprisingly for a psychological paradigm, the
traditional approach paid little attention to inferences from beliefs, except as
a biasing factor, and made no attempt at all to explain changes in degrees of
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

belief. In contrast, the new paradigm focuses on the connection between

reasoning and degrees of belief and changes in degrees of belief. There is
more to reasoning from degrees of belief and belief change than deductive
inferences in probability theory. There are non-deductive inferences and strict
and Jeffrey conditionalisation with inferences about invariance. However, the
new paradigm does not expect people to be perfectly in line with any normative
principles, with no simplifying cognitive processes, or for their degrees of belief
always to match objective probabilities. It does expect to provide a better
scientific account of people’s reasoning by studying their degrees of belief
and how these change over time.

4 Non-deductive Reasoning
We introduced Bayesian reasoning in Section 3.4 and explained that the proof of
Bayes’ theorem is itself an instance of deductive reasoning from the axioms of
probability theory. That it is the starting point for Bayesian accounts of
 24 Philosophy of Mind

non-deductive reasoning, resulting in belief updating using strict or Jeffrey

conditionalisation when invariance holds. Deductive reasoning by itself cannot
give us a full account of belief change, and the traditional deductive paradigm
had nothing to say about updating degrees of belief. In this section, we discuss
in detail more examples of how human reasoning has to go beyond deduction, to
select hypotheses to test and to make inferences that are not deductive.

4.1 Wason’s 2–4–6 Task

Wason (1960) introduced the 2–4–6 task as an experiment on whether people try
to falsify hypotheses, as Popper would recommend (see Evans, 2016, for
detailed review of psychological work with this task). It is also an illustration
of the importance of selecting hypotheses to test and of not focusing on a single
hypothesis. Wason told his participants that he had a mind a rule for generating
whole number triples, and he gave them an example of a triple that conformed to
his rule: 2–4–6. He asked them to try to discover his rule by producing a new
triple themselves and asking him whether it conformed to his rule. He would
answer ‘Yes’ or ‘No’. After receiving his answer, they were able to choose
another triple, and the process would continue. He asked them as well to write
down their reasons for selecting their triples as guesses and to state the rule
when they thought they knew it.
Wason’s hidden rule was that any ascending sequence was acceptable, and it
is intuitive that most people would not think of this rule as their first hypothesis
about what Wason had in mind. What many appeared to do was to jump to the
conclusion that the rule was to ‘increase the three numbers by two’ or at least by
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

equal intervals, and they typically asked about three numbers consistent with
their hypothesis, for example, 5 7 9, in a positive test strategy. In such case they
would invariably receive a ‘Yes’ response, appearing to confirm their hypoth-
esis. The participants did not seem to Wason to be trying to falsify the hypoth-
eses that they considered. They could additionally be said to have assigned too
high a prior probability, P1(h), to the hypothesis that appeared to be suggested
by the context.
Influenced by Popper and helping to initiate the traditional paradigm, Wason
inferred that his participants had a serious failing in rationality. When Wason
and Johnson-Laird (1972) reviewed the 2–4–6 task, they charged the partici-
pants with having a confirmation bias, the tendency to seek verification in the
task and not falsification. They pointed out that a negative test of their hypoth-
esis, such as 1 5 6 would have falsified it. But the problem is that, in general,
positive testing can lead to falsification of hypotheses and is frequently used in
science as well as normal life (see Evans, 1989, 2016). If one predicts an
 Human Reasoning 25

experimental result that is not observed, that is a falsification. Hence, one cannot
infer a confirmatory attitude simply from the observation of positive testing, as
a number of later authors have noted. The difficulty of the 2–4–6 task arises
from the fact that the participant is cued by the biased example to adopt
a hypothesis which is a more specific case of the actual rule of any ascending
sequence. In this particular situation, all positive tests will get positive feed-
back. The hypothesis is not so much incorrect as insufficiently general.
Their mistake is like stating, truly, that iron expands when heated, without
testing if this property applies to metals in general.
Consider a Bayesian analysis of this task in the new paradigm. Because of the
way it is presented to them, the participants may assign a high prior probability
that Wason’s hidden rule was that to increase the three numbers by two. Letting
this hypothesis be ht, their implicit judgement was that P1(ht) was high at the
start of the task. The conditional probability of a ‘Yes’ answer, y, given ht and
the use of 5 7 9 as a guess is 1, for y can be logically deduced from ht and the use
of 5 7 9. Let us just express this judgement as P1(y|ht) = 1, with the understand-
ing that it has made in the context of 5 7 9 as the guess. The participants might
have been so focused on their hypothesis ht that they made no judgement about
the probability of y given its negation, P1(y|not-ht), but it is clear that a ‘No’,
not-y, reply to 5 7 9 should result, by strict conditionalisation, in a posterior
judgement that ht was certainly false, P2(ht) = 0, as P1(not-y|ht) should be 0
when P1(y|ht) is 1. The hypothesis would be falsified, and a high degree of prior
belief in it should fall to 0. A ‘Yes’ answer would confirm ht in a Bayesian
analysis, increasing the degree of belief in it, in this positive test strategy.
A negative test strategy would be to guess a triple that should receive a ‘No’
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

answer given the hypothesis, like 1 2 3. With the understanding that this triple is
the guess, P1(not-y|ht) = 1, and then falsification, from a ‘Yes’ reply, or
confirmation, from a ‘No’ reply, could again result.
More precision in a Bayesian analysis of the task is impossible when P1(ht)
and P1(y|not-ht) are not specified. At Wason’s suggestion, some studies of it did
eventually help participants to think about both a hypothesis, ht, and its neg-
ation, not-ht, or an instance of its negation, with dual goal instructions (Evans,
2016). This resulted in much higher solution rates on the 2–4–6 task, although
participants still followed a positive test strategy for a complementary hypoth-
esis which was only implicitly negative for the focal hypothesis. Note that
a negative test strategy is not normally needed, or even advantageous, in
many contexts (see Klayman & Ha, 1987; Poletiek, 2001). In a fully Bayesian
and new paradigm version of the task, the participants’ judgements P1(ht) and
P1(y|not-ht) would have to be measured, and the question would be how closely
their posterior judgement would be to the P2(ht) that was implied by Bayes’
 26 Philosophy of Mind

theorem and conditionalisation. But the existing studies do support the implica-
tion of the singularity principle (Evans, 2007a): that people will tend to focus on
a single hypothesis that is of importance to them, until they have a reason to give
it up. This is also consistent with the cognitive miser hypothesis (Stanovich,
2018), positing a system which evolved to minimise cognitive effort.
Of course, conforming to probability theory and being logically coherent
are not enough for effective reasoning and belief updating in the real world.
People must not only go beyond probability theory by using strict or Jeffrey
conditionalisation and making inferences about invariance. They must also
have a strategy for considering hypotheses and updating them. There is no
proof in probability theory that obliges us, on pain of incoherence, to have
a positive or negative test strategy or some combination of the two. Someone
could be logically coherent forever by asking Wason about numbers that
increased by two.
There are then lessons to be learned from Wason’s 2–4–6 task, but it can still
be argued that it does not correspond well to scientific investigations of hypoth-
eses or to most ordinary belief updating. It could be charged that it is artificial
for people to have to trust someone to tell them whether a hidden rule had been
conformed to by some data. But we do have to depend in ordinary reasoning on
experts to tell us what has been confirmed and well established. We face too
many complex questions and problems in our contemporary world, even in our
everyday reasoning and decision-making, to rely only on whatever highly
limited expertise we may possess (see Sloman & Fernbach, 2017, on how we
‘never think alone’). Even if we derive P1(h|e) using Bayes’ theorem, we may
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

well have to trust an expert to tell us that e holds and concluding that someone is
trustworthy has to be mainly the result of non-deductive reasoning. In the next
section, we will introduce experiments on relying on experts in arguments.

4.2 Argumentation
The study of argumentation in psychology seeks to account for people’s infor-
mal and ordinary use of arguments in reasoning (Hahn & Oaksford, 2007;
Oaksford & Chater, 2020). Many common arguments are deductively valid,
with both classical and p-validity. The best example is simply MP, which is so
common, and often so automatic, that its use can pass unnoticed. But the
premises of MP and other valid inferences usually have to be supported by
arguments that are not deductively valid. A Bayesian new paradigm approach
reveals that some of these invalid inferences can be strong, well-justified
reasoning in some contexts, with highly probable conclusions, and not the
unqualified ‘fallacies’ of a traditional classification. For example, the classical
 Human Reasoning 27

fallacy of Affirmation of the Consequent (AC), inferring h from if h then e and e,

is classically invalid and p-invalid, but it can sometimes be given
a straightforward justification using Bayes’ theorem and conditionalisation,
leading to updating of the degree of belief in h.
One way we all have to simplify and widen our reasoning is by relying on
experts to a large extent. To decide whether to take Ivermectin to try to prevent
COVID-19, we cannot run the relevant scientific experiments ourselves but
must try to identify the appropriate experts who are also trustworthy. Harris
et al. (2016) studied what was sometimes classically called the appeal to
authority argument. This argument form is clearly logically invalid. Relying
on even a well-qualified and honest authority can lead to false beliefs. But
Harris et al. (2016) proposed a Bayesian model of this form and considered the
conditional probability that a hypothesis will be asserted by an expert given that
it holds, P1(ha|h), as we will symbolise it here. If P1(ha|h) is much greater than
P1(ha|not-h), that is, the likelihood ratio is high, and our prior probability
judgement for h, P1(h), is also not very low, then P1(h|ha), the probability that
h holds given that the possible expert has asserted it, will be relatively high, and
this ‘appeal to authority’ argument will be a relatively strong one, leading to
a higher posterior degree of belief in h, P2(h), when the expert does assert h, that
is, ha holds.
To refine their Bayesian model, Harris et al. introduced the expertise, ex, and
the trustworthiness, tr, of the supposed expert who asserts the hypothesis as
factors in a Bayesian analysis of this argument form. All the relevant condi-
tional probabilities are represented, e.g., P1(ha|h & ex & tr) is the conditional
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

probability that the hypothesis is asserted given that it holds, that the asserter is
an expert, and that the asserter is trustworthy. The model is simplified by
assuming that h, ex, and tr are independent of each other. Independence, and
conditional independence, assumptions are of importance in simplifying
Bayesian models. In this instance, the independence assumption allows Harris
et al. to predict what the P1(h|ha) judgements should be, according to the model,
from judgements about P1(ha|h & ex & tr) and the other relevant probabilities.
Their two experiments gave good, though somewhat noisy, support to their
model.
This model is computational, as Harris et al. emphasise. They do not try to
give an account of the underlying psychological processes. These may
be relatively simple in some cases. We might believe that, if Ivermectin is
not effective for treating COVID-19, then the USA Food and Drug
Administration (FDA) will not approve it for treating COVID-19, and
infer from this that, if the FDA does not approve it for treating COVID-
19, then it is not effective for treating COVID-19. Inferring high confidence
 28 Philosophy of Mind

in if p then q from high confidence in if q then p can be a strong inference in

some contexts (justified by Bayes’ theorem). We might infer that the FDA
gives expert advice because it is similar to the European Medicines Agency
(EMA), and we already believe that the EMA gives expert advice (see
Feeney, 2018, on this kind of inference based on similarity). There are
always many questions for psychologists to answer to complete Bayesian
computational theories of reasoning.

4.3 Base Rates and Natural Sampling

In spite of the successes of Bayesian computational models (Oaksford & Chater,
2020), there is reason to think that it can sometimes be difficult for people to use
Bayesian reasoning. Suppose that participants in a study are given the following
Bayesian problem:

There is a .02 probability that a medical expert will be picked out from
a group in a certain context, P1(hmex) = .02. The probability is 1 the person
picked out will state that Veklury can be an effective treatment for COVID-19
if they are a medical expert, P1(hsv|hmex) = 1, and a probability of .5 that the
picked out person will state Veklury can be an effective treatment for
COVID-19 if they are not a medical expert, P1(hsv|not-hmex) = .5. What is
the probably the person is a medical expert if they state that Veklury can be an
effective treatment for COVID-19, P1(hmex|hsv)?

The correct answer to the above question, using Bayes’ theorem, is about .04,
that is, P1(hmex|hsv) is about .04. Studies of problems like the above have been
run for a long time, following on from Tversky & Kahneman (1974), and it has
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

been found that participants tend not to answer them correctly, responding that
P1(hmex|hsv) is much higher than .04. They seem to be suffering from the
fallacy of base-rate neglect. In above example, this fallacy would be not to
take account, or insufficient account, of the fact that the prior probability, the
base rate, that the person picked out is a medical expert is so low, with P1
(hmex) = .02. It is true that medical experts will apparently make the statement
about Veklury with certainty, and the non-experts will respond with this
statement at random, but there are many more non-experts than experts in
this sample, and there will be many more statements made about Veklury as an
effective treatment for COVID-19 by non-experts than by experts. The fact
that someone makes this statement will therefore not be strong evidence that
they are a medical expert.
Influenced by Kleiter (1994), Gigerenzer and Hoffrage (1995) found that
Bayesian reasoning can be improved if experimenters use what they called
natural frequencies in the Bayesian word problems given to their participants.
 Human Reasoning 29

For the above problem, a natural sampling (Kleiter, 1994) version using natural
frequencies would go like this:

In a sample of 100 people, 2 are medical experts and 98 are not medical
experts. The two who are medical experts will state that Veklury can be an
effective treatment for COVID-19. Out of the 98 who are not medical experts,
49 will state that Veklury can be an effective treatment for COVID-19. Out of
all those who state that Veklury can be an effective treatment for COVID-19
how many will be medical experts, ___ out of ___?

The correct answer to this problem is 2 out of 51, and it is clearly easier for most
people to come by than .04, or more accurately .039, or an even more accurate
longer decimal expansion. Cosmides and Tooby (1996) used an evolutionary
argument to try to explain why such versions of Bayesian problems
are relatively easy to solve. They argued that our human ancestors, who lived
as hunter-gatherers for almost all of our evolutionary history, acquired, by
natural selection, an adaptive dedicated module for easily processing ‘natural
frequencies’. For example, hunter-gatherers might, arguably, have concluded
that 47 out of 52 of their group members who ate a certain red berry recovered
from a cold within a week.
We doubt that an evolutionary hypothesis about a dedicated module for
natural sampling can explain how a ‘natural frequency’ format facilitates
Bayesian reasoning in a psychological word problem (see Evans, 2007a and
Over, 2003). Even if a group of hunter-gatherers had the linguistic means to
represent a number like 47 or 52, this sample from their group might well not tell
them that eating the red berry explains a recovery in a week (Sebben & Ullrich,
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

2021). The sample could be biased, and even if it is not, people may recover
from a cold within a week whether they eat the berry or not.
The term ‘natural frequency’ suggests that there is an objective frequency in
play here. But natural sampling only gives us sample frequencies, which must
be processed further in non-deductive reasoning for high degrees of confidence
in objective frequencies and causes. Bayesian belief updating, with its non-
deductive aspects, is itself needed to take us from a sample frequency of, for
instance, five heads in a row to a high degree of confidence that a coin has a bias
which results in a high objective probability of coming up heads.
We do not doubt, of course, that hunter-gatherers inferred the existence of some
objective frequencies and causes through non-deductive reasoning that increased
their reproductive success. They could notice that there was rapid new berry and
other growth that attracted game after a fire cleared a forest. They apparently
started fires themselves for this purpose for tens of thousands of years (Zimmer,
2021). These interventions could have, in effect, confirmed for them a Bayesian
 30 Philosophy of Mind

model for such causal reasoning (Pearl, 2000; Sloman & Lagnado, 2015). Some
simple natural sampling would have played a part in this process. Still, solving
a word problem in a psychological experiment is a long way from starting fires to
encourage the growth of blueberries, and a dedicated module for natural sampling
could help with the latter, but not the former.
For us, Tversky and Kahneman (1983) had the key insight for explaining
why participants in psychological experiments can solve Bayesian word
problems when these are in a ‘natural frequency’ format. In their article,
as we have reported in Section 3.2, they found that people sometimes
commit the conjunction fallacy of, for example, judging that it is more
probable than a man is over 55 and has had one or more heart attacks than
that the man has had one or more heart attacks. But they also found that
instances of this fallacy significantly decreased when it was expressed using
sample frequencies. In one study, they told their participants that there was
a sample of 100 males, asked how many of these 100 men had one or more
heart attacks, and how many of these 100 men both were over 55 years and
had had one or more heart attacks. Given these questions, we would avoid
the conjunction fallacy by replying, for example, that 10 of these 100 men
had had one or more heart attacks and that 8 of these had had 1 or more heart
attacks and were over 55. Tversky and Kahneman pointed out that, in this
version of the task, the class inclusion relation can be ‘readily perceived and
appreciated’. The class of men who have had one or more heart attacks and
are over 55 is necessarily included in the class of men who have had one or
more heart attacks, and this relationship can be literally seen when we draw
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

Euler circles in which the circle for former class is inside the circle for the
latter class (Barbey & Sloman, 2007).
In mathematical set theory, when one class includes another, the latter is also
said to be a subset of the former, and we can say that, when this relation holds,
the subset and the set that includes it are ‘nested’, as Euler circles can be nested.
The nested-sets hypothesis implies that a nested-sets format is what helps
people to answer questions about sample frequencies, avoiding the conjunction
fallacy and base-rate neglect. This hypothesis has been confirmed in many
studies (Barbey & Sloman, 2007). The natural sampling format puts its samples
in a nest-sets relationship, and it now seems to be generally accepted that
a ‘natural frequency’ problem has to have a nested-sets structure to be trivially
easy to solve (McDowell & Jacobs, 2017).
There is a necessary connection between elementary set theory and logical
relations, as we effectively illustrated when we introduced the conjunction
fallacy in Section 3.2. This can be seen in the Euler circle representation of
logical relations. The circle of p & q truths is inside the circle of p truths,
 Human Reasoning 31

representing the logical fact that the set of cases in which p & q is true is necessarily
a subset of the cases in which p is true. Specifically, the rows of a truth table in which
p & q is true is a subset of the rows of a truth table in which p is true. Making logical
relations clear with a set-subset and nested-sets presentation helps people to avoid
fallacies, both in classical logic and the logic of probability. When we give our
participants the free gift of a nested-sets representation in an experimental word
problem, they can do well. But in the real world, people do not usually receive this
free gift, and then how often they do well with non-deductive reasoning about
sample frequencies is a different question.

4.4 Conclusion
Effective non-deductive reasoning for reproduction, argumentation, or other
goals has to be more than logical inferences about sets and subsets. This is just
to repeat, in other words, that effective reasoning from degrees of belief and for
belief change has to be more than deductive inferences in probability theory.
There are inferences about invariance, which are usually non-deductive, to
account for. And Wason’s participants could have been logically coherent by
naively accepting the hypothesis suggested by the context of his 2–4–6 task and
asking indefinitely about triples increasing by two. Choices must be made about
which hypotheses to test and how to test them. But no matter what normative
principles are proposed for people’s reasoning, they will have to simplify it.
They can sometimes do this by trying to rely on experts, but identifying experts
is sometimes itself hard for people, and they can fail miserably at it. There can
be effective and efficient ways or heuristics to make inferences, but these are not
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

perfect and can also cause problems. That must have been true even when we
were in the happy state of being hunter-gatherers.
We have continued to assume in this section that the Equation, P(if p then q) =
P(q|p), holds for natural language conditionals. Deductive and non-deductive
inferences from p to q, and the strength of these inferences, can be ‘summed up’
in if p then q, and in P(q|p). It is now time for a fuller account of if p then q which
implies the Equation.

5 Conditionals
The new paradigm in the psychology of reasoning identifies the probability of
the natural language conditional, P(if p then q), with the conditional subjective
probability of q given p, P(q|p), and aims to integrate the study of conditional
reasoning with studies of subjective probability and utility judgements. The
origin of the new paradigm can be traced back to de Finetti’s and Ramsey’s
foundational studies of subjective probability (Section 3.3).
 32 Philosophy of Mind

5.1 De Finetti’s Conditional in Psychology

Wason (1966) found that participants in an experiment did not fully conform to the
truth table for the material conditional. They agreed that if p then q is true when p is
true and q is true and is false when p is true and q is false. But when they were given
a false antecedent not-p case, they tended to respond that this was ‘irrelevant’ to
assessing if p then q for truth and falsity. The resulting three-valued table came to be
known as the ‘defective’ truth table in psychology (Over & Baratgin, 2017).
The new probabilistic paradigm in the psychology of reasoning has led to
a strong interest in de Finetti’s (1936/1995) three-valued table for his condi-
tional and in his associated logic of subjective probability, with a third value, u,
indicating doubt or uncertainty (see Egré et al., 2021, on de Finettian three-
valued semantics). The subjective focus of de Finetti’s approach makes it
particularly relevant to psychology. Table 2 is the de Finetti table for the
conditional.
In de Finetti’s account, if p then q is judged to ‘true’ when p and q are known to
be true, and ‘false’ when p is known to be true and q false. But when p is known to
be false, de Finetti held that if p then q is doubtful or uncertain (‘douteux’ in the
original French of Finetti, 1936/1995), and ‘void’ (to use the technical term of de
Finetti, 1937/1964). There have been many logical proposals for three-valued
analyses of conditionals, and these can be studied in psychological experiments to
explore how far ordinary’s people’s judgements correspond to them (Baratgin
et al., 2018). But in de Finetti’s account, the third value u expresses doubt or
uncertainty as a subjective mental state. It can be refined into a more precise
conditional subjective probability judgement at a higher level of understanding
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

(Baratgin et al., 2018; Over & Baratgin, 2017).

Skovgaard-Olsen et al. (2017) studied the ‘defective’ truth table and found
that there were not as many ‘irrelevant’ or ‘neither true nor false’ responses as
would be predicted from the de Finetti table. But they noted that the de Finetti
table might do better if it were combined with the hypothesis that people

Table 2 The de Finetti Table for if

p then q

pq if p then q
TT T
TF F
FT u
FF u

T = true, F = false, and u = uncertain.

 Human Reasoning 33

sometimes pragmatically interpret the assertion of a conditional, if p then q, as

conveying a biconditional p if and only if q. For instance, people might
pragmatically infer ‘If the daughter became a millionaire, then she inherited
all the money’ from an assertion of ‘If the daughter inherits all the money, then
she will become a millionaire’.
As Baratgin et al. (2018) pointed out, de Finetti’s third value expresses a state
of subjective uncertainty and not some other vaguer notion of ‘irrelevance’. They
confirmed that his three-valued table does better than other proposed tables at
accounting for participants’ responses in experiments where there is such state of
doubt, and the conditionals could not be interpreted as biconditionals. In their
materials, they used round and square chips that could be black or white, with
conditionals like the following about a randomly selected chip:

(5.1) If the chip is square (s) then it is black (b).

In their design, s and b could be uncertain as well as if s then b. The uncertainty

was visual. A ‘filter’ made it impossible at times to tell whether or not a given
chip was round or square or black or white.
A de Finetti analysis of the natural language conditional if p then q implies
further that there is a close relation between someone’s assertion of it and
a conditional bet if p then I bet that q. It would make no real difference whether
one asserted (5.1) or ‘If the chip is square then I bet it is black’. The conditional
bet is won when p and q are true and lost when p is true and q is false. Table 2
could equally represent a conditional bet, with T and F in the final column being
‘won’ and ‘lost’. As we pointed out in Section 3.3 for the type of conditional bet
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

described there, the probability it will be won is P(q|p), which is its expected
value. Baratgin et al. (2013) and Politzer et al. (2010) confirmed this close
relationship in people’ judgements.
In an extension of de Finetti’s analysis, in which the third value u expresses
uncertainty about if p then q, u becomes the conditional probability P(q|p) itself
(Sanfilippo et al., 2020). It is the expected value, or ‘prevision’ to use de

Table 3 The Jeffrey table for if p then q

pq if p then q
TT T
TF F
FT P(q|p)
FF P(q|p)

T = true, F = false, and P(q|p) = the conditional

subjective probability of q given p
 34 Philosophy of Mind

Finetti’s term, of an assertion of the conditional. Table 3 has come to be known

as the Jeffrey table (Jeffrey, 1991; Over & Cruz 2018, 2023) for if p then q. The
Jeffrey table avoids a problem with the de Finetti table. The trivial logical truth
if p & q then p, for example, should not have the value u when p is false and
u means ‘doubtful’ or ‘uncertain’. It should have the value 1, and P(p|(p & q)) is
of course 1.
The Jeffrey table combines very well with a pragmatic account of some uses
of ‘true’ and ‘false’ (Adams, 1998; Over & Cruz, 2023), to complement
objective interpretations of truth and falsity in the p & q and the p & not-q
cases. In the pragmatic use, the assertion of if p then q would be termed ‘true’
when P(q|p) at or near 1, and ‘false’ when P(q|p) was at or near 0. With this
supplementary hypothesis about pragmatic uses of ‘true’ and ‘false’, asserting
that if p then q is ‘true’, or ‘false’, does not necessarily imply, for followers of de
Finetti, that p is true, but only that P(q|p) is high for the pragmatic context.
We introduced the Equation, P(if p then q) = P(q|p), for the conditional in
natural language, in Section 2.6. Psychologists of reasoning have highly confirmed
it in experiments as the conditional probability hypothesis for people’s judgements
about a wide range of indicative conditionals and some counterfactuals (see the
reviews by Evans & Over, 2004, Over, 2020, and Over & Cruz, 2018, 2023).
Supporters of the Equation commonly combine it with the Ramsey test (Ramsey,
1929/1990) as a psychological process for determining P(if p then q). According to
this ‘test’, as extended by Stalnaker (1968), we are to make a judgement about
if p then q by supposing p is the case, while making minimal changes to our beliefs
to preserve consistency, and then judging our degree of belief in q under this
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

supposition. The result is then a judgement about the conditional subjective

probability of q given p, P(q|p), and this is identified with P(if p then q).
A conditional satisfying the Equation has been called a ‘probability condi-
tional’ (Adams, 1998), and a ‘conditional event’ (de Finetti, 1937/1964), but in
this Element we will use suppositional conditional as a general term for such
a conditional, and suppositional theories (like that in Evans & Over, 2004) for
accounts that imply that the Equation holds for the natural language conditional.
A suppositional conditional if p then q can also be read informally as
q supposing p (see Evans, 2020, on pragmatic features of a suppositional
conditional). Lewis (1976) proved that a conditional if p then q in the bivalent
and modal theories of Lewis (1973) and Stalnaker (1968) cannot generally have
a probability that is equal to the conditional probability, P(q|p). In other words,
a Lewis or Stalnaker conditional cannot be thought of as a suppositional
conditional. (See Edgington, 1995, for an intuitive account of Lewis’s result,
and Cantwell, 2020, for further points). The proof does not apply to de Finetti
approaches, which are not bivalent (Lassiter, 2019; Sanfilippo et al., 2020).
 Human Reasoning 35

5.2 Inferential connections

Suppositional accounts of the conditional fit perfectly into the new Bayesian
paradigm in the psychology of reasoning. As we have explained in Section 3,
this new approach does not restrict itself to inferences from arbitrary premises
that are to be assumed true, outside of any context that makes them relevant, but
focuses much more on inferences from degrees of belief to degrees of belief in
static reasoning and on dynamic inferences through time leading to belief
revisions or updating. We will cover static reasoning next in this section and
move on to dynamic reasoning in the next.
Let us recall the definition of p-validity from Section 3.2. An inference is
p-valid if and only if the uncertainty of its conclusion cannot be coherently
greater than the sum the uncertainties of its premises (Adams, 1998). Consider
this example as a suppositional conditional:

(5.2) If Anne invests in the cryptocurrency scheme (a), then Carol will too (c).

Suppose our degree of belief in (5.2) is P(c|a) = .6, and our degree of belief in a is
P(a) = .7, and consider MP, inferring c from if a then c and a. Let us also say that,
after making this inference, we judge P(c) to be .6. Our degree of belief in c, .6, is
coherent and does not violate the p-validity of MP as the uncertainty of c, (1 – .6),
is not greater than the sum of the uncertainties of the premises, (1 – .6) + (1 – .7).
The paradoxes of the material conditional are prime examples of inferences
that are not p-valid for suppositional conditionals. P(q|p) can be low when P
(not-p) is high, and when P(q) is high, and experiments have shown that people
do not at all accept these inferences as p-valid in their probability judgements
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

(Cruz et al., 2017). The formal system for p-validity (Adams, 1998) is some-
times called System P (Gilio et al., 2020; Pfeifer & Kleiter, 2009). It descends
from de Finetti’s (1936/1995) logic of probability and Ramsey’s (1926/1990)
logic of partial belief.
As we saw in Section 3.3, coherence in the logic of probability corresponds to
consistency in classical logic but is more general. It is of particular relevance in
the new paradigm, as are the logical coherence intervals that can be derived for
inferences (Pfeifer & Kleiter, 2009). A coherence interval for an inference
specifies the range in which the probability of the conclusion must fall for
coherence. For example, the coherence interval for MP follows from the total
probability theorem:

PðcÞ ¼ PðaÞPðcjaÞ þ Pðnot-aÞPðcjnot-aÞ

By probability theory, we know that P(not-a) = 1 – P(a), but we might be unable

to make a definite judgement about P(c|not-a), which has to be used to
 36 Philosophy of Mind

determine the precise ‘total’ probability of c. But instead of a precise value, we

can infer an interval for the probability of c. Carol may have a reason for
investing in the scheme even if Anne does not invest in it, or Carol might not
invest if Carol does not. In any event, P(c|not-a) can have a maximum value of
1, and a minimum value of 0, by probability theory. If P(c|not-a) = 1, then
P(c) = P(a)P(c|a) + 1 – P(a) = (.6)(.7) + .3 = .72. If P(c|not-a) = 0, then
P(c) = P(a)P(c|a) = (.6)(.7) = .42. The coherence interval in this case of MP is
then from .42 to .72, and our judgement that P(c) = .6 is coherent. Recall the
singularity principle from Section 3.4 and note that people will still be coherent,
at the minimum of the interval, even if singularity causes them to focus on
P(c|a) and in effect judge that P(c|not-a) = 0. This principle does not necessarily
make people incoherent.
But a judgement that P(c) = .95 would be incoherent and a kind of overconfi-
dence, and one that P(c) = .25 would be incoherent and a kind of under-
confidence. These judgements could also be called a type of ‘belief bias’
resulting from too much belief in c, or too little belief in c, than MP justifies.
There is also a coherence interval for AC, inferring a from if a then c and c, and
other conditional inferences. Evans at al. (2015) found that people stay within
these coherence intervals at an above chance rate, just as they conform to
p-validity above chance (Section 3.2), and they do even better by both measures
when the premises and conclusions are presented together for explicit
inferences and not judged separately.
Edgington (1995) argued strongly for the Equation, but she suggested that it
might be disconfirmed for pragmatically problematic conditionals, as she pre-
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

supposed, like the following:

(5.3) If Napoleon is dead (n), Oxford is in England (o).

She remarked that many people might judge (5.3) as ‘ . . . not acceptable, or
even, false’. The Equation/conditional probability hypothesis implies that
(5.3) will be assessed as highly probable and not definitely false, since
P(o|n) is high. In general, P(q|p) is high when P(q) is high and p and q are
independent. Pragmatically unacceptable conditionals like (5.3) are some-
times called ‘missing-link’ conditionals, but this term can be misleading, as
there are pragmatically acceptable conditionals with ‘missing-links’ in some
sense, for example, (3.3) above, ‘If Arthur is not in Edinburgh, then he is in
Scotland’. In sharp contrast, conditionals like (5.3) are not pragmatically all
right, and we will call them Walrus conditionals (Cruz & Over, 2024; Over,
2023b). In Lewis Carroll’s nonsense poem, ‘The Walrus and the carpenter’,
the Walrus says, ‘ . . . the time has come to talk of many things: of shoes, and
ships, and sealing-wax, of cabbages, and kings’. The Walrus’s ‘conversation’
 Human Reasoning 37

is a pragmatically bizarre jumble, and a Walrus conditional is similar. Its

antecedent and consequent do not have a pragmatic or semantic relation to
each other in a given context.
Skovgaard-Olsen et al. (2016) confirmed Edgington’s suggestion about what
we are calling Walrus conditionals. They tested the conditional probability
hypothesis, that P(if p then q) = P(q|p), and found that participants in an
experiment did judge P(if p then q) to be lower than P(q|p) for Walrus condi-
tionals (see Over & Cruz, 2023, on Skovgaard-Olsen et al.’s findings about non-
Walrus conditionals). In reply, Cruz et al. (2016) hypothesised that the problem
with Walrus conditionals is that they do not contain a common topic of
discourse, and they provided experimental support for this hypothesis (as do
Bourlier et al., 2023). Lassiter (2023) develops the concept of pragmatic
discourse coherence. This is much wider than logical coherence, which is
consistency with probability theory. He explains why (to use our terms)
Walrus ‘conversations’ and Walrus conditionals do not have pragmatic discourse
coherence, unless a special context is given for them.
It was a significant step forward for Skovgaard-Olsen et al. to investigate
Walrus conditionals and establish their result about them, but compare these
three conditionals:

(5.4) If Mark presses the power button on his TV, then the TV will be turned on.

(5.5) If Mark is wearing socks (w), then his TV will be working (v).

(5.6) If Mark presses the power button on his TV, then the screen will remain
blank.
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

Conditionals like (5.4) are dependence conditionals: their antecedents increase

the probability of their consequents. Conditionals like (5.5) and (5.6) could be
vaguely said to have missing causal or other ‘links’ and can be more precisely
termed independence conditionals (Cruz & Over, 2023). In these conditionals,
the consequent is independent of the antecedent, for example, P(v|w) = P(v|not-w).
But there is a big difference between (5.5) and (5.6). For (5.5) does not have
pragmatic discourse coherence, and that makes it a Walrus conditional.
Skovgaard-Olsen et al. compared conditionals like (5.4) and (5.5) with each
other, finding that only probability judgements about the dependence conditional
(5.4) complied with the conditional probability hypothesis. But going from (5.4) to
(5.5), they changed a pragmatically acceptable dependence conditional into
a pragmatically unacceptable independence conditional, that is, a Walrus condi-
tional (Cruz & Over, 2024; Over, 2023b). This confound could be avoided in the
future by asking for probability judgements about conditionals like (5.4) and (5.6),
which are both pragmatically acceptable. In (5.6), we have an example of a
 38 Philosophy of Mind

non-Walrus independence conditional that could pragmatically convey useful

information about Mark’s TV.
There is a debate about the significance of experiments on ‘missing-link’ or
Walrus conditionals. There is an inferentialist view that a ‘standard’ conditional
if p then q is true if and only if there is a non-redundant deductive, or sufficiently
strong inductive, abductive, or other non-deductive, relation between p and
q (Douven et al., 2023; van Rooij et al., 2023). This semantic proposal implies
that the conditional probability hypothesis fails for Walrus conditionals because
these conditionals are supposedly not true. On the other side of the debate,
theorists argue for Edgington’s original presupposition that pragmatics can fully
explain the problem with Walrus conditionals (Bourlier et al., 2023; Cruz &
Over, 2023, 2024; Lassiter, 2023; Over & Cruz, 2023). There is experimental
evidence that people classify Walrus conditionals if p then q as ‘true’ when p and
q are true (see Skovgaard-Olsen et al., 2017, on ‘irrelevance’ cases), and also
strong support for the conditional probability hypothesis in experiments where
the materials are pragmatically acceptable (see Sections 2.6 and 5.1 and add-
itionally Kleiter et al., 2018, Oberauer et al., 2007, Pfeifer, 2023, Singmann
et al., 2014, and Wang et al., 2022).
The relation P(q|p) > P(q|not-p) holds for dependence conditionals. This is
the formal equivalent of stating that p raises the probability of q. But there are
grounds for concluding that this relation is neither necessary nor sufficient for
pragmatic acceptability. The de Finetti (1937/1964) normal form for if p then
q is if p then (p & q). Note that, although P(q|p) and P((p & q)|p) are necessarily
equal, if p then (p & q) can be dependence conditional when if p then q is an
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

independence conditional. Consider the normal form of (5.5), that is, if w then
(w & v).Now P(v|w) = P(v|not-w) = P((w & v)|w) = P(v), and yet we have
P((w & v)|w) > P((w & v)|not-w) when P(v) is not 0. Therefore, ‘If Mark is
wearing socks, then he is wearing socks and his TV will be working’ is
a dependence conditional, but that is intuitively not sufficient to make it more
pragmatically acceptable than ‘If Mark is wearing socks, then his TV will be
working’.
As we have already illustrated with (3.3) and (5.6), there are pragmatically
acceptable conditionals which are independence and not dependence condition-
als. For another example, doctors could say to parents whose children are
known not to have autism:

(5.7) If your children are vaccinated, they will not get autism.

The pragmatic implication of (5.7) is that whether children will get autism is
independent of whether they have vaccinations. Knowing about independence,
and conveying information about it using conditionals, is of great utility in
 Human Reasoning 39

human reasoning and decision-making (Cruz & Over, 2023; Over, 2023b).
It is to the advantage of suppositional theories that they give unified accounts
of dependence and independence conditionals.

5.3 Dynamic Reasoning and Counterfactuals

Experiments have strongly supported the conditional probability hypoth-
esis, P(if p then q) = P(q|p) for conditionals with pragmatic discourse
coherence, and subjective conditional probability theory can guide us in
our static and dynamic conditional reasoning. There are the logical coher-
ence intervals we have described for reasoning in a static state, when our
degrees of belief do not change over time, and on top of these, we have the
Bayesian notions of strict and Jeffrey conditionalisation for dynamic rea-
soning when our beliefs are updated over time (Section 3.4; Oaksford &
Chater, 2020).
We can illustrate some more points about dynamic reasoning using the example
of an investigation of a crime. Suppose that a precious jewel has apparently been
stolen at a country house, and the police have high confidence that:

(5.8) The butler stole the jewel (b) or the cook did (k).

On this basis, they can also infer with some confidence that:

(5.9) If the butler did not steal the jewel, the cook did.

Let us use P1 for the police’s probability judgements early in their investigations
of this case, and P2 for their later probability judgements. Notice immediately
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

that, assuming (5.9) is a suppositional conditional, the inference from (5.8) to

(5.9) is not p-valid and will only be probabilistically strong in some contexts.
(5.8) will be highly probable when it is inferred simply from strong evidence
that the butler stole the jewel, making P1(b) high, but then (5.9) can have a very
low probability, with P1(k|not-b) at or near 0 in this context (Cruz et al., 2015;
Gilio & Over, 2012).
However, let us suppose that the police have general grounds for judging that
the probability of (5.8) is relatively high. For example, the butler and cook are
the only suspects without alibis. The police can then have relatively high
confidence in (5.9). Let us also say P1(b) = P1(k) = .5, and P1(k|not-b) = .9.
After later finding definite evidence that the butler did not steal the jewel,
P2(not-b) = 1 the police could use Bayesian strict conditionalisation
(Section 3.4) to infer P2(k) = P1(k|not-b) = .9, provided that invariance in the
conditional probabilities holds, P2(k|not-b) = P1(k|not-b). Invariance can
fail when new information is acquired in the change from P1 to P2, as
 40 Philosophy of Mind

would happen if the police were to discover that the jewel was not stolen, but
only misplaced. In that case, P2(k) and P2(k|not-b) would both be 0.
Suppose, however, that the police become more and more convinced in
dynamic reasoning that the butler stole the jewel. At some point in this process,
they might start to use counterfactual conditionals, beginning with ‘If the butler
had not stolen the jewel . . .’, but they could reject as improbable:

(5.10) If the butler had not stolen the jewel, then the cook would have.

Using the Ramsey test (Section 5.1), the police could reason that the cook would
not be turned into a criminal supposing something intervened to prevent the
butler from stealing the jewel (see Pearl, 2013, on Ramsey’s ‘idea’, and
Kaufmann, 2023, for a temporal analysis of counterfactuals). If they got still
more evidence and started to go backwards, losing their confidence that the
butler is the culprit, their counterfactual thought (5.10) could return to the
indicative (5.9) and their previously high confidence in that. In recent years,
psychologists of reasoning have used Bayesian networks (Pearl, 2000) to
represent people’s conditional reasoning about causes (Oaksford & Chater,
2020), but much more research is needed on how people move in their reasoning
from indicative conditionals to counterfactuals and, sometimes, back again
(Over & Cruz, 2023).
There is much research in psychology on counterfactuals and the emotions,
especially regret (see Corbett et al., 2023, on how the two are linked in human
beings even at a young age). Recall yet again the cryptocurrency scheme and
suppose it does turn out to be a scam. We could express regret in this way:
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

(5.11) If we had not invested in that scheme, we would not look like fools.

But assume our decision about the scheme was a long time ago, and we have
numerous investments, making it hard to remember the details. When we consult
our accountants, they tell us that we did not in fact make that investment, and we use
(5.11) as the major premise of MP to infer that we do not look like fools. Then our
regret vanishes. People do conform to MP when the major premise is
a counterfactual (see Thompson & Byrne, 2002, and Over & Cruz, 2023, for
comment). But psychologists do not have an account of how, or why, logical
reasoning with counterfactuals is so closely tied to people’s emotions, and philo-
sophers have so far paid no attention to this link in their theories of counterfactuals.

5.4 Conclusion
The compliance of humans with the Equation or conditional probability hypoth-
esis, P(if p then q) = P(q|p), has been strongly confirmed in psychological
 Human Reasoning 41

experiments for pragmatically acceptable conditionals. This provides empirical

support for suppositional theories of conditionals in which if p then q can be
read as q supposing p, and P(if p then q) = P(q|p). Ramsey and de Finetti laid the
theoretical foundations for these theories, which cover both dependence and
independence conditionals and can be extended to dynamic reasoning and belief
updating. There is an above chance tendency for people to reason within the
logical coherence intervals of suppositional conditionals, and also comply with
p-validity, especially in their explicit conditional reasoning, but whether this
tendency makes them rational is a question for Section 7.

6 Dual Processes in Reasoning

Dual-process theories have played – and continue to play – a significant role in the
psychology of reasoning. As we shall see, these accounts form part of a family of
theories whose origins predate modern cognitive psychology. They are independ-
ent of the new paradigm, which reflects a shift in thinking about how people
reason and the relevant normative standards to apply. Dual-process theory is more
concerned with the question of when people apply explicit reasoning to produce
their answers, as opposed to faster and lower effort intuitive processes. In this
section, we will briefly review the dual-process framework and more specifically
the history of dual-process theories in the psychology of reasoning.
A distinction between two kinds of thinking, one slow and reflective and the other
fast and intuitive has been around for hundreds of years in philosophical writing and
has featured heavily in ‘modern’ cognitive and social psychology (Frankish &
Evans, 2009). By modern, we mean roughly post-1960, which is about the date of
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

the cognitive revolution in psychology, in which information processing models of

the mind displaced the behaviourism that had been dominant for the previous fifty
years. The idea of nonconscious and conscious thinking has figured in many
accounts. Important origins in modern psychology include work on implicit learning
by Reber, which originally predated the cognitive revolution (see Reber, 1993, for an
overview of this research programme). While the cognitive revolution largely
caused a switch of focus from implicit to explicit memory systems (Frankish &
Evans, 2009), some authors such as Reber and others continued to emphasise the
contrast between the two, and the fact that some kinds of knowledge can be acquired
implicitly without any awareness of a learning process (see Cleeremans, 2015;
Cleeremans & Kuvaldina, 2019 for overviews of this field of work).
A number of the dichotomies associated with broad dual-process accounts are
shown in Table 4, but these are by no means exhaustive. We have used the labels
Type 1 and Type 2 here. The terms System 1 and System 2, introduced by Stanovich
(1999) became popular and are still used by some authors (e.g., Kahneman, 2011),
 42 Philosophy of Mind

Table 4 Dichotomies and features that have been associated with dual processes
in thinking by various authors

Type 1 Type 2
Intuitive Reflective
Non-conscious Conscious
Fast Slow
Automatic Controlled
Associative Rule based
Implicit knowledge Explicit knowledge
Contextualised Abstract
High capacity Low capacity
Independent of general intelligence Correlated with general intelligence
Independent of working memory Imposes load on working memory

but the notion of dual systems is problematic. A number of authors have suggested
that the two systems could have roots in evolutionary distinct systems, System 1
being associated with more animal-like cognition and System 2 more distinctively
human (e.g., Epstein, 1994; Evans & Over, 1996; Reber, 1993; Stanovich, 2004),
but this seems to us now to be an unhelpful way to use the word ‘system’. We now
prefer to describe such approaches as two minds theories, a concept explored in
detail by Evans (2010). System 1 cannot be a single system as there are multiple
kinds of Type 1 process (Evans & Stanovich, 2013), and there are also arguments
for multiple Type 2 systems (Evans, 2019). Hence, the Type 1 and 2 terminology
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

seems to be the clearest.

Theories of dual processes, systems, and minds raise a very large and complex set
of questions which we cannot possibly deal with in this short section. We will focus
here primality on dual-process accounts within the psychology of reasoning. First,
however, we draw attention to the distinction between unconscious and precon-
scious processes. There are very many wholly unconscious inferential processes in
the brain, such as those used for visual perception and language comprehension,
which are not part of our story here. Type 1 processes (see Table 4) are better
thought of as preconscious in that they post some final product into consciousness
(or working memory), typically a putative intuitive judgement accompanied by
a feeling of confidence (Thompson et al., 2011). Type 2 processes, in contrast, post
intermediate products into working memory so that we have some conscious sense
of the process of reasoning. It is the required use of a capacity-limited singular
working memory system (Baddeley, 2007, 2020) that makes Type 2 processes
relatively slow. Some theorists (e.g., Engle, 2002) also equate working memory
 Human Reasoning 43

with controlled attention, linking with the automatic-controlled dichotomy, popular

in dual-process writing.

6.1 Matching and Belief Biases: Origins of Modern

Dual-Process Theories
Peter Wason is often regarded as the founder of the modern psychology of
reasoning (Manktelow, 2021), and the second author was fortunate enough to
have had him as his PhD supervisor. Among other things, Wason was an
inventor of reasoning problems which have led to hundreds of published studies
in the literature. We have already met his 2–4–6 task (Wason, 1960) in
Section 4.1. His most famous problem, however, is the four-card selection
task, usually known just as the selection task, first introduced to an unsuspecting
audience in an early book chapter (Wason, 1966). It has since become the single
most cited and investigated task in the whole psychology of reasoning (Evans,
2022). The original problem used abstract materials, and a typical example is
shown in Figure 2. In essence, people are asked to test the truth of a conditional
statement about cards which have a letter on one side and a number on the other.
In the example shown, the claim is that if there is an A one side then there is a 3
on the other. In Wason’s original experiment four actual cards were used with
a human experimenter sitting across at a table, but in most later experiments,
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

Figure 2 Standard abstract version of the Wason selection task

 44 Philosophy of Mind

a diagram similar to that in Figure 2 is shown on a computer screen and cards

can be selected by pointing and clicking with a mouse. The important feature is
that each card has a visible facing side and a hidden side. All we know for sure is
that a letter has a number hidden on the back and vice versa. The visible values
are A, D, 3 and 7.
The task is not purely deductive as it involves hypothesis testing, but it does test
whether people understand that the statement could be disproved by finding a card
which has an A on one side, but a number other than a 3 on the other. Logically,
they could only discover this if the A card was turned over to display some other
number, or if the 7 card (an example of a number which is not a 3) turned out to
have an A on the back. Given that the task is stated to be only about the four
displayed cards, Wason and most (but not all) later researchers consider the correct
choice to be the A and the 7. The claim will be true for these four cards unless
proven false, and only turning the A and the 7 could do that. Many studies have
shown that participants, typically but not always undergraduate students, make this
choice only about 10 per cent of the time. As Wason discovered, most people select
either the A card alone, or the A and the 3. Choosing 3 could not disprove the
statement, of course which makes no claim that an A must be on the back of a 3.
The A could, but so could the 7, so both need to be chosen to comply with the
instructions. So elusive is the correct solution that Wason (1966) claimed that even
professors of logic were known to get it wrong!
Wason’s original explanation of the problem was a verification bias, better
known now as confirmation bias. (He maintained the view that he had demon-
strated such a bias with his 2–4–6 problem, a claim now seen as highly dubious,
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

as discussed in Section 4.) He believed that people were trying to prove the
statement true rather than false, and hence looking for the confirming case of
A and 3. In the original paper he also proposed that people have a ‘defective
truth table’ for the conditional (see Section 5), in which the statement was seen
as irrelevant to letters other than A. This explains why the D card is hardly ever
selected, but not why people often choose the 3 card, unless they take the
statement as a biconditional. The alternative account was matching bias. That
is, people chose A and 3 simply because they matched the content of the
conditional statement. In order to demonstrate this, negations have to be
introduced into the conditional. For example, suppose the problem is the same
as shown in Figure 2 except that the conditional statement is ‘If there is an A on
one side of the card, then there is not a 3 on the other side of the card’.
If cards are chosen to verify, then the participants should select the A and 7
(not the 3) cards, but if they are matching, they will continue to choose A and 3,
even though this is now the logically correct choice, as it can uncover
a falsifying case. The latter is what participants actually do, as originally
 Human Reasoning 45

demonstrated by Evans and Lynch (1973) and replicated many times since.
Good Popperian as he was, Wason immediately accepted that his verification
bias account must be wrong, he but was nevertheless puzzled by the finding.
The main reason for his curiosity was that he had recently shown that when
asked to provide a short verbal justification for their card choices, people
appeared to show understanding of the logic (Goodwin & Wason, 1972). That
is, they described their choices as verifying or falsifying the statement, consist-
ent with their card selections. As a result, Wason and Evans (1975) ran an
experiment combining the two methods. They used both the affirmative and
negative conditional forms, but also asked participants to provide a short,
written justification for each card choice. Participants performed both tasks,
half affirmative followed by negative and half vice versa. The results of this
simple and single experiment launched the dual-process theory in the psych-
ology of reasoning. (Their findings were never contested and were eventually
replicated in all essential details by Lucas and Ball, 2005.)
First, Wason and Evans replicated the matching bias account. Participants’
predominant choices were the equivalent of A and 3 on both versions of the task
(of course, actual lexical content was varied randomly). However, they also
replicated Goodwin and Wason, in that explanations offered were always
logically consistent with the card choices made. This meant that on the negative
tasks, people generally showed ‘insight’, saying that they were choosing A and
3 in order to prove the statement false, but on the usual affirmative versions they
continued to justify their choices as seeking a confirming combination of A and
3. This applied equally to those receiving the negative version first. So how
could a negative stimulate ‘insight’ which was then immediately lost when the
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

negative was removed?

The explanation offered by Wason and Evans was that participants’ card
choices were influenced by an unconscious Type 1 process (they actually used
the now modern terms Type 1 and 2) but that the verbal justifications were
generated by a second, separate process of Type 2 reasoning. In effect, they
were rationalisations. In a second study, (Evans & Wason, 1976) showed that
participants would provide a rationalisation for any of several common patterns
of card choices that were implied to be the correct choice by the experimenters,
even though only one was. They gave a justification of verification and/or
falsification with high confidence, regardless of the ‘solution’ being justified.
Hence, the original form of the dual-process theory of reasoning was that of
preconscious intuition followed by conscious rationalisation. This is important,
because later versions mutated into a different account in which Type 1 and
Type 2 processes compete for control of the response, with the suggestion that
Type 2 reasoning was required for correct logical reasoning. Evans, (2019) has
 46 Philosophy of Mind

discussed in detail the origin and nature of these two forms of dual-process
theory and their potential for reconciliation. The original rationalisation form of
the theory is still preferred by contemporary authors who believe that reasoning
evolved for the purpose of argumentation rather than problem-solving (Mercier
& Sperber, 2011, 2017).
The second form, which we might call the conflict theory, was stimulated by
the study of belief bias in syllogistic reasoning already described (Sections 2.3
and 2.4; Evans et al., 1983). In view of our discussion of the new paradigm, the
reader may wonder if it is still right to refer to this effect as a cognitive ‘bias’. It
is now held to be permissible, and even obligatory, for people to take account of
their prior beliefs in reasoning, as long as this is consistent with Bayesian
principles (Section 3). But the term ‘belief bias’ persists in the contemporary
literature and strictly speaking it is a bias when observed in experiments with
instructions requiring participants to disregard prior belief. It also turns out that
the ability to comply with such instructions – or not – is a key factor in dual-
process accounts. These are the main reasons ‘belief bias’ continues to be used
in some contemporary studies. In deference to the new paradigm, however, we
will refer to the belief effect in this section, avoiding the term ‘bias’.
The reader may recall that in the study of Evans et al. (1983) participants were
influenced by both the logical validity of the conclusions offered and their prior
believability. Also, this belief effect was stronger for invalid than valid argu-
ments. One account, favoured by the original authors as well as later in a mental
model theory (Oakhill et al., 1989), is known as selective scrutiny. That is,
participants are more motivated to examine the logical basis for unbelievable
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

conclusions, or to seek counterexample mental models which demonstrate their

invalidity. This means that reasoning is motivated by disbelief in line with
cognitive miser principle (Stanovich, 2018). Why bother challenging the evi-
dence for an assertion you already believe? This explains why the belief effect is
stronger for invalid arguments, but not why it is still significantly present for
valid arguments. Nevertheless, the belief effect studies stimulated the view that
belief-based intuitions may compete with logical reasoning, a conclusion sup-
ported by some experimental findings. For example, the belief effect is stronger
and logical accuracy weaker when very short time limits are given (Evans &
Curtis-Holmes, 2005), presumably because there is insufficient time for Type 2
reasoning.
These positions have led to some detailed debate and investigation in the
recent psychological literature (see Evans et al., 2022, pp. 161–164). We do not
have space to review these studies here but in essence there is a distinction being
drawn between a Type 1 belief effect (across the board tendency towards
believable conclusions) and a Type 2 belief effect, which is motivated reasoning
 Human Reasoning 47

selectively applied when conclusions are unbelievable. An important recent

contribution has been the idea that on reasoning and judgement tasks, a quick
intuitive answer comes to mind with a degree of confidence known as feeling of
rightness (FOR) (Thompson et al., 2011). The evidence shows that people are
more likely to rethink or change an intuitive answer if it has low FOR, which is
a kind of selective scrutiny. Correspondingly, participants have been shown to
have higher FOR for believable than unbelievable conclusions (Thompson
et al., 2011) and also for matching than mismatching cards on the selection
task (Thompson et al., 2013). There has also been a good deal of interest in the
Cognitive Reflection Test (Frederick, 2005; Toplak et al., 2011), a set of simple
looking problems with intuitively compelling answers that are actually wrong.
Such tasks are often failed by those of high intelligence who do not look beyond
intuition. Intuitive confidence and selective processing accord with the ‘cogni-
tive miser’ hypothesis, which postulates that in a world of massive information
overload we evolved to use our central cognitive resources as selectively as
possible (Stanovich, 2018). Intuitive confidence can, however, easily be mis-
placed in the modern world.

6.2 Dual Processes and Individual Differences

Individual differences in cognitive ability have been studied by psychologists
for over a century, with Spearman (1904) presenting the first theory of general
intelligence – also known as just g – which led to a major field of study and the
later invention of IQ testing. For a history, and review of recent research on
intelligence see Deary (2020). In essence, there is strong evidence that there is
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

indeed a general factor of intelligence, largely hereditary, that correlates with

academic attainment and all manner of cognitive tasks, especially those involv-
ing reasoning and calculation.
Dual-process theory is mostly studied by experimental means. It is widely
accepted that Type 2 processing (or explicit reasoning) requires use of short-term
and capacity-limited working memory store (Evans & Stanovich, 2013). Working
memory has been subject to an enormous number of studies in the memory
literature, largely separated from the study of reasoning and decision making (see
Baddeley, 2007, 2020, for the history of this research programme). However,
cognitive psychologists studying reasoning are well aware of this work and its
methods. One can try to inhibit Type 2 reasoning, for example, by asking people to
reason with concurrent working memory loads, or by instruction to respond
immediately without thought. However, one can also apply a psychometric
approach, arguing that tasks that require Type 2 reasoning should be correlated
with individual differences in general intelligence. This is the basis of a long-
 48 Philosophy of Mind

standing research programme by Keith Stanovich and Ricard West from the 1990s
onwards (Stanovich, 1999, 2011; Stanovich & West, 2000). We should also note
that there have now been many studies of individual differences in measured
working-memory capacity (WMC), which also correlated with performance on
a huge range of cognitive tasks (see Baddeley, 2007). These measures are them-
selves highly correlated with general intelligence, although the size of the correl-
ation is disputed. At the very least, however, we can say that those high in general
intelligence are likely to have higher working memory capacity and are therefore
also likely to be better at Type 2 reasoning.
Stanovich and West have studied a large range of reasoning and decision
tasks drawn from cognitive and social psychology usually using SAT scores as
a surrogate for general intelligence (SAT is, in psychometric speak, highly
g-loaded). Among many other similar findings, they have shown that perform-
ance on the abstract Wason selection task is related to general intelligence
(Stanovich & West, 1998) as in the ability to resist the influence of belief in
syllogistic reasoning (Stanovich et al., 1999). In a recent huge scale and
comprehensive study, they showed that general intelligence is implicated in
the vast majority of known cognitive tasks on which well-established errors and
cognitive biases are demonstrated (Stanovich et al., 2016). However, there is
one important exception: myside bias. This is the tendency to value and propose
only arguments which accord with one’s core values and worldview. Myside
bias is equally marked in those of higher intelligence, education level and
rational thinking style. This exceptionality is so important that Stanovich has
recently written a whole book about myside thinking (Stanovich, 2021). We
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

discuss myside bias in Section 7.

From the beginning, Stanovich and West have argued that, although
general intelligence facilitates Type 2 reasoning and the avoidance of cog-
nitive biases, it is not the only basis for rational thinking. Indeed, Stanovich
has written another book on what he sees as the fallacy that measured
intelligence equates with rationality (Stanovich, 2009). In all of their stud-
ies, they show that when the variance due to intelligence is removed, there
remains a second important predictor, which is self-reported rational think-
ing style. In essence, it is not enough to have a high IQ to be smart, you must
also have the disposition and motivation to apply Type 2 reasoning to the
problem in hand. People high in rational thinking style do not rely on
intuition when solving novel or difficult problems but check their potential
solutions out with explicit reasoning. It is well known that when problems
solicit particularly plausible but incorrect solutions, even those of high
intelligence will often fail to solve them. We will return to Stanovich’s
writing about rationality in Section 7.
 Human Reasoning 49

6.3 Debate and Controversy about Dual-Process Theory

In hindsight, both Keith Stanovich and Jonathan Evans realised that they inad-
vertently contributed to the creation of myth in dual-process theory, which we
term the normative fallacy. The fallacy is that Type 1 processes are responsible for
cognitive biases and errors and Type 2 processes generate normatively correct
answers. In the most extreme application of the fallacy, some authors write as
though one could diagnose the type of processing underlying a judgement or
inference from the normative accuracy of the answer given.
The normative fallacy must be wrong for a number of reasons. Type 1
processes certainly can lead to cognitive biases, especially with abstract and
novel laboratory tasks such as the Cognitive Reflection Test. Moreover, our
subjective feelings of confidence are far from reliable and can be attached to
wrong as well as right answers (Thompson et al., 2011). However, experts with
experience and others can develop accurate intuitions which often allow them to
respond rapidly and appropriately to real world situations, as a number of
authors have argued (Gigerenzer, 2007; Gladwell, 2005; Klein, 1998). Much
of this helpful intuition takes the form of rapid pattern recognition. The evi-
dence for such accurate intuitions is beyond dispute, so Type 1 intuitive
processes alone cannot be identified with cognitive biases.
Equally, Type 2 processing does not and cannot guarantee correct responding
(see Elqayam & Evans, 2011, and Section 7). Type 2 reasoning can fail for
a number of reasons. People may have insufficient cognitive capacity for the
problem or lack relevant knowledge or ‘mindware’. They may also try to
simplify their reasoning by focusing on only a single possibility, when more
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

are relevant, as described by the singularity principle (Section 3.4) and consist-
ent with the cognitive miser hypothesis (Stanovich, 2018). All of these things
have been pointed out repeatedly by both Evans and Stanovich (e.g., Evans,
2007b; Stanovich, 2011). So how did these authors contribute to the fallacy?
The answer is by focussing much dual-process research on particular laboratory
tasks that are novel, abstract, and difficult, and to which real world experience is
hard to apply. Deductive reasoning can be difficult, for example, and prior
knowledge is biasing by definition in the old, traditional paradigm (see
Section 2). Without logical training, people will frequently commit fallacies,
and without high general intelligence, they will find it hard to follow instruc-
tions to assume premises and disregard their prior beliefs (see Section 2). These
laboratory tasks are selected for dual-process research precisely because Type 2
reasoning is necessary (but not sufficient) for their solution, so that we can make
it more or less easy to apply, for example by the use of time limits, instructional
sets or working memory loads.
 50 Philosophy of Mind

Misunderstanding of dual-process theory goes much deeper than the nor-

mative fallacy, however. There has been a tendency for critics to assume that
the typical correlated features of the two types of processing shown in Table 4
are all necessary and defining features. Such a position is easily refutable, of
course, but has been assumed by a number of leading critics (e.g., Keren &
Schul, 2009; Kruglanski & Gigerenzer, 2011). However, to our knowledge, no
particular dual-process theorist has proposed this perfect alignment of attri-
butes. This is part of what Evans and Stanovich (2013) call the ‘received’
version of dual-process theory, a generic and simplified version which is
attacked by critics while being proposed by no author in particular. Other
criticisms considered by Evans and Stanovich include vague and multiple
definitions, the idea that processing varies on a continuum rather than
a dichotomy (Kruglanski & Gigerenzer, 2011; Osman, 2004), and a lack of
convincing empirical evidence. Evans and Stanovich offer rebuttals of all
these arguments and particularly the last.
An important distinction between dual-process theories lies in whether they
propose a parallel or serial architecture. The two main types are known as
parallel-competitive and default-interventionist (Evans, 2007b). In the parallel
form (Sloman, 1996; Smith & DeCoster, 1999), what they term ‘rule-based’
(Type 2) and ‘associative’ (Type 1) processes proceed in a parallel and the
individual may become aware of a conflict between the two. More commonly
advanced, however, is the default-interventionist form, favoured by Evans and
Stanovich and other leading theorists (e.g., Kahneman, 2011). In this version,
rapid Type 1 processes generate a default intuitive response which may or may
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

not be overridden by subsequent, slower Type 2 reasoning. The tendency to

override can be influenced by motivation provided by either the reasoner
(rational thinking style) or the intuition itself (feeling of rightness).
Intervention which successfully avoids a cognitive bias also depends upon
having sufficient time and cognitive resources available and on the possession
of prior relevant knowledge or ‘mindware’ (Stanovich, 2011, 2018). Evans
(2019) has proposed that Type 1 intuitions are always scrutinised to some extent
by Type 2 processes, but the latter often serve only to rationalise the intuition.
The scrutiny can, however, lead to a rethink in which Type 2 reasoning substi-
tutes an alternative response. The complexity of these accounts exposes the
simplistic nature of the normative fallacy.
In recent years, there have been questions for standard dual-processing
accounts raised by some reasoning researchers, who have suggested that
people might have logical intuitions (De Neys, 2012, 2022). On a variety of
problems, people have been shown to detect the normatively correct answer
very rapidly, as evidenced by a variety of cognitive and neuroscientific
 Human Reasoning 51

measures. It has also been claimed that the logic of a reasoning problem can
interfere with judgements to decide the believability of a conclusion (Handley
et al., 2011), in a reversal of the evidence normally claimed to show that
‘belief bias’ research supports dual-process theory. There are also papers
showing that people of higher intelligence may have more accurate intuitions,
which cannot be attributed to better Type 2 reasoning ability (Thompson et al.,
2018).
In commenting on these studies (Evans, 2018, 2019) has pointed to the very
simple nature of the tasks used in these experiments, in which the normative
response could easily be produced rapidly. For example, Handley et al.’s
findings showed that it was more difficult to judge a conclusion unbelievable
when it was derived by a Modus Ponens inference. But MP is arguably
provided automatically by the language module (Braine & O’Brien, 1991).
To infer that a correct answer necessarily results from Type 2 reasoning is, of
course, the normative fallacy. A recent paper by Ghasemi et al. (2022)
confirms this interpretation of the studies (see also Ghasemi et al., 2023).
They showed that intuitive inferences can be rapidly generated from structural
features (as well as beliefs), but that such inferences may or may not be
logically accurate. For example, the conclusion of AC has ‘pseudovalidity’
such that it will interfere with conflicting belief judgements, in just the same
way as occurs for the actually valid MP inference. Intuitions flow from
structure as well as beliefs, but of course we already knew that from examples
such as atmosphere effects in syllogistic reasoning (Section 2.2) and matching
bias on the selection task.
In conclusion, we agree with critics that the ‘received’ version of dual-
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

process theory is wrong. There is no Type 2 process which is invariably slow

and normatively correct, for example. Type 1 intuitions need not be belief
based and can be accurate. However, neither we nor other authors have
proposed such a theory. The dual-process theory of reasoning has evolved
into a much more nuanced and complex account as a mass of evidence has
been collected. Simple generalisations about the two types of processing are
not sustainable, and some would argue that this makes it hard to falsify the
theory. However, it is important to understand the framework which has
motivated so much research over the past forty years or so. Type 2 reasoning
corresponds to the kinds of reasoning normally discussed in philosophy, but
psychologists have shown that this reasoning is also influenced by rapid
preconscious processes, and that a large number of factors can affect its
conclusions, including those related to the presentation of the reasoning
problem and its context, and the individual nature of the person doing the
reasoning.
 52 Philosophy of Mind

7 Rationality and Reasoning

Experimental studies of reasoning, as well as of judgement and decision-
making (JDM, not generally covered in this Element), have evoked a ‘great
rationality debate’ (Stanovich, 2011). This debate in psychology was started by
a philosopher, L. Jonathan Cohen (1981), and many philosophers and psych-
ologists have contributed to it. Psychologists of reasoning and JDM cannot
escape discussion of rationality owing to their common use of normative
theories to assess error and bias. We saw in Section 2 that the deductive
paradigm, based on classical logic, was eventually challenged and modified
by the new paradigm, which we have discussed at length in this Element.
Essentially, the inability of participants to comply with classical reasoning
from assumptions, and with the material conditional analysis of the natural
language conditional, forced psychologists either to declare people inherently
irrational, or else to question the normative accounts that judged them to be in
error. As we have shown, most took the latter path. In this section, we will
explore several aspects of rationality and human reasoning. But we will begin
with a brief historical review of the rationality debate as it developed.

7.1 The Great Rationality Debate

Evans (2021) has recently reviewed the history of the rationality debate in
detail, and we will give only a brief summary here. As we have already
shown, Peter Wason was publishing studies of reasoning from the 1960s
onwards, leading to a widely read and influential book summarising much of
this work and its conclusions (Wason & Johnson-Laird, 1972). Wason’s claims
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

of confirmation bias (later disputed) and other evidence of logical error and bias
led him directly to conclusions of irrationality, writing later, ‘It could be argued
that irrationality rather than rationality is the norm. People all too readily
succumb to logical fallacies’ (Wason, 1983). From the early 1970s onwards,
the revolutionary research of Daniel Kahneman and Amos Tversky had massive
influence on the study of judgement and decision making. JDM researchers in
the 1960s thought that human decision-making was fairly rational, provided that
probabilities were allowed to be subjective, but Kahneman and Tversky dem-
onstrated a whole range of cognitive biases in a series of highly influential
articles.
Their general thesis was that people judge probabilities by applying heur-
istics such as those based on representativeness (Section 5.3) and availability
(Tversky & Kahneman, 1974). Although often useful, these heuristics could
lead to biases. For example, the availability heuristic is used when we judge
frequency by the ease with which examples can be brought to mind. Thus,
 Human Reasoning 53

doctors when confronted with a patient’s symptoms might recall similar cases in
their experience and use these to determine a probable diagnosis. However, as
Kahneman and Tversky demonstrated, this heuristic can often lead to biases due
to how our memories work. If shown a list of names that includes famous as
well as unknown people, for example, participants will overestimate the fre-
quency of the famous ones as they are easier to recall. Doctors could also be
biased if they recalled a patient with similar symptoms but an unusual diagnosis
who was especially memorable, say, as a close family friend. Following the
early papers of Kahneman and Tversky, a very large field of research built up
demonstrating many judgemental biases both in the laboratory and in real world
settings (Gilovich et al., 2002; Kahneman et al., 1982).
Kahneman and Tversky were more cautious than Wason about attributing
irrationality to their participants, but many of their followers were less
restrained. Cohen (1981) read widely in the psychology of reasoning and
JDM and came to the conclusion that psychologists were overstating their
case. In fact, he claimed that irrationality could never be demonstrated by
such experiments, a position later dubbed Panglossian (Stanovich, 2011). As
the second author has observed (e.g., Evans, 2021) there were three main
arguments in Cohen’s paper which Evans terms:

1. The normative system problem

2. The interpretation problem
3. The external validity problem

The normative system problem is that psychologists were wont to adopt

https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

standard normative theories, such as classical logic, whereas philosophers

and others had provided many alternative normative accounts of reasoning
and decision-making. One should not therefore simply judge people right or
wrong by uncritically presupposing some norm. Of course, Cohen was right
about this and the problem of alternative norms for psychological research
was discussed in detail more recently by Elqayam and Evans (2011). We
have already shown in this Element how the new paradigm psychology of
reasoning emerged due to dissatisfaction with classical logic as a normative
account of deduction. We should also note the claim of Stanovich (e.g., 2011)
that it is compliance with classical norms of logic that generally correlates
with intelligence and rationality thinking measures, while also recalling that
there is considerable overlap in the norms of classical logic and probability
logic. For example, as we have pointed out above, in Section 2.4, MP and MT
are both valid in classical logic and p-valid (Section 3.2) in probability logic,
and AC and DA are both invalid in classical logic and p-invalid in probability
logic.
 54 Philosophy of Mind

The interpretation problem is that the participants may not understand the
problem in the way in which the experimenter intended. Take as an example,
the conjunction fallacy described in Sections 3.2 and 4.3 (Tversky &
Kahneman, 1983). Given a stereotype suggesting that Linda is a radical
young woman, people may judge that ‘Linda is a feminist and a bank teller’
to be more probable than ‘Linda is a bank teller’. This judgement might seem
to be logically incoherent and a fallacy, but when the two sentences are
included in the same list, it is possible that participants take the second to
mean implicitly ‘Linda is a bank teller who is not a feminist’, in which case
there is no fallacy. The interpretation problem is real and must be attended to
by the experimentalists, but it cannot plausibly explain away the vast
amounts of evidence for cognitive biases. The weakest of Cohen’s claims,
in our view, is the external validity problem: that psychological experiments
are unrepresentative of the real world. They can be artificial, of course, but
there are very many reported experiments that are carefully controlled
pointing to clear evidence of cognitive biases. There have also been many
demonstrations that biases demonstrated in the laboratory occur in real world
contexts (see a number of the papers in the collection edited by Gilovich
et al., 2002).
How far psychologists of reasoning should make normative evaluations of
people’s reasoning is a debatable question in psychological research, as we
first discussed many years ago (Evans & Over, 1996). However, when
Elqayam and Evans (2011) went so far as to propose a descriptivist approach
in which we would essentially do away with normative evaluations in the
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

psychology of reasoning, few of the commentators agreed with this, psych-

ologists and philosophers alike. But since Cohen, psychologists have felt
obliged to comment on rationality in a less simplistic manner. We explore
some of the many ideas about rationality that have arisen in the remainder of
this section.

7.2 Rationality and Dual Processing

Dual-process theory, the focus of Section 6, has been linked in various ways
with human rationality. Consider the basic distinction between epistemic and
instrumental rationality (Kolodny & Brunero, 2023). Organisms have instru-
mental rationality to the extent that they can reliably achieve their individual
goals. Epistemic rationality is the acquisition and maintenance of well-
supported beliefs. Epistemic rationality was long thought of as a human char-
acteristic, but more recent studies of non-human animal cognition cast some
doubts on this tradition. The notion that Type 2 or System 2 thinking is uniquely
 Human Reasoning 55

human, which has been suggested by various authors (e.g., Evans & Over, 1996;
Reber, 1993; Stanovich, 2004) is not strictly correct as pointed out by the
biologist Toates (2004, 2006). Higher animals have working memory and two
different kinds of learning, one based on association and the other on encoding
of individual, episodic events (Schacter, 1987). We think it more accurate to say
that the second system (which does Type 2 processing) became uniquely
developed in human beings, due the evolution of greatly enlarged frontal
lobes, high working memory capacity, and capacities for language and suppos-
itional reasoning (Evans, 2010).
We refer to this development as the new mind, which was added to the old
mind which still shares many features with the minds of other animals, includ-
ing basic emotions and associative learning systems. Humans consequently
have a unique ability to reason about hypothetical possibilities. Stanovich
(2011) has made a similar distinction between an autonomous and a reflective
mind. Our instrumental rationality derives partially from associative learning
systems in the old mind that we share with other animals, as well as innate
cognitive modules underlying, for example, the visual system and language
processing which operate entirely outside of consciousness. But it also depends
on the epistemic rationality of the new mind with its highly developed capacity
for acquiring, storing, retrieving, and reasoning about explicit knowledge and
belief.
Evans (2013) has written about the differences in rationality of the two
minds. Old mind rationality in humans and other animals is driven by the
past. We evolved cognitive modules by adaptation to past environments, and
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

we learn within our lifetimes to adapt our behaviour to the current environment.
Much of this learning is Skinnerian: we repeat activities which have been
reinforced in the past and avoid those which were punishing. This learning
process can lead to helpful habits as well as destructive addictions, for example,
pathological gambling, and phobias. New mind rationality is facilitated by
hypothetical thinking or ‘cognitive decoupling’ (Evans & Stanovich, 2013)
meaning that it can ignore the current context, and imagine scenarios not
grounded in current beliefs about the actual state of the world. Such thinking
can engage with both counterfactual possibilities and yet to be determined
future possibilities. Hence we can, in the manner envisaged by decision theory,
anticipate the future consequences of our actions and change our behaviour
accordingly. But when the two minds are in conflict, the old often wins (Evans,
2010). For example, we go for short-term gratification and not the future benefit
we prefer at a higher level. Yet, it is the new mind that makes us distinctively
human and able to do many things unknown elsewhere in the animal kingdom,
for example in the arts and sciences. To do these things we must maintain, at
 56 Philosophy of Mind

least to some extent, logically coherent belief systems and reason with them
well enough for our purposes.
Keith Stanovich (e.g., Stanovich, 2004, 2011, 2018) has written extensively
about rationality within a dual-process framework. His autonomous mind con-
tains a mixture of innate cognitive modules, implicit knowledge acquired by
associative learning, and knowledge which was once explicit and has become
automated by practice. He has argued, intriguingly, that other animals are often
more instrumentally rational than humans, at least as conventionally measured by
normative decision theory (Stanovich, 2013). This is because animals always
pursue immediate goals, while humans have layers of complexity of higher-order
values and conflicting goals in their reflective minds. A good example of a human
bias is sunk costs. Imagine you bought expensive opera tickets months ago but
when the night comes you are exhausted, and your favourite football team is on
TV. Right now, you would rather stay home and watch the game than go to the
opera. In normative decision theory, that is what you should do because the cost of
the tickets is sunk, the same whether you go out or stay in. But most people would
nevertheless go to the opera because they would not like the thought of ‘wasting’
money, or perhaps because they like to think they are sophisticated people who
prefer opera to football. Of course, we could argue that their choice is rational in
a sense not captured by the instrumental rationality of pursuing immediate goals.
Stanovich has argued that normative problem-solving and decision-making
depends on a number of factors. You must first recognise that the problem in
front of you cannot be solved by habit or pattern recognition. Then you must
apply the necessary cognitive effort, you must have sufficient cognitive
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

capacity to solve the problem and you must have relevant mindware, or
knowledge of how to solve it. Mindware is a key concept (Stanovich, 2018):
rationality depends on education and training as well. Other authors have, by
contrast, praised the role of intuition and ‘gut feelings’ in rational decision
making (e.g., Gigerenzer, 2007; Gladwell, 2005). Both perspectives have their
merits. Intuitions can save lives (see Klein, 1998, for examples) and pattern
recognition, derived from experiential learning, has a key role in expert
problem-solving. But Stanovich emphasises that in a modern technological
world, reliance on intuitions can be dangerous and a rational thinking style
highly advantageous. Our cognitive miserliness evolved in a very different
world from that in which we live now.

7.3 Bayesian Reasoning and Instrumental Rationality

It is impossible to check our beliefs, above the level of very restricted and
limited ones, for full logical coherence, as defined by normative Bayesianism.
 Human Reasoning 57

It is computationally too difficult, and sometimes logically impossible, to

ensure coherence and classical consistency in our beliefs in general. The
singularity principle (Evans, 2007a), which posits that people tend to focus on
one hypothesis at a time, can give them local coherence in immediate reasoning,
allowing them to reach some goals and have instrumental rationality to this
extent, despite incoherence or inconsistency with other beliefs they hold. It is
equally true that full logical coherence is too weak as a condition to guarantee
that we achieve our real-world goals. Some people with severe clinical delu-
sions, for example, of a paranoid nature, could be coherent in their subjective
delusional beliefs and yet be failing miserably to achieve their goals of having
a happy and fulfilling life. Thus, full logical coherence is neither necessary nor
sufficient for instrumental rationality.
There is much more, however, to Bayesianism than being coherent, and it is
far more difficult to specify exactly how it as a whole is related to instrumental
rationality. In Section 3.4, we saw that Bayesian conditionalisation is
a normative proposal about how we should update or revise our beliefs. There
are arguments and even proofs, under certain assumptions, that we will come to
agree in our beliefs if we use strict Bayesian conditionalisation to update them.
In 3.4 we used the example of trying to find out whether a certain coin was
double-headed or fair. Some of us might start out by assigning an equal prior
probability of .5 to these two possibilities, while other people judge that the coin
is quite likely to be fair and assign a probability of .75 to that. But now if the coin
is repeatedly spun and comes up heads again and again, and we all follow the
procedure of strict conditionalisation, agreeing that invariance holds, and
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

remaining coherent in our beliefs, then we will all get closer and closer to
agreeing that the coin is double-headed. Our separate priors will be ‘washed
out’, and there will be convergence in our beliefs. There are even arguments that
such convergence is what is meant by scientific objectivity or truth (see
Sprenger & Hartmann, 2019, and the references given there for these conver-
gence arguments and proofs and critical evaluations of them).
Well-confirmed scientific hypotheses and theories, by Bayesian belief
updating, can obviously be used to achieve many goals, some of which have
the greatest importance for our survival as individuals and a species. Even so,
not everyone has the goal of scientific truth in the first place. Scientists should
have it but sometimes do not. Fanatics and propagandists can be indifferent or
even hostile to it. Perhaps propagandists have to use some Bayesian belief
updating to find the best ways of persuading people of some point of view.
But fanatics could simply refuse to acknowledge the existence of scientific
evidence. They might, for example, refuse to accept the word of scientific
experts and appeal to their own ‘experts’, as defined by their myside bias.
 58 Philosophy of Mind

Social media algorithms also have the effect of encouraging users to access only
the opinions of others that accord with their own, with or without probative
evidence.
Stanovich (2021) considers at length whether myside bias is really a bias at
all and whether it should be considered to be irrational. As mentioned earlier
(Section 6.2), what sets myside bias measures apart from all other cognitive
biases is that they are not related to general intelligence or rational thinking
measures. The extent of the ‘bias’ is actually predicted by something different:
the strength of the core belief which is influencing the thinking. Referring to the
Bayesian framework, Stanovich points out that the diagnosticity (Section 3.5)
of the evidence should be assessed independently of prior belief, before the two
are combined to conform to Bayes’ theorem (Section 3.4). But whether or not
Bayesian procedures are properly followed, prior beliefs can vary widely
between people and have a major impact on their reasoning.
Many psychologists, for example, would reject apparently strong statis-
tical evidence for extra-sensory perception due to a firm prior belief that ESP
is intrinsically most unlikely. Myside bias is particularly evident for topics
on which there is a deep binary division of opinion, such as the case for the
UK leaving or remaining in the European Union in 2016. In such cases,
Stanovich notes that each side often holds that the other must be less
educated or intelligent, or gullible in accepting weak or unsubstantiated
arguments. But the evidence can suggest that they are not: they may simply
have differing beliefs and values. Myside bias is not about factual beliefs.
Compare our personal beliefs that (a) the UK has left the EU and (b) that the
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

UK should one day re-join the EU. Only the former is verifiable against
objective evidence. There are no clear objective grounds for establishing the
‘should’ in (b).
Rationality can reside in the beliefs people hold rather than the reasoning they
do. Consider again people with actual clinical delusions such as those who have
paranoia (Rhodes et al., 2020). They may ‘hear’ voices, and their new mind,
Type 2 processing may have the goal, which only a human being could have, of
explaining these occurrences. The result could be the inference that they are
being spied on, giving them the satisfaction of goal achievement and to that
extent instrumental rationality. Experiments have found that people with delu-
sions can tend to jump to conclusions given limited evidence (see Rhodes et al.,
2020, for points about these results), but in some instances, the inference to the
explanation of being spied on could satisfy Bayesian principles, with the
problem simply being hallucinatory premises. As we have already noted,
paranoid delusions may prevent some people from achieving many life goals,
but that might not hold for other individuals if their paranoia is not very severe.
 Human Reasoning 59

How far people with even clinical delusions fail to achieve instrumental ration-
ality, caused by violating Bayesian or other normative principles, has to be
investigated on an individual basis.
Some limited conformity to logical and Bayesian principles is a necessary
condition for us to have hypothetical thoughts and Type 2 processing to begin
with, before these are assessed for their contribution to instrumental rationality.
As we reported in Section 3.2, people do have an above chance tendency to
conform to p-validity and be coherent, and MP has an especially high endorse-
ment rate. The existence of hypothetical thought most clearly depends on some
correct use of MP in particular, even though that that might sometimes take
place when there is not full logical coherence.

7.4 Conclusions
Psychologists and philosophers alike have been engaged in a debate about
the implications for human rationality of experimental research on human
reasoning and judgement and decision-making. The frequently reported
conflicts between human responses and presupposed normative systems
initially led to concerns that humans might be fundamentally irrational.
However, the past fifty years or so that debate has become much more
complex and nuanced, in part triggered by the contribution of philosophers
(Cohen, 1981), but also by a growing conviction among psychologists that
humans could not have achieved what they have if they were irrational. This
has led to detailed scrutiny of presupposed normative systems, as evidenced
by the new paradigm psychology of reasoning. It is now more common to
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

consider Bayesian principles, rather than classical logic, as yardsticks for

normative reasoning. But this general tendency leaves many questions open.
Good (1983) characterised 46,656 different varieties of normative
Bayesians. Elqayam and Evans (2013) distinguished between strict and
(their preferred) soft Bayesian approaches in the scientific study of human
reasoning, and there is not yet wide agreement in the psychology of reason-
ing on the best specific approach to take.
Whether our reasoning has instrumental rationality in any given case can
sometimes be extremely difficult to determine, partly due to the complexity of
the human brain/mind. In some respects, people act like other animals, learning
by experience what is good and bad for them and achieving instrumental
rationality in this way. But the evolved addition of the distinctively human
‘new mind’ adds layers of complexity in which people can have multiple goals,
sometimes in conflict with one another and reflecting uniquely human value
systems. That is why it can be so hard to decide if someone’s myside thinking is
 60 Philosophy of Mind

a cognitive bias. A key aim in the new mind is the attempt to maintain at least
a locally coherent belief system, and it appears that cannot be achieved without
creating a cognitive frame which encourages myside thinking to some extent.
On the other hand, it is clear that myside thinking can undermine instrumental
rationality, for example, when people act on the advice of ‘experts’ defined by
their myside thinking instead of experts more reliably identified (Section 4.2) or
seek out only the views of those who already agree with them, a process
facilitated in the modern age by social media.

8 Overall Conclusions
Psychologists have been studying human reasoning intensively since the 1960s
with much use of the traditional deduction paradigm for the first thirty to forty
years that period. This method instructed people to ignore prior beliefs and draw
only conclusions that necessarily followed from assumed premises, and it used
standard binary logic as a normative system for deciding accuracy. The review
of hundreds of such studies in Evans et al. (1993) made it clear that humans are
poor reasoners by those standards, even those of higher intelligence and educa-
tion levels. First, they make many logical errors, especially by endorsing
inferences with conclusions that do not necessarily follow given the premises.
Second, they are systematically biased by irrelevant linguistic and structural
features of the arguments presented. Finally, people have great difficulty in
ignoring the prior beliefs that are stimulated by the content and context of the
problems presented. The last finding is strikingly the case for the those of
measured lower cognitive ability, even within university student populations.
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

These kinds of results caused a major debate about the rationality of human
reasoning. Philosophers and psychologists were both engaged, and they pre-
sented a paradox to reasoning researchers. How could human have achieved
their astonishing advances in science, engineering, and the arts, which so clearly
set them apart from other animals, if they were fundamentally irrational? The
conclusion eventually reached by most psychologists was that the problem lies
in the presupposed normative system: binary logic is simply not a good yard-
stick for rational reasoning in the real world. Human reasoning is in fact
inherently belief based and probabilistic. This conclusion led in turn to what
is known as the new paradigm in the psychology of reasoning.
The new paradigm recognises that human reasoning usually consists of
drawing probable conclusions from degrees of belief as premises, or from
some contextually relevant hypothetical possibilities, and is often directed
towards revising or updating degrees of belief. It aims to account for this
reasoning using subjective probability theory and the Bayesian account of belief
 Human Reasoning 61

updating. It is supported by experimental results broadly confirming that people

judge the probability of a natural language conditional, P(if p then q), to be the
conditional probability of q given p, P(q|p). The introduction of Bayesian
accounts of reasoning into the psychology of reasoning has yielded many new
results in the field. There is some tendency for people to be coherent and to
conform to probabilistic validity, p-validity, in their reasoning, especially when
the inferences are explicitly presented as listed premises with a conclusion. But
they are still not perfect, and sometimes their reasoning is fallacious, and their
judgements biased.
The broader questions associated with human reasoning discussed over the past
fifty years or so have concerned dual-process theory and rationality, each given
a section at the end of this Element. Dual-process theory has ancient roots in
philosophy and has also been manifested in many independent versions throughout
cognitive and social psychology (Frankish & Evans, 2009). The version that
developed within the psychology of reasoning started out largely as an attempt to
explain the frequent observation of cognitive fallacies and biases, with the idea that
unconscious Type 1 processes responsible for biases somehow compete with or pre-
empt more rational Type 2 processes. The second author, who helped to develop
these early accounts, came to realise later that there is a much broader family of such
theories be found in diverse fields, including the studies of cognitive social psych-
ology, implicit learning, decision making, attention and working memory and other
fields. A possible explanation of many findings is the evolution of a new mind,
distinctively developed in human beings for Type 2 processing, which is added to
and may compete with an older mind similar to that found in other higher animals.
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

At the end of this Element, we focused on rationality and studies of human

reasoning, a topic which fascinates philosophers and psychologists alike. It has
been debated extensively since the 1980s. The first major challenge came from
Cohen (1981), who pointed out, as a philosopher, that psychologists were relying
on classical logic and not considering other logics. He also questioned the validity
of laboratory experiments on several grounds. His arguments certainly helped
stimulate the dissatisfaction with the traditional methods which later resulted in
the new paradigm. The development of dual-process and two-minds theories also
complicated the debate about rationality, which is really much more complex than
the search for the right normative system. The human mind is multi-layered, with
both specialised and general inferential systems, and is capable of having many
goals and values which may conflict with one another. It needs to strive for
coherence to avoid chaos, and in the process manifests myside thinking. All of
these factors can compromise instrumental rationality, defined as the successful
pursuit of goals. But for all its flaws, the ability of the human mind to reason
effectively in many domains is an extraordinary achievement.
 References
Adams, E. (1965). The logic of conditionals. Inquiry, 8, 166–197.
Adams, E. (1998). A primer of probability logic. Stanford: CLSI.
Baddeley, A. (2007). Working memory, thought and action. Oxford: Oxford
University Press.
Baddeley, A. (2020). Exploring working memory: Selected works of Alan
Baddelely. London: Routledge.
Baratgin, J., Over, D. E., & Politzer, G. (2013). Uncertainty and the de Finetti
tables. Thinking & Reasoning, 19, 308–328.
Baratgin, J., Politzer, G., Over, D. E., & Takahashi, T. (2018). The psychology of
uncertainty and three-valued truth tables. Frontiers in Psychology, 9, 1479.
Barbey, A. K., & Sloman, S. A. (2007). Base-rate respect: From ecological
rationality to dual processes. Behavioral and Brain Sciences, 30, 241–297.
Boissin, E., Josserand, M., De Neys, W., & Caparos, S. (2024). Debiasing
thinking among non-WEIRD reasoners. Cognition, 243(3), 105681.
Bonnefon, J.-F. (2013). New ambitions for a new paradigm: Putting the
psychology of reasoning at the service of humanity. Thinking &
Reasoning, 19(3–4), 381–398.
Bourlier, M., Jacquet, J., Lassiter, D., & Baratgin, J. (2023). Coherence, not
conditional meaning, accounts for the relevance effect. Frontiers in
Psychology, 14, 1150550.
Braine, M. D. S., & O’Brien, D. P. (1991). A theory of if: A lexical entry, reasoning
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

program, and pragmatic principles. Psychological Review, 98, 182–203.

Byrne, R. M. J. (1989). Suppressing valid inferences with conditionals.
Cognition, 31, 61–83.
Cantwell, J. (2020). Revisiting McGee’s probabilistic analysis of conditionals.
Journal of Philosophical logic, 5, 1–45.
Chater, N., & Oaksford, M. (1999). The probability heuristics model of
syllogistic reasoning. Cognitive Psychology, 38, 191–258.
Cleeremans, A. (2015). Implicit learning and consciousness. London: Routledge.
Cleeremans, A., & Kuvaldina, M. (2019). Implicit learning: 50 years on.
London: Routledge.
Cohen, L. J. (1981). Can human irrationality be experimentally demonstrated?
Behavioral and Brain Sciences, 4, 317–370.
Cosmides, L., & Tooby, J. (1996). Are humans good intuitive statisticians after
all? Rethinking some conclusions from the literature on judgment under
uncertainty. Cognition, 58, 1–73.
 References 63

Corbett, B., Feeney, A., & McCormack, T. (2023). Prosocial risk taking and
interpersonal regret in children: An individual differences study. Social
Development, 32, 171–187.
Cruz, N. (2020). Deduction from uncertain premises? In S. Elqayam, I. Douven,
J. St. B. T. Evans, & N. Cruz (Eds.), Logic and uncertainty in the human mind
(pp. 27–41). Abingdon: Routledge.
Cruz, N., Baratgin, J., Oaksford, M., & Over, D. E. (2015). Bayesian reasoning
with ifs and ands and ors. Frontiers in Psychology, 6, 192.
Cruz, N., & Over, D. E. (2023). Independence conditionals. In S. Kaufmann,
D. E. Over, & G. Sharma (Eds.), Conditionals – Logic, linguistics and
psychology (pp. 223–233). London: Palgrave Macmillan.
Cruz, N., & Over, D. E. (2024). From de Finetti’s three values to conditional
probabilities in the psychology of reasoning. In P. Egré, & L. Rossi (Eds.),
Handbook of trivalent logics. Cambridge, MA: MIT Press.
Cruz, N., Over, D., Oaksford, M., & Baratgin, J. (2016). Centering and the
meaning of conditionals. In A. Papafragou, D. Grodner, D. Mirman, &
J. C. Trueswell (Eds.), Proceedings of the 38th annual conference of the
cognitive science society (pp. 1104–1109). Austin, TX: Cognitive Science
Society.
Cruz, N., Over, D., & Oaksford, M. (2017). The elusive oddness of or-
introduction. In G. Gunzelmann, A. Howes, T. Tenbrink, & E. Davelaar
(Eds.), The 39th annual meeting of the cognitive science society (pp.
663–668). London: Cognitive Science Society.
De Finetti, B. (1936/1995). The logic of probability. Philosophical Studies, 77,
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

181–190.
De Finetti, B. (1937/1964). Foresight: Its logical laws, its subjective sources.
In H. E. Kyburg, & H. E. Smokier (Eds.), Studies in subjective probability
(pp. 55–118). New York: Wiley.
De Neys, W. (2012). Bias and conflict: A case for logical intuitions.
Perspectives on Psychological Science, 7, 28–38.
De Neys, W. (2022). Advancing theorizing about fast-and-slow thinking.
Behavioral and Brain Sciences, 46, e111.
Doherty, M. E., Mynatt, C. R., Tweney, R. D., & Schiavo, M. D. (1979).
Pseudodiagnosticity. Acta Psychologica, 43, 11–21.
Douven, I., Elqayam, S., & Krzyżanowska, K. (2023). Inferentialism:
A manifesto. In S. Kaufmann, D. E. Over, & G. Sharma (Eds.),
Conditionals – Logic, linguistics and psychology (pp. 175–221). London:
Palgrave Macmillan.
Deary, I. J. (2020). Intelligence: A very short introduction. Oxford: Oxford
University Press.
 64 References

Demey, L., Kooi, B., & Sack, J. (2023). Logic and probability. The Stanford
Encyclopedia of Philosophy. E. N. Zalta, & U. Nodelman (Eds.), https://
plato.stanford.edu/archives/fall2023/entries/logic-probability/.
Edgington, D. (1995). On conditionals. Mind, 104, 235–329.
Egré, P., Rossi, L., & Sprenger, J. (2021). De Finettian logics of indicative
conditionals. Journal of Philosophical Logic, 50, 187–213.
Elqayam, S., & Evans, J. St. B. T. (2011). Subtracting ‘ought’ from ‘is’:
Descriptivism versus normativism in the study of human thinking.
Behavioral and Brain Sciences, 34, 233–290.
Elqayam, S., & Evans, J. St. B. T. (2013). Rationality in the new paradigm: Strict
versus soft Bayesian approaches. Thinking & Reasoning, 19, 453–470.
Engle, R. W. (2002). Working memory capacity as executive attention. Current
Directions in Psychological Science, 11, 19–23.
Epstein, S. (1994). Integration of the cognitive and psychodynamic
unconscious. American Psychologist, 49, 709–724.
Evans, J. St. B. T. (1989). Bias in human reasoning: Causes and consequences.
Brighton: Erlbaum.
Evans, J. St. B. T. (2002). Logic and human reasoning: An assessment of the
deduction paradigm. Psychological Bulletin, 128, 978–996.
Evans, J. St. B. T. (2005). How to do research: A psychologist’s guide. Hove:
Psychology Press.
Evans, J. St. B. T. (2006). The heuristic-analytic theory of reasoning: Extension
and evaluation. Psychonomic Bulletin & Review, 13, 378–395.
Evans, J. St. B. T. (2007a). Hypothetical thinking: Dual processes in reasoning
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

and judgement. Hove: Psychology Press.

Evans, J. St. B. T. (2007b). On the resolution of conflict in dual-process theories
of reasoning. Thinking & Reasoning, 13, 321–329.
Evans, J. St. B. T. (2010). Thinking twice: Two minds in one brain. Oxford:
Oxford University Press.
Evans, J. St. B. T. (2014). Two minds rationality. Thinking & Reasoning, 20,
129–146.
Evans, J. St. B. T. (2016). Reasoning, biases and dual processes: The lasting
impact of Wason (1960). The Quarterly Journal of Experimental Psychology,
69(10), 1–17.
Evans, J. St. B. T. (2018). Dual-process theory: Perspectives and problems. In
W. De Neys (Ed.), Dual process theory 2.0 (pp. 137–155). London:
Routledge.
Evans, J. St. B. T. (2019). Reflections on reflection: The nature and function of
type 2 processes in dual-process theories of reasoning. Thinking &
Reasoning, 25(4), 383–415.
 References 65

Evans, J. St. B. T. (2021). The rationality debate in the psychology of reasoning:

A historical review. In M. Knauff, & W. Spohn (Eds.), The handbook of
rationality (pp. 71–86). Cambridge, MA: MIT Press.
Evans, J. St. B. T. (2022). Wason selection task. In R. F. Pohl (Ed.), Cognitive
illusions: Intriguing phenomena in thinking, judgment and memory
(pp. 140–153). New York: Routledge.
Evans, J. St. B. T. (2020). The suppositional conditional is not (just) the
probability conditional. In S. Elqayam, I. Douven, J. St. B. T. Evans, &
N. Cruz (Eds.), Logic and uncertainty in the human mind (pp. 57–70).
Abingdon: Routledge.
Evans, J. St. B. T., Ball, L. J., & Thompson, V. A. (2022). Belief bias in deductive
reasoning. In R. F. Pohl (Ed.), Cognitive illusions: Intriguing phenomena in
thinking, judgement and memory (pp. 154–172). New York: Routledge.
Evans, J. St. B. T., Barston, J. L., & Pollard, P. (1983). On the conflict between
logic and belief in syllogistic reasoning. Memory & Cognition, 11, 295–306.
Evans, J. St. B. T., & Curtis-Holmes, J. (2005). Rapid responding increases
belief bias: Evidence for the dual-process theory of reasoning. Thinking &
Reasoning, 11(4), 382–389.
Evans, J. St. B. T., Handley, S., Neilens, H., & Over, D. E. (2007). Thinking
about conditionals: A study of individual differences. Memory & Cognition,
35, 1772–1784.
Evans, J. St. B. T., Handley, S., Neilens, H., & Over, D. E. (2010). The influence
of cognitive ability and instructional set on causal conditional inference.
Quarterly Journal of Experimental Psychology, 63(5), 892–909.
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

Evans, J. St. B. T., Handley, S. J., Harper, C., & Johnson-Laird, P. N. (1999).
Reasoning about necessity and possibility: A test of the mental model theory
of deduction. Journal of Experimental Psychology: Learning Memory and
Cognition, 25, 1495–1513.
Evans, J. St. B. T., Handley, S. J., & Over, D. E. (2003). Conditionals and
conditional probability. Journal of Experimental Psychology: Learning
Memory and Cognition, 29, 321–335.
Evans, J. St. B. T., & Lynch, J. S. (1973). Matching bias in the selection task.
British Journal of Psychology, 64, 391–397.
Evans, J. St. B. T., Newstead, S. E., & Byrne, R. M. J. (1993). Human reason-
ing: The psychology of deduction. Hove: Erlbaum.
Evans, J. St. B. T., & Over, D. E. (1996). Rationality and reasoning. Hove:
Psychology Press.
Evans, J. St. B. T., & Over, D. E. (2004). If. Oxford: Oxford University Press.
Evans, J. St. B. T. (2013). Two minds rationality. Thinking & Reasoning, 20,
129–146.
 66 References

Evans, J. St. B. T., & Stanovich, K. E. (2013). Dual process theories of higher
cognition: Advancing the debate. Perspectives on Psychological Science, 8,
223–241.
Evans, J. St. B. T., Thompson, V., & Over, D. E. (2015). Uncertain deduction
and conditional reasoning. Frontiers in Psychology, 6, 398.
Evans, J. St. B. T., & Wason, P. C. (1976). Rationalisation in a reasoning task.
British Journal of Psychology, 63, 205–212.
Frankish, K., & Evans, J. St. B. T. (2009). The duality of mind: An historical
perspective. In J. St. B. T. Evans, & K. Frankish (Eds.), In two minds: Dual
processes and beyond (pp. 1–30). Oxford: Oxford University Press.
Frederick, S. (2005). Cognitive reflection and decision making. Journal of
Economic Perspectives, 19(4), 25–42.
Feeney, A. (2018). Forty years of progress on category-based inductive reason-
ing. In L. J. Ball, & V. A. Thompson (Eds.), International handbook of
thinking and reasoning (pp. 167–185). Hove: Psychology Press.
Ghasemi, O., Handley, S. J., Howarth, S., Newman, I. R., & Thompson, V. A.
(2022). Logical intuition is not really about logic. Journal of Experimental
Psychology: General, 151, 2009–2028.
Ghasemi, O., Handley, S. J., Howarth, S. (2023). Illusory intuitive inferences:
Matching heuristics explain logical intuitions. Cognition, 235, 105417.
Gigerenzer, G. (2007). Gut feelings: The intelligence of the unconscious.
London: Penguin.
Gigerenzer, G., & Hoffrage, U. (1995). How to improve Bayesian reasoning
without instruction: Frequency formats. Psychological Review, 102, 684–704.
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

Gilio, A., & Over, D. E. (2012). The psychology of inferring conditionals from
disjunctions: A probabilistic study. Journal of Mathematical Psychology, 56,
118–131.
Gilio, A., Pfeifer, N., & Sanfilippo, G. (2020). Probabilistic entailment and
iterated conditionals. In S. Elqayam, I. Douven, J. St. B. T. Evans, & N. Cruz
(Eds.), Logic and uncertainty in the human mind (pp. 71–101). Abingdon:
Routledge.
Gilovich, T., Griffin, D., & Kahneman, D. (2002). Heuristics and biases: The
psychology of intuitive judgement. Cambridge: Cambridge University Press.
Gladwell, M. (2005). Blink. London: Penguin.
Good, I. J. (1983). Good thinking: The foundations of probability and its
applications. Minneapolis: University of Minnesota Press.
Goodwin, R. Q., & Wason, P. C. (1972). Degrees of insight. British Journal of
Psychology, 63, 205–212.
Grice, P. (1989). Studies in the way of words. Cambridge, MA: Harvard
University Press.
 References 67

Hadjichristidis, C., Sloman, S. A., & Over, D. E. (2014). Categorical induction

from uncertain premises: Jeffrey’s doesn’t completely rule. Thinking &
Reasoning, 20, 405–431.
Hahn, U., & Oaksford, M. (2007). The rationality of informal argumentation:
A Bayesian approach to reasoning fallacies. Psychological Review, 114,
704–732.
Handley, S. J., Newstead, S. E., & Trippas, D. (2011). Logic, beliefs, and
instruction: A test of the default interventionist account of belief bias.
Journal of Experimental Psychology-Learning Memory and Cognition,
37(1), 28–43.
Harris, A., J. L., Hahn, U., Madsen, J. K., & Hsu, A. S. (2016). The appeal to
expert opinion: Quantitative support for a Bayesian network approach.
Cognitive Science, 40, 1496–1533.
Henle, M. (1962). On the relation between logic and thinking. Psychological
Review, 69, 366–378.
Howson, C., & Urbach, P. (2006). Scientific reasoning: The Bayesian approach
(third edition). Chicago, IL: Open Court.
Inhelder, B., & Piaget, J. (1958). The growth of logical thinking. New York:
Basic Books.
Jeffrey, R. C. (1983). The logic of decision (second edition). Chicago, IL:
University of Chicago Press.
Jeffrey, R. C. (1991). Matter of fact conditionals. Aristotelian Society
Supplementary Volume, 65, 161–183.
Johnson-Laird, P. N. (1983). Mental models. Cambridge: Cambridge University
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

Press.
Johnson-Laird, P. N., & Bara, B. G. (1984). Syllogistic inference. Cognition, 16,
1–61.
Johnson-Laird, P. N., & Byrne, R. M. J. (1991). Deduction. Hove: Erlbaum.
Johnson-Laird, P. N., Khemlani, S., & Goodwin, G. P. (2015). Logic, probabil-
ity, and human reasoning. Trends in Cognitive Science, 19, 201–214.
Kahneman, D. (2011). Thinking, fast and slow. New York: Farrar, Straus and
Giroux.
Kahneman, D., Slovic, P., & Tversky, A. (1982). Judgment under uncertainty:
Heuristics and biases. Cambridge: Cambridge University Press.
Kaufmann, S. (2023). How fake is fake past? In S. Kaufmann, D. E. Over, &
G. Sharma (Eds.), Conditionals – Logic, linguistics and psychology (pp.
389–423). London: Palgrave Macmillan.
Keren, G., & Schul, Y. (2009). Two is not always better than one: A critical
evaluation of two-system theories. Perspectives on Psychological Science, 4,
533–550.
 68 References

Klayman, J., & Ha, Y. W. (1987). Confirmation, disconfirmation and informa-

tion in hypothesis testing. Psychological Review, 94, 211–228
Klein, G. (1998). Sources of power. Cambridge, MA: MIT Press.
Kleiter, G. D. (1994). Natural sampling: Rationality without base rates. In
G. H. Fisher, & D. Laming (Eds.), Contributions to mathematical psych-
ology, psychometrics, and methodology (pp. 375–388). New York: Springer.
Kleiter, G. D., Fugard, J. B., Pfeifer, N. (2018). A process model of the
understanding of uncertain conditionals. Thinking & Reasoning, 24,
386–422.
Kneale, W., & Kneale, M. (1962). The development of logic. Oxford: Oxford
University Press.
Kolodny, N., & Brunero, J. (2023). Instrumental rationality. The Stanford
Encyclopedia of Philosophy. E. N. Zalta, & U. Nodelman (Eds.), https://
plato.stanford.edu/archives/sum2023/entries/rationality-instrumental/.
Kruglanski, A. W., & Gigerenzer, G. (2011). Intuitive and deliberative judgements
are based on common principles. Psychological Review, 118(1), 97–109.
Lassiter, D. (2019). What we can learn from how trivalent conditionals avoid
triviality. Inquiry, 63, 1087–1114.
Lassiter, D. (2023). Decomposing relevance in conditionals. Mind & Language,
38(3), 644–668.
Lassiter, D., & Goodman, N. D. (2015). How many kinds of reasoning?
Inference, probability, and natural language semantics. Cognition, 136,
123–134.
Lewis, D. (1973). Counterfactuals. Cambridge, MA: Harvard University Press.
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

Lewis, D. (1976). Probabilities of conditionals and conditional probabilities.

Philosophical Review, 85, 297–315.
Lucas, E. J., & Ball, L. J. (2005). Think-aloud protocols and the selection task:
Evidence for relevance effects and rationalisation processes. Thinking and
Reasoning, 11(1), 35–66.
Manktelow, K. I. (2021). Beyond reasoning: The life, time and work of Peter
Wason, pioneering psychologist. Abingdon: Routledge.
McDowell, M., & Jacobs, P. (2017). Meta-analysis of the effect of natural
frequencies on Bayesian reasoning. Psychological Bulletin, 143, 1273–1312.
Mercier, H., & Sperber, D. (2011). Why do humans reason? Arguments for an
argumentative theory. Behavioral and Brain Sciences, 34, 57–111.
Mercier, H., & Sperber, D. (2017). The enigma of reason: A new theory of
human understanding. Cambridge, MA: Harvard University Press.
Morley, N. J., Evans, J. St. B. T., & Handley, S. J. (2004). Belief bias and figural
bias in syllogistic reasoning. Quarterly Journal of Experimental Psychology,
57(4), 666–692.
 References 69

Newstead, S. E., Handley, S. J., Harley, C., Wright, H., & Farelly, D. (2004).
Individual differences in deductive reasoning. Quarterly Journal of
Experimental Psychology, 57A, 33–60.
Oakhill, J., Johnson-Laird, P. N., & Garnham, A. (1989). Believability and
syllogistic reasoning. Cognition, 31, 117–140.
Oaksford, M., & Chater, N. (1994). A rational analysis of the selection task as
optimal data selection. Psychological Review, 101, 608–631.
Oaksford, M., & Chater, N. (2007). Bayesian rationality: The probabilistic
approach to human reasoning. Oxford: Oxford University Press.
Oaksford, M., & Chater, N. (2020). New paradigms in the psychology of
reasoning. Annual Review of Psychology, 71, 305–330.
Oberauer, K., & Wilhelm, O. (2003). The meaning(s) of conditionals:
Conditional probabilities, mental models and personal utilities. Journal of
Experimental Psychology: Learning Memory and Cognition, 29, 680–693.
Oberauer, K., Weidenfeld, A., & Fischer, K. (2007). What makes us believe a
conditional? Thinking & Reasoning, 13(4), 340–369.
Osman, M. (2004). An evaluation of dual-process theories of reasoning.
Psychonomic Bulletin and Review, 11(6), 988–1010.
Over, D. E. (2003). From massive modularity to metarepresentation: The evolution
of higher cognition. In D. E. Over (Ed.), Evolution and the psychology of
thinking: The debate (pp. 121–144). Hove: The Psychology Press.
Over, D. E. (2009). New paradigm psychology of reasoning. Thinking &
Reasoning, 15, 431–438.
Over, D. E. (2020). The development of the new paradigm in the psychology of
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

reasoning. In S. Elqayam, I. Douven, J. St. B. T. Evans, & N. Cruz (Eds.),

Logic and uncertainty in the human mind (pp. 431–438). Abingdon:
Routledge.
Over, D. E. (2023a). The new paradigm and massive modalization: Commentary
on Knauff and Gazzo Castañeda. Thinking & Reasoning, 29(3), 389–395.
Over, D. E. (2023b). Independence and rationality. The Reasoner: Focus issue
on the Handbook of Rationality, 17(6), 47–48. www.thereasoner.org/.
Over, D. E., & Baratgin, J. (2017). The ‘defective’ truth table: Its past, present,
and future. In N. Galbraith, D. E. Over, & E. Lucas, (Eds.), The thinking
mind: A Festschrift for Ken Manktelow (pp. 15–28). Hove: Psychology
Press.
Over, D. E., & Cruz, N. (2018). Probabilistic accounts of conditional reasoning.
In L. J. Ball, & V. A. Thompson (Eds.), International handbook of thinking
and reasoning (pp. 434–450). Hove: Psychology Press.
Over, D. E., & Cruz, N. (2023). Indicative and counterfactual conditionals in the
psychology of reasoning. In S. Kaufmann, D. E. Over, & G. Sharma (Eds.),
 70 References

Conditionals – Logic, linguistics and psychology (pp. 139–174). London:

Palgrave Macmillan.
Over, D. E., Hadjichristidis, C., Evans, J. St. B. T., Handley, S. J., &
Sloman, S. A. (2007). The probability of causal conditionals. Cognitive
Psychology, 54, 62–97.
Pearl, J. (2000). Causality: Models, reasoning, and inference. Cambridge:
Cambridge University Press.
Pearl, J. (2013). Structural counterfactuals: A brief introduction. Cognitive
Science, 37, 977–985.
Pfeifer, N. (2023). The logic and pragmatics of conditionals under uncertainty:
A mental probability logic perspective. In S. Kaufmann, D. E. Over, &
G. Sharma (Eds.), Conditionals – Logic, linguistics and psychology (pp.
73–102). London: Palgrave Macmillan.
Pfeifer, N., & Kleiter, G. D. (2009). Framing human inference by coherence
based probability logic. Journal of Applied Logic, 7, 206–217.
Poletiek, F. (2001). Hypothesis-testing behaviour. Hove: Psychology Press.
Politzer, G., Over, D. E., & Baratgin, J. (2010). Betting on conditionals.
Thinking & Reasoning, 16, 172–197.
Popper, K. R. (1959). The logic of scientific discovery. London: Hutchinson.
Ramsey, F. P. (1926/1990). Truth and probability. In D. H. Mellor (Ed.),
Philosophical papers (pp. 52–94). Cambridge: Cambridge University
Press.
Ramsey, F. P. (1929/1990). General propositions and causality. In D. H. Mellor
(Ed.), Philosophical papers (pp. 145–163). Cambridge: Cambridge
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

University Press.
Reber, A. S. (1993). Implicit learning and tacit knowledge. Oxford: Oxford
University Press.
Rhodes, S., Galbraith, N., & Manktelow, K. (2020). Delusional rationality. In
S. Elqayam, I. Douven, J. St. B. T. Evans, & N. Cruz (Eds.), Logic and
uncertainty in the human mind (pp. 178–191). Abingdon: Routledge.
Rips, L. J. (1994). The psychology of proof. Cambridge, MA: MIT Press.
Rothschild, D. (2023). Living in a material world: A critical notice of Suppose
and tell. Mind, 132, 208–233.
Sanfilippo, G., Gilio, A., Over, D. E., & Pfeifer, N. (2020). Probabilities of
conditionals and previsions of iterated conditionals. International Journal of
Approximate Reasoning, 121, 150–173.
Schacter, D. L. (1987). Implicit memory: History and current status. Journal of
Experimental Psychology: Learning Memory and Cognition, 13, 501–518.
Sebben, S., & Ullrich, J. (2021). Can conditionals explain explanations?
A modus ponens model of B because A. Cognition, 215, 104812.
 References 71

Shapiro, S., & Kouri Kissel, T. (2022). Classical logic. The Stanford Encyclopedia
of Philosophy, E. N. Zalta & U. Nodelman (Eds.), https://plato.stanford.edu/
archives/win2022/entries/logic-classical/.
Singmann, H., Klauer, K. C., & Over, D. E. (2014). New normative standards
of conditional reasoning and the dual-source model. Frontiers of
Psychology, 5, 316.
Skovgaard-Olsen, N., Kellen, D., Krahl, H., & Klauer, K. C. (2017). Relevance
differently affects the truth, acceptability, and probability evaluations of
‘and’, ‘but’, ‘therefore’, and ‘if then’. Thinking & Reasoning, 23, 449–482.
Skovgaard-Olsen, N., Singmann, H., & Klauer, K. C. (2016). The relevance
effect and conditionals. Cognition, 150, 26–36.
Sloman, S. A. (1996). The empirical case for two systems of reasoning.
Psychological Bulletin, 119, 3–22.
Sloman, S. A., & Fernbach, P. (2017). The knowledge illusion: Why we never
think alone. New York: Riverhead.
Sloman, S. A., & Lagnado, D. A. (2015). Causality in thought. Annual Review
of Psychology, 66, 223–247.
Smedslund, J. (1970). Circular relation between understanding and logic.
Scandinavian Journal of Psychology, 11, 217–219.
Smith, E., & DeCoster, J. (1999). Associative and rule-based processing:
A connectionist interpretation of dual-process models. In S. Chaiken, &
Y. Trope (Eds.), Dual-process theories in social psychology (pp. 323–360).
New York: The Guildford Press.
Spearman, C. (1904). General intelligence, objectively determined and
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

measured. American Journal of Psychology, 15, 201–293.

Sprenger, J., & Hartmann, S. (2019). Bayesian philosophy of science. Oxford:
Oxford University Press.
Stalnaker, R. (1968). A theory of conditionals. In N. Rescher (Ed.), Studies in
logical theory (pp. 98–112). Oxford: Blackwell.
Stanovich, K. E. (1999). Who is rational? Studies of individual differences in
reasoning. Mahway, NJ: Lawrence Elrbaum Associates.
Stanovich, K. E. (2004). The robot’s rebellion: Finding meaning the age of
Darwin. Chicago, IL: University of Chicago Press.
Stanovich, K. E. (2009). What intelligence tests miss. London: Yale university
press.
Stanovich, K. E. (2011). Rationality and the reflective mind. New York: Oxford
University Press.
Stanovich, K. E. (2013). Why humans are (sometimes) less rational than other
animals: Cognitive complexity and the axioms of rational choice. Thinking &
Reasoning, 19, 1–26.
 72 References

Stanovich, K. E. (2018). Miserliness in human cognition: The interaction of

detection, override and mindware. Thinking & Reasoning, 24, 423–444.
Stanovich, K. E. (2021). The bias that divides us: The science and politics of
myside thinking. Cambridge, MA: MIT Press.
Stanovich, K. E., Sÿ, W. C., & West, R. F. (1999). The domain specificity and
generality of belief bias: Searching for a generalizable critical thinking skill.
Journal of Educational Psychology, 91(3), 497–510.
Stanovich, K. E., West, C., & Toplak, M. E. (2016). The rationality quotient:
Towards a test of ratiional thinking. Cambridge, MA: MIT Press.
Stanovich, K. E., & West, R. F. (1998). Cognitive ability and variation in
selection task performance. Thinking & Reasoning, 4, 193–230.
Stanovich, K. E., & West, R. F. (2000). Advancing the rationality debate.
Behavioral and Brain Sciences, 23(5), 701–726.
Stevenson, R. J., & Over, D. E. (1995). Deduction from uncertain premises.
Quarterly Journal of Experimental Psychology, 48A, 613–643.
Thompson, V. A. (1994). Interpretational factors in conditional reasoning.
Memory & Cognition, 22, 742–758.
Thompson, V. A., & Byrne, R. M. J. (2002). Reasoning counterfactually: Making
inferences about things that didn’t happen. Journal of Experimental
Psychology: Learning, Memory, and Cognition, 28, 1154–1170.
Thompson, V. A., Evans, J. St. B., & Campbell, J. I. D. (2013). Matching bias on
the selection task: It’s fast and feels good. Thinking & Reasoning, 19(3–4),
431–452.
Thompson, V. A., Pennycook, G., Trippas, D., & Evans, J. St. B. T. (2018). Do
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

smart people have better intuItions? Journal of Experimental Psychology:

General, 147, 945–961.
Thompson, V. A., Prowse Turner, J. A., & Pennycook, G. (2011). Intuition,
reason, and metacognition. Cognitive Psychology, 63(3), 107–140.
Toates, F. (2004). In two minds – Consideration of evolutionary precursors
permits a more integrative theory. Trends in Cognitive Sciences, 8(2), 57.
Toates, F. (2006). A model of the hierarchy of behaviour, cognition and
consciousness. Consciousness and Cognition, 15, 75–118.
Toplak, M. E., West, R. F., & Stanovich, K. E. (2011). The cognitive reflection
test as a predictor of performance on heuristics-and-biases tasks. Memory &
Cognition, 39(7), 1275–1289.
Tversky, A., & Kahneman, D. (1974). Judgement under uncertainty: Heuristics
and biases. Science, 185, 1124–1131.
Tversky, A., & Kahneman, D. (1983). Extensional vs intuitive reasoning: The
conjunction fallacy in probability judgment. Psychological Review, 90,
293–315.
 References 73

van Rooij, R., Krzyżanowska, K., & Douven, I. (2023). Williamson’s abductive
case for the material conditional account. Studia Logica, 111, 653–685.
Wang, M., Over, D. E., & Liang, L. (2022). What is required for the truth of
a general conditional? Quarterly Journal of Experimental Psychology, 75, 11.
Wason, P. C. (1960). On the failure to eliminate hypotheses in a conceptual task.
Quarterly Journal of Experimental Psychology, 12, 12–40.
Wason, P. C. (1966). Reasoning. In B. M. Foss (Ed.), New horizons in psych-
ology I (pp. 106–137). Harmandsworth: Penguin.
Wason, P. C. (1983). Realism and rationality in the selection task. In
J. St. B. T. Evans (Ed.), Thinking and reasoning: Psychological approaches
(pp. 44–75). London: Routledge & Kegan Paul.
Wason, P. C., & Evans, J. St. B. T. (1975). Dual processes in reasoning?
Cognition, 3, 141–154.
Wason, P. C., & Johnson-Laird, P. N. (1972). Psychology of reasoning:
Structure and content. London: Batsford.
Wilkins, M. C. (1928). The effect of changed material on the ability to do formal
syllogistic reasoning. Archives of Psychology, 16, 102.
Williamson, T. (2020). Suppose and tell: The semantics and heuristics of
conditionals. Oxford: Oxford University Press.
Woodworth, R. S., & Sells, S. B. (1935). An atmosphere effect in syllogistic
reasoning. Journal of Experimental Psychology, 18, 451–460.
Zhao, J., & Osherson, D. (2010). Updating beliefs in light of uncertain evidence:
Descriptive assessment of Jeffrey’s rule. Thinking & Reasoning, 16(4),
288–307.
Zimmer, K. (2021). People may have used fire to clear forests more than 80,000
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

years ago. Scientific American. www.scientificamerican.com/article/people-

may-have-used-fire-to-clear-forests-more-than-80-000-years-ago/.
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press
 Philosophy of Mind

Keith Frankish
The University of Sheffield
Keith Frankish is a philosopher specializing in philosophy of mind, philosophy of
psychology, and philosophy of cognitive science. He is the author of Mind and Supermind
(Cambridge University Press, 2004) and Consciousness (2005), and has also edited or
coedited several collections of essays, including The Cambridge Handbook of Cognitive
Science (Cambridge University Press, 2012), The Cambridge Handbook of Artificial
Intelligence (Cambridge University Press, 2014) (both with William Ramsey),
and Illusionism as a Theory of Consciousness (2017).

About the Series

This series provides concise, authoritative introductions to contemporary work in
philosophy of mind, written by leading researchers and including both established and
emerging topics. It provides an entry point to the primary literature and will be the
standard resource for researchers, students, and anyone wanting a firm grounding
in this fascinating field.
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press
 Philosophy of Mind

Elements in the Series

Imagination and Creative Thinking
Amy Kind
Attention and Mental Control
Carolyn Dicey Jennings
Biological Cognition
Bryce Huebner and Jay Schulkin
Embodied and Enactive Approaches to Cognition
Shaun Gallagher
Mental Content
Peter Schulte
Affective Bodily Awareness
Frédérique de Vignemont
The Computational Theory of Mind
Matteo Colombo and Gualtiero Piccinini
Memory and Remembering
Felipe De Brigard
Non-physicalist Theories of Consciousness
Hedda Hassel Mørch
Neurolaw
Gregg D. Caruso
https://doi.org/10.1017/9781009495349 Published online by Cambridge University Press

Animal Minds
Marta Halina
Human Reasoning
David E. Over and Jonathan St B. T. Evans

A full series listing is available at: www.cambridge.org/EPMI