Quantifying the Benefit of Artificial Intelligence for Scientific Research

Jian Gao1,2,3, Dashun Wang1,2,3,4*

1
Center for Science of Science and Innovation, Northwestern University, Evanston, IL, 60208, USA
2
Kellogg School of Management, Northwestern University, Evanston, IL, 60208, USA
3
Northwestern Institute on Complex Systems, Northwestern University, Evanston, IL, 60208, USA
4
McCormick School of Engineering, Northwestern University, Evanston, IL, 60208, USA
* Correspondence to: [email protected]

The ongoing artificial intelligence (AI) revolution has the potential to change almost every
line of work. As AI capabilities continue to improve in accuracy, robustness, and reach, AI
may outperform and even replace human experts across many valuable tasks. Despite
enormous effort devoted to understanding the impact of AI on labor and the economy and
AI’s recent successes in accelerating scientific discovery and progress, we lack a systematic
understanding of how AI advances may benefit scientific research across disciplines and
fields. Here, drawing from the literature on the future of work and the science of science, we
develop a measurement framework to estimate both the direct use of AI and the potential
benefit of AI in scientific research, applying natural language processing techniques to 74.6
million publications and 7.1 million patents. We find that the use of AI in research is
widespread throughout the sciences, growing especially rapidly since 2015, and papers that
use AI exhibit a citation premium, more likely to be highly cited both within and outside
their disciplines. Moreover, our analysis reveals considerable potential for AI to benefit
numerous scientific fields, yet a notable disconnect exists between AI education and its
research applications, highlighting a mismatch between the supply of AI expertise and its
demand in research. Lastly, we examine demographic disparities in AI’s benefits across
scientific disciplines and find that disciplines with a higher proportion of women or Black
scientists tend to be associated with less benefit, suggesting that AI’s growing impact on
research may further exacerbate existing inequalities in science. As the connection between
AI and scientific research deepens, our findings may become increasingly important, with
implications for the equity and sustainability of the research enterprise.

1
Main text
The rapid advances in artificial intelligence (AI) may lead to massive value creation and capture

across many facets of human society1-5, creating enormous social and economic opportunities6-8,

and just as many challenges9-17. While extensive efforts have been devoted to understanding the

impact of AI on the labor market and the economy18-22, the impact of AI on the growing research

enterprise remains unclear. Indeed, recent AI advances have shown promise to achieve and, in

some cases, exceed expert-level performance across many economically valuable tasks23-30. As

society prepares for the moment when AI may outperform or even replace human recruiters,

bankers, doctors, lawyers, composers, and drivers, an important question arises: What is the impact

of AI in advancing scientific research across different disciplines and fields?

A better understanding of the impact of AI on science may not only help guide AI development,

bridging AI innovations more closely with scientific research but also hold implications for science

and innovation policy. Understanding the impact of AI on science is both timely and important

given AI’s recent remarkable success in advancing research frontiers across several fields31-44,

from predicting the structure of proteins in biology45-47 to designing new drug candidates in

medicine48-51, from discovering natural laws in physics52-54 to solving complicated equations and

discovering new conjectures in mathematics55-57, from controlling nuclear fusion58 to predicting

new material properties59-62, from designing taxation policy63 to suggesting democratic social

mechanisms64, and many more65-70. These advances raise the possibility that as AI continues to

improve in accuracy, robustness, and reach71-76, it may bring meaningful benefits to science,

propelling scientific progress across a range of research areas while significantly augmenting

researchers’ innovative capacities.

Yet, despite the rapid progress of AI and its broad applications in several domains, there is

substantial skepticism about whether today’s AI is capable or significant enough to advance

2
scientific research. Indeed, most current AI applications belong to the category of “narrow AI,”77-
79
which tackles specifically defined problems, and hence may not be suitable to fulfill the broad

range of tasks that scientific research demands2,18,80. Further, to the extent that AI may provide

automated solutions to an existing problem, science is about not only solving well-defined

problems but also spotting new frontiers and generating novel hypotheses81. These views paint a

more nuanced picture of AI’s applicability to advancing science, suggesting that AI may be better

suited to perform some research tasks than others2,10,82.

Building on the growing literature on the future of work83-87 and the science of science88-96, here

we develop a quantitative framework for estimating the impact of AI on scientific research (see

Methods for details). Our primary dataset contains 74.6 million publications from 1960 to 2019

from the Microsoft Academic Graph (MAG) dataset97, spanning 19 disciplines and 292 fields (see

Supplementary Note 1.1 for details). We integrate this dataset with 7.1 million patents granted

between 1976 and 2019 by the U.S. Patent and Trademark Office (USPTO) (see Supplementary

Note 1.2). We then follow prior studies to identify AI publications and AI patents using a keyword-

based approach (see Supplementary Notes 2.1 and 3.1 for details)91,97,98, allowing us to measure

AI’s impact on science at two levels. First, we quantify the direct impact of AI using an “AI n-

gram framework” (Fig. 1a), which estimates the relative frequency of the use of AI in a field (see

Supplementary Notes 2.3). Specifically, we extract AI n-grams (bigrams and trigrams; e.g., “deep

learning” and “convolutional neural network”) from both the titles and abstracts of AI publications

and calculate the frequency of their occurrences to approximate AI advances91,96. We then repeat

this n-gram measurement for publications in each field and year, allowing us to calculate the

weighted frequency of AI n-grams appearing in a paper to approximate the direct impact of AI on

each field and year. Second, motivated by the future of work literature82-84, we quantify the

3
potential impact of AI using an “AI capability–field task framework” (Fig. 2a), which measures

the alignment between AI capabilities and the tasks of a field (see Supplementary Notes 3.3). In

particular, we infer the capabilities of AI (i.e., what AI can do) by extracting verb-noun pairs (e.g.,

“learn representation”) from the titles of AI publications and AI patents using natural language

processing (NLP) techniques and calculating their relative frequency99-101. Here, following prior

work82, we rely only on titles as they have a higher signal-to-noise ratio than abstracts. We then

estimate the tasks of each field (i.e., what a field does) by calculating the relative frequency of

verb-noun pairs extracted from the titles of publications in each field and year. Calculating the

overlap between the prevalent tasks in a field and the inferred AI capabilities allows us to

approximate the potential impact of AI on each field and year (see Methods for more details).

Results

The widespread use of AI across the sciences

Overall, AI research presents a dynamically evolving landscape (Figs. 1b,c). While the frequency

of certain dominant AI n-grams in 2019 (e.g., “machine learning,” “convolutional neural network,”

“deep learning,” “deep neural network,” and “artificial intelligence”) shows an overall upward

trend (Fig. 1b), some AI n-grams emerged only recently (e.g., “generative adversarial network”),

some rose to prominence after a long period of dormancy (e.g., “deep learning”), and some were

popular a decade ago but have become less prevalent in recent years (e.g., “support vector

machine”; Fig. 1c). Amidst this rapidly evolving AI research landscape, there has been a

precipitous rise in the use of AI by many disciplines, as proxied by the mention of AI n-grams in

the titles and abstracts of publications (Fig. 1d; see Supplementary Fig. 2 for details).

This increase in the use of AI by different disciplines raises an interesting question: How do the

citations of papers that use AI compare to those of other papers in the same field? To answer this

question, we define hit papers as those in the top 5% by total citations in the same field and year

4
and calculate the likelihood that a paper is a hit paper. We find that for a majority of disciplines,

papers that mention AI n-grams tend to be associated with a higher probability of being hit papers

within their disciplines (Fig. 1e), and they also receive more citations from other disciplines

compared to papers that do not mention AI n-grams (Fig. 1f). This citation premium of papers that

mention AI n-grams appears to be stronger in disciplines with a lower propensity to use AI (see

detailed correlation analysis in Supplementary Fig. 4), suggesting that disciplines that seem distant

from AI may see substantial benefits from using AI to advance their research.

The dynamic AI landscape also prompts us to explore changes in the direct impact of AI on

scientific research over time. Specifically, we calculate the direct AI impact score using the “AI

n-gram framework” for each discipline between 2000 and 2019 (see Supplementary Fig. 3 for the

results for 1960-2019), extracting AI n-grams from the titles and abstracts of AI publications and

calculating their frequency of occurrence in the publications in each discipline (see Methods and

Supplementary Note 2 for details). We find that disciplines overall used more AI in their research

over the past two decades (e.g., in computer science, the direct AI impact score increased from

0.5% in 2000 to 1.3% in 2019; Fig. 1g, solid lines). This increase occurs not only in computer

science but also in a wide range of other disciplines (Fig. 1h), including, for example, physics,

biology, and economics. Moreover, this increase has not been linear; there has been a notably sharp

increase in the use of AI since 2015 across many disciplines (Fig. 1g).

To better understand whether the recent rise in the direct impact of AI on science is associated

with changes in AI capabilities or field-specific shifts in research direction, we calculate an

alternative measure for AI impact scores by keeping AI capabilities fixed in 2015 and apply this

alternative measure to each discipline and year between 2015 and 2019 (see Supplementary Note

4.2). More specifically, we use (1) AI n-grams extracted from AI publications before 2015 without

5
changing either their terms or their frequency for 2015-2019, and (2) AI n-grams extracted from

publications in each discipline and year during the period 2015-2019 (Fig. 1g, dashed lines). We

find that the new scores deviate substantially from the original scores (Fig. 1g, solid lines, and

Supplementary Fig. 7), which indicates that across disciplines, sciences benefit more from cutting-

edge AI advances. Overall, these results suggest that new AI capabilities play an important role in

contributing to the recent, sharp increase in the use of AI across scientific disciplines (see

Supplementary Note 4 for more details and results related to AI’s growing impact).

The potential impact of AI

While explicit mentions of AI n-grams by publications signal the direct use of AI in research, AI

may also exert potential impacts on scientific research beyond these direct uses. In particular, the

growing AI capabilities may help perform some core tasks that a research field demands. Here we

build on the future of work literature, which suggests that AI capabilities and field tasks can be

captured by verb-noun pairs (e.g., “learn representation”)82-84, prompting us to develop an “AI

capability–field task framework” to quantify the potential impact of AI on scientific research (Fig.

2a). We apply NLP algorithms to extract verb-noun pairs from the titles of AI papers and AI

patents to estimate AI capabilities99-101 (see Methods for details and Supplementary Note 8.3 for

validations of the approach). Applied systematically to all disciplines and fields, this framework

allows us to estimate which subfields within a discipline may benefit most from AI. Taking biology

as an example, the results suggest that the subfield with the largest potential AI impact is

“biological system” (Fig. 2c), as many of its basic tasks appear aligned with inferred AI

capabilities (e.g., “extract feature,” “detect object,” and “improve prediction”). Interestingly, the

“biological system” field, ranked seventh among all non-computer science fields for the potential

AI impact (Fig. 2d), also happens to be the field for the AlphaFold paper45, which Science called

the 2021 Breakthrough of the Year102.

6
While there are considerable differences in the direct impact of AI on scientific disciplines (Fig.

1h), the differences in the potential AI impact are relatively small across disciplines

(Supplementary Fig.7), suggesting the potentially widespread applicability of AI in science. We

further study within-discipline heterogeneity by examining the percentiles of the direct and

potential AI impact scores for each discipline’s subfields (see Supplementary Note 4.1 for details).

We find that the two percentiles in each field are highly correlated with each other (Fig. 2e;

Pearson’s r = 0.891 and P-value < 0.001), especially among the top-ranked subfields (Fig. 2f; the

top 3 subfields within each discipline according to the two percentiles are entirely overlapped in

over half disciplines), indicating that the two measurements are consistent in identifying fields

most benefit from AI (see Supplementary Table 1 for the list of three subfields within each

discipline that have the highest direct AI impact scores and potential AI impact scores).

Nevertheless, almost every discipline has some subfields with a significant AI impact, which holds

robust even for disciplines with a low AI impact overall, such as sociology and economics (Fig.

2g; see Supplementary Fig. 8 for the results for all disciplines). Taken together, these results

suggest that the direct impact of AI on research is pervasive across disciplines and fields, and its

potential impact may extend beyond its current uses in science.

Growing knowledge demands for AI

The rapidly expanding AI frontier and its increasing impact on science may lead to growing

demands for AI expertise from domain experts, raising the question of whether the current

education and training on AI skills are commensurate with AI’s impact. To answer this question,

we analyze 4.2 million university course syllabi from the Open Syllabus Project (OSP) database103

and estimate the level of AI education in each discipline (see Methods and Supplementary Note 5

for details). We find that, excluding the top three computational disciplines (i.e., computer science,

mathematics, and engineering), the correlation between the AI education level in a discipline and

7
the impact of AI on the discipline decreases, as well as its significance (Figs. 3a,b; Pearson’s r =

0.493 with P-value = 0.074 for direct impact and Pearson’s r = 0.263 with P-value = 0.363 for

potential impact). The results suggest that the supply of AI talent and knowledge in most

disciplines appears to be incommensurate with the benefits these disciplines may extract from AI

capabilities, highlighting a substantial AI impact–AI training gap. This result is robust under

alternative measures of AI education levels (Supplementary Fig. 12).

To meet the growing knowledge demands on AI, domain experts may rely on cross-discipline

collaborations to access AI capabilities. We analyze collaboration patterns for AI publications in

domains other than computer science that are co-authored by domain experts and/or computer

scientists (as a proxy for AI researchers; see Methods and Supplementary Notes 6.1 and 6.2 for

detailed methods and alternative proxies of AI researchers). We find that, in aggregate, about 40%

of AI publications are published by domain experts and about one-third are collaborative works

(Fig. 3c). In disciplines where AI has a greater impact, we see a larger propensity for domain

experts to collaborate with computer scientists (Figs. 3d,e; Pearson’s r = 0.841 and P-value <

0.001 for direct impact; Pearson’s r = 0.802 and P-value < 0.001 for potential impact). Moreover,

we see faster growth in collaborative AI publications than in domain-only AI publications

(Supplementary Fig. 13). Overall, the share of collaborative AI publications in major disciplines

is increasing over time (Fig. 3f; see Supplementary Fig. 14 for the results for other disciplines),

suggesting that domain experts’ reliance on AI expertise is growing. These results are robust when

we use an alternative method of determining AI researchers (see Supplementary Note 6.3 for

detailed methods and results).

We further explored discrepancies between the potential and direct AI impact scores in different

fields, as well as temporal changes in AI impacts on a field. More specifically, we dig deeper into

8
fields with “higher-direct-impact” than typical and those with “higher-potential-impact” than

typical. We also compare fields that underwent “upshifts” in AI benefits, going from “low-AI-

benefit” to “high-AI-benefit” over time. For both cases, we find that those “higher-direct-impact”

fields and those fields that experienced “upshifts” in AI benefits tend to feature more cross-domain

collaborations with AI researchers and higher education levels in AI (see Supplementary Notes

4.3 for details on measures and statistics). Taken together, these findings highlight the importance

of teamwork and cross-domain collaborations amidst AI’s potentially increasing impact on

scientific research and the narrowing of individual domain expertise across the sciences104-106.

Demographic disparities

As the connection between AI and scientific research deepens, it is important to understand who

benefits from AI, which has implications for the equity and sustainability of the research enterprise.

Here we study the gender and racial/ethnic composition of each discipline and further examine

potential differences in the distribution of the benefits from AI across demographic groups.

Specifically, we leverage the de-identified Survey of Doctorate Recipients (SDR) data to solicit

demographic information on U.S.-trained doctoral scientists and engineers by the discipline of

doctorate, sex, and race/ethnicity. We then crosswalk the SDR disciplines of doctorate to the

disciplines in the MAG data to estimate the share of women scientists and underrepresented

minorities (URM) scientists in each discipline (see Methods and Supplementary Note 7 for details).

We find a negative correlation between the share of women scientists within each discipline and

its AI impact scores for both the direct impact (Pearson’s r = –0.555; P-value = 0.032; Fig. 4a)

and the potential impact (Pearson’s r = –0.593; P-value = 0.020; Fig. 4b). Aggregating the AI

impact scores of all disciplines by their gender composition (see Methods for details), we find that

women scientists tend to be associated with a smaller score and thus less benefit from AI (Figs.

9
4cd). Studying the racial and ethnic composition across disciplines, we find another negative

relationship between the share of URM scientists in each discipline and its AI impact score, a

pattern that is again robust for both the direct impact of AI (Pearson’s r = –0.734; P-value = 0.002;

Fig. 4e) and its potential impact (Pearson’s r = –0.711; P-value = 0.003; Fig. 4f). This pattern

appears especially strong for Black scientists within the URM group (Figs. 4gh). On average,

women and URM researchers benefit less from AI. We further performed career-level analysis,

looking at what happens when one starts to engage in AI research (see Supplementary Note 9 for

details). We find that while the average hit rate of a researcher’s papers tends to increase

immediately after engaging in AI research, this citation premium is less concentrated among

underrepresented groups, with women and URM researchers appearing to profit less from AI

engagement compared with their counterparts. Together, these results suggest that while AI has

the potential to bring benefits to all disciplines, the benefits may be distributed unequally across

demographic groups. Hence as AI’s impact on science continues to grow, these unequal career

effects may further amplify existing disparities in science107,108.

Discussion

In this study, we develop a measurement framework to estimate the extent to which AI may benefit

scientific research, aiming to quantify both the direct impact of AI and the potential impact of AI

across a range of scientific disciplines and research fields. Here we use impact to broadly represent

the benefits of AI to science. We find that scientific disciplines are increasingly using AI, as

proxied by the mention of AI-related terms in publication titles and abstracts, with especially sharp

growth in recent years. Publications that use AI tend to see a citation premium, as they are more

likely to be cited both within and outside their disciplines. While there is substantial heterogeneity

in the impact of AI across different disciplines, almost every discipline includes some subfields

that see great benefits from AI. For example, the medicine discipline as a whole is not ranked

10
among the highest in terms of AI benefits, but some of its subfields (e.g., “nuclear medicine,”

“optometry,” and “medical physics”) show substantial AI benefits (Supplementary Fig. 8). Overall,

these results suggest that the benefits that AI may bring to scientific research are widespread across

a range of disciplines and fields, potentially extending beyond the current uses of AI in science.

A systematic understanding of the impact of AI on scientific research may better inform science

and education policy. Our research suggests that the impact of AI on scientific disciplines has

raced ahead across science, facilitated in part by cross-discipline collaborations, while the

educational focus on AI to upskill future scientists within each discipline has lagged. This

misalignment between the impact of AI and AI education (i.e., the AI impact–AI training gap) has

important implications for best practices in preparing next-generation scientists to fully leverage

AI advances. While these analyses are correlational by nature, they support the hypothesis that

collaboration between domain experts and AI researchers may represent an important way to

facilitate the use of AI across science. They also suggest a further benefit of increasing AI training

across disciplines, which would likely help the disciplines to develop domain-specific AI expertise,

allowing them to enjoy greater and timelier benefits from AI advances.

It is also important to recognize that as AI becomes increasingly capable of performing research

tasks, it may have an unequal impact on the research workforce. There are long-standing concerns

about demographic disparities in science109-111. Our results suggest that the groups that have been

historically underrepresented in science are also the groups that may benefit less from AI in

scientific research. These results are somewhat expected, given that gender disparities tend to

correlate with technical fields, which tend to be dominated by men112,113. Nevertheless, our

analysis highlights that as AI plays more important roles in accelerating science, it may exacerbate

existing disparities in science, with implications for building a diverse, equitable, and inclusive

11
research workforce. It thus underscores the importance of expanding the AI-related professoriate

by broadening participation and opportunities in AI research and increasing funding and

educational programs targeted toward women and underrepresented groups in AI-related fields114.

While this study takes an initial step toward quantifying the impact of AI on scientific research, it

has several limitations that are important to consider when interpreting the results. First, our

analyses build on the future of work literature and rely on publication and patent data. Given its

multidimensional nature, however, the benefits of AI for science may go beyond the advantages

that can be estimated from such datasets. These frameworks, in fact, may underestimate the full

range of benefits that AI may bring to research. AI may, for example, optimize the research process

by powering new tools and systems that improve the efficiency of doing science, including

improving access to information, reducing the knowledge burden, guiding human intuition,

automating routine research tasks, and more115,116. Second, AI research evolves rapidly, suggesting

the need for continuous monitoring and updates to the estimates of its benefits to science. As our

datasets trace publications and patents to the end of 2019, they cannot fully capture newer

developments, such as the recent rise of foundation models in AI research117-120. Given that these

foundation models, such as large language models, can be adapted to a wide range of downstream

tasks through fine-tuning, it is conceivable that they may play a significant role in augmenting

research. Third, as a general-purpose technology71,121, AI may generate downstream spillover

effects, with indirect impacts on various domains. For example, by discovering faster matrix

multiplication algorithms122, AI may have indirect effects on disciplines that would benefit from

such advances. Fourth, although the direct mention of AI n-grams in publication titles and abstracts

is suggestive of the use of AI in research, the same n-gram may have different meanings in different

contexts. Also, the same AI capability may bring different benefits to different fields, amidst

alternative ways to define AI terms (see Supplementary Note 8.1 for details)63, suggesting fruitful

12
future directions to further improve our frameworks for understanding AI capabilities and their

impacts on scientific research. Lastly, as AI’s capabilities and its impact on science continue to

grow, it will become ever more crucial to understand the impact of AI on fairness and equity in

research16,123. Equally important is to understand how AI may introduce potential biases or

otherwise create unintended consequences in the genesis of scientific knowledge, especially given

the “black box” nature of many leading AI tools124-127.

Overall, these findings based on large-scale quantitative analyses may prove useful to the AI

research community, helping us better understand the AI capabilities that may be most fruitful for

scientific research. At the same time, the misalignment between the level of AI education and its

impact on research suggests that collaborations between domain experts and AI researchers may

be especially productive, bridging deep domain expertise and new AI advances. Given that

tomorrow’s technological developments often begin upstream from basic scientific research128-130,

a more robust understanding of the impact of AI on science may further inform a range of

important policy considerations for the future of education, research, and innovation2-4.

Methods

Data sources. To estimate the impact of AI on science, we use a variety of datasets that include
information regarding scientific publications, patents, course syllabi, and the demographics of
researchers (see Supplementary Note 1 for details). We introduce two primary datasets. (1) We
use the Microsoft Academic Graph (MAG) database for publication data. We collect information
on 74.6 million publications between 1960 and 2019 of various types (“journal,” “conference,”
“book,” or “book chapter”). These publications are categorized into 19 disciplines (e.g., “computer
science”) and 292 fields (e.g., “machine learning”) under the MAG “field of study” taxonomy, in
which one discipline contains several child fields (see Supplementary Note 1.1 for details). For
each publication, we collect the title, abstract, year, discipline, and field information. (2) We use
PatentsView for patent data. We collect information on 7.1 million patents granted between 1976
and 2019 from PatentsView, a data platform based on bulk data from the U.S. Patent and

13
Trademark Office (USPTO). Each patent is associated with a list of patent classification codes and
keywords. Using these codes and keywords, we identify AI-related patents (see Supplementary
Note 1.2 for details). Together, the MAG publication data and USPTO patent data allow us to
estimate the direct and potential impacts of AI on disciplines and fields.

We supplement the analysis with two more datasets to examine the alignment of the impact of AI
with the level of AI education and to study the gender, racial, and ethnic composition in science.
(1) We use syllabus data that is sourced from the Open Syllabus Project (OSP), the world’s first
large-scale database of university course syllabus documents. Our syllabus dataset contains 4.2
million English-language syllabi published between 2000 and 2018 (see Supplementary Note 1.3
for details). Each syllabus is associated with a list of Classification of Instructional Programs (CIP)
codes representing its academic fields and a list of referenced publications103. We manually
crosswalk CIP codes to MAG fields, and we link syllabus references to MAG publications (see
Supplementary Note 5.1 for details). (2) We use the Survey of Doctorate Recipients (SDR) for de-
identified demographic data regarding individuals with a U.S. research doctoral degree in a science,
engineering, or health field. We use the 2017 SDR data on scientists and engineers, including the
discipline of their doctorate, their sex, and their race and ethnicity. We manually crosswalk the
SDR doctorate disciplines to the MAG disciplines (see Supplementary Note 7.1 for details).

Calculation of AI impact scores. We estimate the direct impact of AI by implementing the “AI
n-gram framework.” Specifically, following prior studies91, we identify AI-related publications
using the five MAG field categories (“machine learning,” “artificial intelligence,” “computer
vision,” “natural language processing,” and “pattern recognition”). Because MAG used a topic
modeling approach to label each paper’s field categories, the AI-related publications identified
here go beyond the explicit mention of these five keywords. Identifying AI research from large-
scale publication databases remains a challenging task, but the simple approach we use balances
precision and recall in determining AI publications (see Supplementary Note 8.1 for details). There
are also other ways to identify AI research (see Supplementary Note 2.1 for details), and our results
are robust under these alternative approaches (see Supplementary Note 8 for details).

From the titles and abstracts of AI publications, we extract n-grams (bigrams and trigrams; e.g.,
“deep learning” and “deep neural network”) and normalize them by lemmatizing and standardizing
words. From these normalized n-grams, we filter AI n-grams using a list of topics under the five

14
AI field categories in the MAG “field of study” taxonomy. This taxonomy is constructed primarily
based on Wikipedia topics (see Supplementary Note 1.1 for details). We calculate the frequency
of AI n-grams per paper to approximate cumulative AI advances. Formally, the AI n-gram
frequency vector at year 𝑡 is 𝐺#!"
# #
= 𝐺!" #
/𝑁!" #
, where 𝐺!" is the vector that summarizes the counts of
#
AI n-grams extracted from AI publications before year 𝑡, and 𝑁!" is the number of AI publications.
We repeat this process for publications in each field to extract n-grams (both AI n-grams and non-
AI n-grams), and we calculate their frequency to approximate current field development. For
example, the biology n-gram frequency vector at year 𝑡 is 𝐺#$# = 𝐺$# /𝑁$# , where 𝐺$# is the vector
that summarizes the count of n-grams extracted from biology publications at year 𝑡, and 𝑁$# is the
number of these biology publications. The coordinate of the same AI n-gram in the biology
frequency vector and the AI frequency vector is the same. In other words, each coordinate of 𝐺#$#
represents one n-gram, where AI n-grams have the same coordinates as those in 𝐺#!"
#
. Finally, we
calculate the direct AI impact score for biology at year 𝑡 based on the frequency of AI n-grams:

𝑆%# = ∑ 𝐺#$# ∙ 𝐺#!"

#
, (1)
where the symbol “∑∙” represents the dot product of the biology frequency vector and the AI
frequency vector of the same AI n-grams. In this calculation, only their common n-grams in these
two vectors are considered, and the same n-gram has the same coordinate in these two vectors.
There are other ways to calculate the direct AI impact score, and our results are largely robust
under some alternative calculations (see Supplementary Note 8.2 for details). A larger direct AI
impact score indicates that AI has a stronger impact on the field.

We estimate the potential impact of AI by implementing the “AI capability–field task framework,”
which is built on the future of work literature82-84. It assumes that research fields may potentially
benefit from AI if their basic tasks are aligned with AI capabilities (see Supplementary Note 3 for
details on the underlying assumptions). We predict the capabilities of AI (i.e., what AI can do) by
extracting verb-noun pairs (e.g., “learn representation”) from the titles of AI publications and AI
patents using a dependency parsing algorithm developed in NLP (see Supplementary Note 3.2 for
details)99-101. Here, following the prior work82, we only use titles because they have a higher signal-
to-noise ratio than the other text fields. After normalizing verb-noun pairs through lemmatization
and standardization, we calculate their relative frequency to approximate AI capabilities.
Specifically, the AI capability frequency vector for AI papers at year 𝑡 is 𝑃𝑎𝑝𝑒𝑟(𝐶3!"
#
)=
# ∑ # #
𝐶!" / 𝐶!" , where 𝐶!" is the vector that summarizes the counts of verb-noun pairs extracted from

15
AI publications before year 𝑡. We repeat this process for AI patents and calculate the vector
𝑃𝑎𝑡𝑒𝑛𝑡(𝐶3!"
#
). By taking an average of common verb-noun pairs in these two vectors, we calculate
the AI capability frequency vector 𝐶3!"
#
to approximate cumulative AI capabilities at year 𝑡:
𝐶3!"
#
= 6𝑃𝑎𝑝𝑒𝑟7𝐶3!"
#
8 + 𝑃𝑎𝑡𝑒𝑛𝑡7𝐶3!"
#
8:/2, (2)
where the symbol “+” represents summing up the frequencies of the same verb-noun pair in the
two vectors. Analogously, we predict the basic tasks of a research field (i.e., what the field does)
by extracting verb-noun pairs from the titles of publications in the field. Taking the biology field
as an example, the field task frequency vector at year 𝑡 is given by 𝑇#$# = 𝑇$# / ∑ 𝑇$# , where 𝑇$# is the
vector that summarizes the counts of verb-noun pairs extracted from biology publications at year
𝑡. In the calculation, we apply the term frequency-inverse document frequency (tf-idf) to discount
the weight of commonly appearing verb-noun pairs in both AI capability and field task vectors
(see Supplementary Note 3.3 for details). Finally, we calculate the potential AI impact score of
biology at year 𝑡 based on the alignment between its tasks and AI capabilities:
∑ ()!" ∙+,#$
"
𝑆&# = , "
∑ +#$ ∙+#$, " , (3)

where the symbol “∑∙” represents the dot product of the AI vector and the biology vector, and the
denominator is applied to normalize the score for comparison across time. In the calculation, only
common verb-noun pairs in the AI vector and the biology vector are considered, and the same
verb-noun pair has the same coordinate in the two vectors. A larger potential AI impact score
means that AI is predicted to have a stronger impact on the field. There are other ways to calculate
the potential AI impact score, and our findings are largely robust under some alternative
calculations (see Supplementary Note 8.3 for details).

Estimation of AI education levels. We measure the level of AI education in each discipline by

leveraging OSP syllabus data and MAG publication data. This measure assumes that a discipline
has a higher AI education level if a larger fraction of publications referenced by syllabi in the
discipline are AI publications. The OSP dataset categorizes course syllabi by educational fields
and provides a link from syllabi to their referenced publications. As syllabi with more references
more likely correspond to graduate-level or research-oriented courses, we only use syllabi with at
least five references and those in the recent period 2014-2018. First, we crosswalk the taxonomies
of educational disciplines and academic disciplines by mapping OSP fields to MAG disciplines,
and we match syllabi-referenced publications to MAG publications using the digital object
identifier (DOI), title, and year (see Supplementary Note 5.1 for details). From these publications,

16
we identify AI publications based on the MAG five AI field categories (see Supplementary Note
2.1 for details). We then estimate a discipline’s AI education level by calculating the fraction of
citations in the discipline’s syllabi that are citations to AI publications (see Supplementary Note
5.2 for details). As robustness checks, we also use syllabi with at least ten references, calculate an
alternative measure for the level of AI education defined as the fraction of a discipline’s syllabi
that cites at least one AI publication, and repeat the analysis for different time periods between
2000 and 2018 (see Supplementary Note 5.2 for detailed methods and additional results).

Calculation of cross-discipline collaborations on AI. We estimate the level of cross-discipline

collaborations on AI research between domain experts and AI researchers using AI publications
in each discipline other than computer science. Specifically, we first assign a primary discipline to
each researcher based on the discipline in which they published most frequently in the period 1960-
2019 and treat authors whose primary discipline is computer science (CS) as AI researchers (see
Supplementary Note 6 for more details). We then categorize each AI publication in a discipline
into one of the four co-authorship types based on its authors’ primary disciplines: (1) “domain &
CS,” which involves both domain experts and computer scientists; (2) “domain sole,” which
involves only domain experts; (3) “CS sole,” which involves only computer scientists; and (4)
“others,” which involves neither domain experts nor computer scientists. Next, we calculate the
share of collaborative AI publications (i.e., those in the “domain & CS” type) for each discipline
(see Supplementary Note 6.1 for detailed methods). Here the calculation only considers AI
publications with at least two authors that were published in the period 1980-2019. As robustness
checks, we also use an alternative approach to identify primary AI researchers (see Supplementary
Note 6.3 for detailed methods and results).

Data Availability
The MAG data are available at https://zenodo.org/record/6511057. The USPTO patent data are
available at https://patentsview.org. The OSP dataset is available from the paper at
https://www.pnas.org/doi/10.1073/pnas.1804247115. The SDR data are available at
https://www.nsf.gov/statistics/srvydoctoratework, and the datasets used in this study are de-
identified, containing only summary statistics for each discipline. The data necessary to reproduce
all main plots in this paper are available at https://kellogg-cssi.github.io/ai4science.

Code Availability
Data are linked and analyzed with customized code in Python 3 using standard software packages
within these programs. The code necessary to reproduce all main plots and statistical analyses is
available at https://kellogg-cssi.github.io/ai4science.

17
Acknowledgements
We thank Yian Yin, Yifan Qian, Binglu Wang, Nima Dehmamy, Lav Varshney, Lili Miao, Alyse
Freilich, and all members of the Center for Science of Science and Innovation (CSSI) at
Northwestern University for helpful discussions. This work was supported by the Air Force Office
of Scientific Research FA9550-19-1-0354, the National Science Foundation (SBE 1829344 and
TIP 2241237), the Alfred P. Sloan Foundation (G-2019-12485), and the Peter G. Peterson
Foundation (21048).

Author Contributions
J.G. and D.W. conceived the idea. D.W. supervised the project. J.G. collected data and performed
analyses. J.G. and D.W. analyzed the results, interpreted the findings, and wrote the paper.

Competing Interests
The authors declare no competing interests.

18
Figures
a The “AI n-gram framework” b AI publications c support vector machine
wavelet transform
machine learning
convolutional neural network
15 machine learning
AI papers

deep pattern deep convolutional neural network

genetic algorithm deep learning
learning recognition learning

AI n-gram per paper (%)

deep learning image processing feature extraction
structure deep pattern deep neural network feature extraction support vector machine
prediction learning recognition
artificial intelligence image segmentation deep neural network
10 random forest component analysis feature selection
artificial neural network artificial neural network artificial intelligence
(1) N-gram extraction generative adversarial network
feature selection learning model
recurrent neural network
recognition
learning
prediction
pattern
structure
deep

pattern recognition object detection

reinforcement learning
5 color image computer vision
[

3/6 machine learning image processing

2/6 digital image point cloud
[3 1 2] image retrieval random forest
1/6 (3) Direct AI
image sequence image segmentation
impact score 0
]

(2) Normalization 2000 2005 2010 2015 2019 decision tree natural language
image quality image quality
[

biology economics
[3 2 2 1]
3/8 d computer vision artificial neural network

machine learning

machine learning
deep learning

artificial intelligence
random forest

artificial neural network

2/8 feature vector generative adversarial network

convolutional neural network

random forest
artificial neural network

deep learning
prediction

0.6
AI n-gram p.p. (%)
structure

structure

principal component analysis recurrent neural network

learning

2/8
protein

protein
folding

recognition system component analysis

deep

1/8
edge detection decision tree
0.4
]

bayesian network reinforcement learning

(1) N-gram extraction natural language sentiment analysis
Biology papers

0.2 motion estimation term memory

protein structure protein deep object detection
structure prediction folding learning 2005 2010 2015 2019 feature vector
recurrent neural network wavelet transform
deep protein deep structure learning model facial expression
learning folding learning prediction 0.0
2000 2010 2019 2000 2010 2019 point cloud
Publication year principal component analysis
Pub. year Pub. year term memory pattern recognition

e philosophy f mathematics g computer science engineering physics

political science geology 37% 24% 24%

Direct AI impact (trend)

art chemistry
business sociology
geography economics 2000 2010 2019 2000 2010 2019 2000 2010 2019
mathematics philosophy psychology economics biology
sociology art 24% 14% 22%
Scientific discipline
Scientific discipline

engineering biology
history history
geology political science 2000 2010 2019 2000 2010 2019 2000 2010 2019
physics materials science h 1.5
All disciplines
30%
sociology
medicine environmental science
Direct AI impact (%)

psychology medicine
computer science engineering 1.0
economics psychology 2000 2010 2019
environmental science geography political science
biology business 0.5 27%
chemistry physics
materials science computer science
0.0
0 1 2 3 4 0.8 1.0 1.2 1.4 1.6 2000 2005 2010 2015 2019 2000 2010 2019
Ratio (Hit_AI / Hit_non_AI) Ratio (SOC_AI / SOC_non_AI) Publication year Publication year

Fig. 1. Measuring the direct impact of AI on scientific research. (a) The “AI n-gram framework”
for estimating the direct impact of AI. First, AI-related publications are identified by the MAG
five AI fields. Then, n-grams are extracted from the titles and abstracts of AI publications. Next,
the frequency of AI n-grams per paper is calculated after lemmatization and normalization.
Similarly, n-grams are extracted for publications in each field, and the frequency of n-grams per
paper is calculated. Finally, a field’s direct AI impact score for a year is calculated by the dot
product of the frequency of AI n-grams cumulated up to the year and the field’s n-grams at the
year. (b) The frequency of ten AI n-grams in 2019 and the trend in use of these n-grams over the
past two decades. (c) Temporal changes in the rankings of the top 30 AI n-grams in 2019. AI n-
grams are presented in rainbow color order, according to their ranking in 2019. (d) The frequency
over the 2000-2019 time period of the top five AI n-grams in biology and economics in 2019. (e)
The ratio of the hit rate of AI-using papers over non-AI-using papers. Here AI-using papers are
identified as those that mention at least one AI n-gram, and the hit rate of papers (Hit) is defined
as the likelihood that a paper is in the top 5% by total citations within the same field and year. (f)
The ratio of the share of outside-field citations (SOC) for AI-using papers over that for non-AI-
using papers. (g) Temporal trends in the direct AI impact scores of disciplines as shown by solid
color lines. The dashed color line shows the score calculated using each discipline’s yearly n-
grams and AI n-grams fixed at 2015. The percentage change comparing the two scores in 2019 is
shown. Each plot uses its y-axis scale to illustrate the relative change best. (h) The direct AI impact

19
scores of disciplines using the same y-axis scale. Colored lines correspond to disciplines in panel
(g), and gray dotted lines represent other disciplines (see Supplementary Fig. 7 for detailed results).

a The “AI capability−field task framework” b 0.5 AI publications d pattern recognition

Top 20 (all) Top 100 fields Top 20 (non-CS)
nuclear medicine
detect object learn model computer vision industrial engineering

Share of AI capability (%)

AI papers &

analyze predict recognize

AI patents

generate image analyze data artificial intelligence control engineering

data structure pattern 0.4 machine learning process engineering
track object learn representation data mining remote sensing
predict analyze predict
structure data structure process image generate model computer engineering engineering drawing
0.3 pose estimation extract feature operating system biological system
real time computing reliability engineering
natural language processing statistics
(1) Verb-noun extraction 0.2
algorithm operations management
speech recognition systems engineering
information retrieval medical physics
analyze
structure

recognize
data

pattern
predict

nuclear medicine electronic engineering

0.1 embedded system automotive engineering
industrial engineering manufacturing engineering
[

3/6 control engineering marine engineering

0.0 process engineering cartography
[3 2 1] 2/6 remote sensing geodesy
2000 2005 2010 2015 2019
1/6
(3) Potential AI engineering drawing radiology
impact score Publication year computer hardware mechanical engineering
]

(2) Normalization
c 0.5
Biology subfields e 100
Pearson's r = 0.891
f 0.8
[

Direct AI impact (percentile)

3/8 biological system fishery P-value < 0.001
[3 2 2 1]
2/8 comput. biology agronomy 80

Share of disciplines
0.4 0.6
Potential AI impact
proliferation

animal science anatomy

2/8
structure

pathway
regulate
analyze

biotechnology agricultural sci.

predict

1/8 60
inhibit

0.3 food science horticulture

data

0.4
0.2 40
(1) Verb-noun extraction
Biology papers

0.2
predict regulate predict analyze 0.1 20
structure pathway structure data
inhibit analyze regulate predict
proliferation data pathway structure 0.0 0 0.0
2000 2005 2010 2015 2019 0 20 40 60 80 100 0 1 2 3
Publication year Potential AI impact (percentile) #overlapped fields (top 3)

g Discipline's subfields political science sociology biology economics

direct AI impact

0 25 50 75 100 0 25 50 75 100 0 25 50 75 100 0 25 50 75 100

psychology physics engineering computer science

potential AI impact

0 25 50 75 100 0 25 50 75 100 0 25 50 75 100 0 25 50 75 100 0 25 50 75 100

Ranking percentile Ranking percentile Ranking percentile Ranking percentile Ranking percentile

Fig. 2. Measuring the potential impact of AI and discipline heterogeneity. (a) The “AI
capability–field task framework” for estimating the potential impact of AI. First, AI capabilities
are inferred by extracting verb-noun pairs from the titles of AI publications and AI patents using
a dependency parsing algorithm. Then, field tasks are inferred from publications in each field using
the same method. Next, the potential AI impact score is determined by calculating the overlap
between field tasks and AI capabilities, after discounting the frequency of commonly appearing
verb-noun pairs. (b) The frequency of ten AI verb-noun pairs in 2019 and their temporal trends
over the past two decades. (c) The top ten subfields of biology according to potential AI impact in
2019 and their temporal trend over the past two decades. The “biological system” field is
consistently ranked first among all subfields. (d) The top 100 fields by potential AI impact in 2019
are shown with color-coded lines. The top 20 fields are listed on the left, with the five AI fields in
orange, computer science (CS) fields in dark blue, and others in gray. The top 20 non-CS fields
are listed on the right, with “biological system” in red and ranked seventh. (e) The strong
correlation between the direct AI impact and potential AI impact of research fields based on their
percentiles. Linear fit (centre line) with 95% confidence intervals (error bands) is shown. (f) A
large overlap among the top three subfields for each discipline by direct AI impact and potential
AI impact scores. Most disciplines exhibit three overlapped subfields (see Supplementary Table
S1 for details). (g) The substantial heterogeneity of AI’s impact within scientific disciplines. As
illustrated by the legend on the left, each plot shows the percentiles of a discipline’s subfields,
where the percentiles based on the direct AI impact score are in the upper row and those based on
the potential AI impact score are in the lower row. Ten disciplines are presented for illustration;
all other disciplines are shown in Supplementary Fig. 8.

20
a 4 Pearson's r = 0.694
b 4 Pearson's r = 0.609
c

Others (3.7%)
P-value = 0.002 P-value = 0.009
uter uter CS sole
comp nce
Level of AI education (%)

Level of AI education (%)

comp nce

Share of AI papers (%)

3 Pearson's r = 0.493 scie 3 Pearson's r = 0.263 scie (21.6%)

shown in a treemap
P-value = 0.074 cs P-value = 0.363 matic
s
emati mathe
math
Domain sole
2 2
(42.1%)

g
eerin eering
1 engin 1 engin
Domain & CS
y ics physics (32.6%)
philosoph biology phys biology
hology
political medicine psyc
hology philosophy psyc geography
geography art medicine
sci. art chemistrygeology geology
0 cio logy 0 polit ical chemist ry business
so ics
history business econom history sci. sociology economics
−4 −3 −2 −1.5 −1.0 −0.5 Type of coauthorship
Direct AI impact (log) Potential AI impact (log) (each block)

d g
e f
0.4 Pearson's r = 0.841 rin 0.4 Pearson's r = 0.802 0.5 engineering physics
e
P-value < 0.001 g ine P-value < 0.001 mathematics psychology
en s ng
tic eri s
Share of collab. AI papers

Share of collab. AI papers

Share of collab. AI papers

medicine
ma gine atic
the en hem ics
ma gy at hys y 0.4
0.3 in e cs ics olo 0.3 ine y m p g
dic
i
m ys ge ic
d aph olo
o h e ge
me con p ph
y m ogr ics
s
y e y ra e m es
log iolog og tal y g ono
sin
i o e en log c 0.3
c
sohy s
b g
nm bio logy e als buy
l s r o o i g
0.2 op es ria e vi ce 0.2
so
ci ter nce holo ntal
os sin te nc y encien ma cie yc e
hil bu mascie olog s hy al s y ps ironm
a l p h y s op t i c
li nc istre v nce
c i. yc tr l o
liti sc ps emis ph
i o
p cie em n
e cie 0.2
po s ch s
ryart ch
0.1 his
to
0.1 arty
tor
his 0.1
−4.5 −4.0 −3.5 −3.0 −2.5 −1.6 −1.4 −1.2 −1.0 −0.8 −0.6 1990 2000 2010 2019
Direct AI impact (log) Potential AI impact (log) Publication year

Fig. 3. Misalignment between AI education and AI impact, but growing knowledge demand
for AI. (a) The correlation between the direct AI impact score and the AI education level that is
estimated by the share of syllabus references to AI publications. Linear fits (centre lines) with 95%
confidence intervals (error bands) are shown. The red line shows that the correlation loses
significance when excluding the three disciplines with the largest AI impact scores: computer
science, engineering, and mathematics. (b) The correlation between the potential AI impact score
and the AI education level. Linear fits (centre lines) with 95% confidence intervals are shown. (c)
The treemap chart shows the share of AI publications by four co-authorship types, where “domain
& CS” represents collaborative AI publications by domain experts and computer scientists,
“domain sole” represents AI publications by domain experts only, “CS sole” represents AI
publications by computer scientists only, and “others” represents AI publications that are neither
by domain experts nor by computer scientists. Here only AI publications in disciplines other than
computer science with at least two authors are considered. (d) The positive correlation between
the direct AI impact score and the share of collaborative (“domain & CS”) AI publications in each
discipline. Linear fit (centre line) with 95% confidence intervals (error bands) is shown. (e) The
positive correlation between the potential AI impact score and the share of collaborative (“domain
& CS”) AI publications in each discipline. Linear fit (centre line) with 95% confidence intervals
(error bands) is shown. (f) The share of collaborative AI publications (“domain & CS”) in five
disciplines across the period 1990-2019. Curves are smoothed by taking a three-year moving
average. Results for other disciplines are shown in Supplementary Fig. 14.

21
a co
c e c
g
−2.0 sci mput Pearson's r = − 0.555 0.4 −2.0 s om Pearson's r = − 0.734 0.4
enc er P-value = 0.032 macien pute P-value = 0.002
e the ce r
m
Direct AI impact (log)

Direct AI impact (log)

ma

Direct AI impact (%)

Direct AI impact (%)

eng athe en
gin tic
−2.5 geo neer matic
i 0.3 −2.5
ee se 0.3
gra ing s ph rin sc nvir
phy e ys g ien on
phy
sic scni viron psy ge c
ics ce me
n
e
−3.0 s c nce me cho
n l
−3.0 og he ec pstal URM
ma ecohemis bi tal ogy 0.2 rap mi b o
ma h str iol nom m cho
y 0.2
sci eria nom try ology medi
t sc ter y y ogy ics edi logy
enc ls ics cin ien ia
−3.5 e
soc
e −3.5 ce ls socine
cio
iolo log
phi g y 0.1 ph y 0.1
los po ilo
po oph sc litic soph
−4.0 sci litical y −4.0 ien al y
enc ce
e
0.0 0.0
0.2 0.4 0.6 en 0.08 0.12 0.16 . r k
Me
n ian e e
hit URM ativ Hisp Othe Blac
Share of women om Share of URM As W N
W

b co
d f co
h
−0.4 Pearson's r = − 0.593 0.3 −0.4 Pearson's r = − 0.711 0.3
s mp sc mp
ma cienc uter P-value = 0.020 ma ienc uter P-value = 0.003
Potential AI impact (log)

Potential AI impact (log)

the e t he e e
eng matic enviro en ma nvir
Potential AI impact

Potential AI impact
gin tic on
in s nm ee me
phy eerin sci enta m rin ps s
g ych scie nta
sic g enc l ge
−0.8 s e geog e edi
cin
0.2 −0.8 ph ogra olo nce l 0.2
con rap
ma o hy psy e m ysi ph eco gy URM
sciterials ch ics
m
cho sc atercs y no
mi
me
enc em bio log ien ia ch cs dic
e istr log y ce ls em i
so ne
y y ist cio
bio ry log
−1.2 pol soc 0.1 −1.2 log po 0.1
sci itica iolo y sc litic y
enc l gy ien al
e p ph c
hilo ilo e
sop so
hy ph
y
−1.6 0.0 −1.6 0.0
0.2 0.4 0.6 en 0.08 0.12 0.16 . r k
Me
n ian e e
hit URM ativ Hisp Othe Blac
Share of women om Share of URM As W N
W

Fig. 4. Gender and racial disparities in the impact of AI across disciplines. (a) The negative
correlation between the direct AI impact score and the share of women scientists in each discipline.
Linear fit (centre line) with 95% confidence intervals (error bands) is shown. (b) The negative
correlation between the potential AI impact score and the share of women scientists in each
discipline. Linear fit (centre line) with 95% confidence intervals (error bands) is shown. (c) The
average direct AI impact scores for women and men scientists. The average score for men/women
is calculated by weighting the direct AI impact score of each discipline by the share of men/women
in the discipline. (d) The average potential AI scores for women and men scientists. (e) The
negative correlation between the direct AI impact score and the share of URM (underrepresented
minority) scientists. The URM category includes “African American or Black,” “American Indian
or Alaska Native,” “Hispanic or Latino,” and “Native Hawaiians or other Pacific Islanders.” Linear
fit (centre line) with 95% confidence intervals (error bands) is shown. (f) The negative correlation
between the potential AI impact score and the share of URM scientists. Linear fit (centre line) with
95% confidence intervals (error bands) is shown. (g) The average direct AI impact score for each
racial and ethnic group. The average score for each group is calculated by weighting the direct AI
impact score of each discipline by its share of the particular racial and ethnic group in the discipline.
The average score for each racial and ethnic group under the URM category is shown separately
on the right. (h) The average potential AI impact score for each racial and ethnic group.

22
References
1 Herbert, A. S. The Sciences of the Artificial. (The MIT Press, 1969).
2 Brynjolfsson, E. & Mitchell, T. What can machine learning do? Workforce implications.
Science 358, 1530-1534 (2017).
3 Agrawal, A., Gans, J. & Goldfarb, A. The Economics of Artificial Intelligence: An Agenda.
(University of Chicago Press, 2019).
4 Autor, D., Mindell, D. A. & Reynolds, E. B. The Work of the Future: Shaping Technology
and Institutions. (MIT Task Force, 2019).
5 Acemoglu, D., Autor, D., Hazell, J. & Restrepo, P. Artificial intelligence and jobs: Evidence
from online vacancies. Journal of Labor Economics 40, S293-S340 (2022).
6 Aghion, P., Jones, B. F. & Jones, C. I. Artificial intelligence and economic growth. In The
Economics of Artificial Intelligence: An Agenda Ch. 9, 237-290 (University of Chicago
Press, 2019).
7 Cockburn, I. M., Henderson, R. & Stern, S. The impact of artificial intelligence on innovation:
An exploratory analysis. In The Economics of Artificial Intelligence: An Agenda Ch. 4, 115-
148 (University of Chicago Press, 2019).
8 Tomasev, N., Cornebise, J., Hutter, F., Mohamed, S., Picciariello, A., Connelly, B., Belgrave,
D. C. M., Ezer, D., van der Haert, F. C., Mugisha, F., Abila, G., Arai, H., Almiraat, H.,
Proskurnia, J., Snyder, K., Otake-Matsuura, M., Othman, M., Glasmachers, T., de Wever, W.,
Teh, Y. W., Khan, M. E., De Winne, R., Schaul, T. & Clopath, C. AI for social good:
Unlocking the opportunity for positive impact. Nature Communications 11, 2468 (2020).
9 Dwivedi, Y. K., Hughes, L., Ismagilova, E., Aarts, G., Coombs, C., Crick, T., Duan, Y.,
Dwivedi, R., Edwards, J., Eirug, A. & others. Artificial Intelligence (AI): Multidisciplinary
perspectives on emerging challenges, opportunities, and agenda for research, practice and
policy. International Journal of Information Management 57, 101994 (2021).
10 Frey, C. B. & Osborne, M. A. The future of employment: How susceptible are jobs to
computerisation? Technological Forecasting and Social Change 114, 254-280 (2017).
11 Acemoglu, D. & Restrepo, P. The race between man and machine: Implications of technology
for growth, factor shares, and employment. American Economic Review 108, 1488-1542
(2018).
12 Khan, H. N., Hounshell, D. A. & Fuchs, E. R. H. Science and research policy at the end of
Moore's law. Nature Electronics. 1, 14-21 (2018).
13 Iansiti, M. & Lakhani, K. R. Competing in the Age of AI: Strategy and Leadership When
Algorithms and Networks Run the World. (Harvard Business Press, 2020).
14 Eshraghian, J. K. Human ownership of artificial creativity. Nature Machine Intelligence 2,
157-160 (2020).
15 Marcus, G. & Davis, E. Rebooting AI: Building Artificial Intelligence We Can Trust.
(Pantheon Books, 2019).
16 Liang, W. X., Tadesse, G. A., Ho, D., Li, F. F., Zaharia, M., Zhang, C. & Zou, J. Advances,
challenges and opportunities in creating data for trustworthy AI. Nature Machine Intelligence
4, 669-677 (2022).
17 Bengio, Y., Hinton, G., Yao, A., Song, D., Abbeel, P., Darrell, T., Harari, Y. N., Zhang, Y.-
Q., Xue, L., Shalev-Shwartz, S., Hadfield, G., Clune, J., Maharaj, T., Hutter, F., Baydin, A.
G., McIlraith, S., Gao, Q., Acharya, A., Krueger, D., Dragan, A., Torr, P., Russell, S.,
Kahneman, D., Brauner, J. & Mindermann, S. Managing extreme AI risks amid rapid progress.
Science 384, 842-845 (2024).

23
18 Frank, M. R., Autor, D., Bessen, J. E., Brynjolfsson, E., Cebrian, M., Deming, D. J., Feldman,
M., Groh, M., Lobo, J., Moro, E., Wang, D., Youn, H. & Rahwan, I. Toward understanding
the impact of artificial intelligence on labor. Proceedings of the National Academy of Sciences,
U.S.A. 116, 6531-6539 (2019).
19 Agrawal, A., Gans, J. S. & Goldfarb, A. Artificial intelligence: The ambiguous labor market
impact of automating prediction. Journal of Economic Perspectives 33, 31-50 (2019).
20 Koebis, N., Starke, C. & Rahwan, I. The promise and perils of using artificial intelligence to
fight corruption. Nature Machine Intelligence 4, 418-424 (2022).
21 Brynjolfsson, E., Li, D. & Raymond, L. R. Generative AI at work. National Bureau of
Economic Research, NBER Working Paper No. w31161 (2023).
22 Noy, S. & Zhang, W. Experimental evidence on the productivity effects of generative artificial
intelligence. Science 381, 187-192 (2023).
23 Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., van den Driessche, G.,
Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe, D.,
Nham, J., Kalchbrenner, N., Sutskever, I., Lillicrap, T., Leach, M., Kavukcuoglu, K., Graepel,
T. & Hassabis, D. Mastering the game of Go with deep neural networks and tree search.
Nature 529, 484-489 (2016).
24 Geirhos, R., Temme, C. R., Rauber, J., Schütt, H. H., Bethge, M. & Wichmann, F. A.
Generalisation in humans and deep neural networks. In Advances in Neural Information
Processing Systems Vol. 31 7538-7550 (The MIT Press, 2018).
25 Grace, K., Salvatier, J., Dafoe, A., Zhang, B. & Evans, O. When will AI exceed human
performance? Evidence from AI experts. Journal of Artificial Intelligence Research 62, 729-
754 (2018).
26 Liu, X., Faes, L., Kale, A. U., Wagner, S. K., Fu, D. J., Bruynseels, A., Mahendiran, T.,
Moraes, G., Shamdas, M., Kern, C. & Ledsam, J. R. A comparison of deep learning
performance against health-care professionals in detecting diseases from medical imaging: A
systematic review and meta-analysis. Lancet Digital Health 1, e271-e297 (2019).
27 Ishowo-Oloko, F., Bonnefon, J. F., Soroye, Z., Crandall, J., Rahwan, I. & Rahwan, T.
Behavioural evidence for a transparency-efficiency tradeoff in human-machine cooperation.
Nature Machine Intelligence 1, 517-521 (2019).
28 Yang, Y., Wu, Y. Y. & Uzzi, B. Estimating the deep replicability of scientific findings using
human and artificial intelligence. Proceedings of the National Academy of Sciences, U.S.A.
117, 10762-10768 (2020).
29 Wurman, P. R., Barrett, S., Kawamoto, K., MacGlashan, J., Subramanian, K., Walsh, T. J.,
Capobianco, R., Devlic, A., Eckert, F., Fuchs, F., Gilpin, L., Khandelwal, P., Kompella, V.,
Lin, H. C., MacAlpine, P., Oller, D., Seno, T., Sherstan, C., Thomure, M. D., Aghabozorgi,
H., Barrett, L., Douglas, R., Whitehead, D., Durr, P., Stone, P., Spranger, M. & Kitano, H.
Outracing champion Gran Turismo drivers with deep reinforcement learning. Nature 602,
223-228 (2022).
30 Maslej, N., Fattorini, L., Perrault, R., Parli, V., Reuel, A., Brynjolfsson, E., Etchemendy, J.,
Ligett, K., Lyons, T., Manyika, J., Niebles, J. C., Shoham, Y., Wald, R. & Clark, J. The AI
Index 2024 Annual Report. (AI Index Steering Committee, Institute for Human-Centered AI,
Stanford University, 2024).
31 Gil, Y., Greaves, M., Hendler, J. & Hirsh, H. Amplify scientific discovery with artificial
intelligence. Science 346, 171-172 (2014).
32 Wang, H. C., Fu, T. F., Du, Y. Q., Gao, W. H., Huang, K. X., Liu, Z. M., Chandak, P., Liu,
S. C., Van Katwyk, P., Deac, A., Anandkumar, A., Bergen, K., Gomes, C. P., Ho, S., Kohli,
P., Lasenby, J., Leskovec, J., Liu, T. Y., Manrai, A., Marks, D., Ramsundar, B., Song, L., Sun,

24
 J. M., Tang, J., Velickovic, P., Welling, M., Zhang, L. F., Coley, C. W., Bengio, Y. & Zitnik,
M. Scientific discovery in the age of artificial intelligence. Nature 620, 47-60 (2023).
33 Carleo, G., Cirac, I., Cranmer, K., Daudet, L., Schuld, M., Tishby, N., Vogt-Maranto, L. &
Zdeborová, L. Machine learning and the physical sciences. Reviews of Modern Physics 91,
045002 (2019).
34 Rahwan, I., Cebrian, M., Obradovich, N., Bongard, J., Bonnefon, J. F., Breazeal, C., Crandall,
J. W., Christakis, N. A., Couzin, I. D., Jackson, M. O., Jennings, N. R., Kamar, E., Kloumann,
I. M., Larochelle, H., Lazer, D., McElreath, R., Mislove, A., Parkes, D. C., Pentland, A.,
Roberts, M. E., Shariff, A., Tenenbaum, J. B. & Wellman, M. Machine behaviour. Nature
568, 477-486 (2019).
35 Jimenez-Luna, J., Grisoni, F. & Schneider, G. Drug discovery with explainable artificial
intelligence. Nature Machine Intelligence 2, 573-584 (2020).
36 Xu, Y. J., Liu, X., Cao, X., Huang, C. P., Liu, E. K., Qian, S., Liu, X. C., Wu, Y. J., Dong, F.
L., Qiu, C. W., Qiu, J. J., Hua, K. Q., Su, W. T., Wu, J., Xu, H. Y., Han, Y., Fu, C. G., Yin,
Z. G., Liu, M., Roepman, R., Dietmann, S., Virta, M., Kengara, F., Zhang, Z., Zhang, L. F.,
Zhao, T. L., Dai, J., Yang, J. L., Lan, L., Luo, M., Liu, Z. F., An, T., Zhang, B., He, X., Cong,
S., Liu, X. H., Zhang, W., Lewis, J. P., Tiedje, J. M., Wang, Q., An, Z. L., Wang, F., Zhang,
L. B., Huang, T., Lu, C., Cai, Z. P., Wang, F. & Zhang, J. B. Artificial intelligence: A powerful
paradigm for scientific research. The Innovation 2, 100179 (2021).
37 Davies, A., Velickovic, P., Buesing, L., Blackwell, S., Zheng, D. I., Tomasev, N., Tanburn,
R., Battaglia, P., Blundell, C., Juhasz, A., Lackenby, M., Williamson, G., Hassabis, D. &
Kohli, P. Advancing mathematics by guiding human intuition with AI. Nature 600, 70-74
(2021).
38 Peng, H., Ke, Q., Budak, C., Romero, D. M. & Ahn, Y. Y. Neural embeddings of scholarly
periodicals reveal complex disciplinary organizations. Science Advances 7, eabb9004 (2021).
39 Krenn, M., Pollice, R., Guo, S. Y., Aldeghi, M., Cervera-Lierta, A., Friederich, P., Gomes, G.
D., Hase, F., Jinich, A., Nigam, A., Yao, Z. P. & Aspuru-Guzik, A. On scientific
understanding with artificial intelligence. Nature Reviews Physics. 4, 761-769 (2022).
40 Belikov, A. V., Rzhetsky, A. & Evans, J. A. Prediction of robust scientific facts from literature.
Nature Machine Intelligence 4, 445-454 (2022).
41 Grossmann, I., Feinberg, M., Parker, D. C., Christakis, N. A., Tetlock, P. E. & Cunningham,
W. A. AI and the transformation of social science research. Science 380, 1108-1109 (2023).
42 Groh, M., Badri, O., Daneshjou, R., Koochek, A., Harris, C., Soenksen, L. R., Doraiswamy,
P. M. & Picard, R. Deep learning-aided decision support for diagnosis of skin disease across
skin tones. Nature Medicine. 30, 573-583 (2024).
43 Bail , C. A. Can Generative AI improve social science? Proceedings of the National Academy
of Sciences, U.S.A. 121, e2314021121 (2024).
44 Alvarez, A., Caliskan, A., Crockett, M. J., Ho, S. S., Messeri, L. & West, J. Science
communication with generative AI. Nature Human Behaviour. 8, 625-627 (2024).
45 Senior, A. W., Evans, R., Jumper, J., Kirkpatrick, J., Sifre, L., Green, T., Qin, C., Žídek, A.,
Nelson, A. W. R., Bridgland, A., Penedones, H., Petersen, S., Simonyan, K., Crossan, S.,
Kohli, P., Jones, D. T., Silver, D., Kavukcuoglu, K. & Hassabis, D. Improved protein structure
prediction using potentials from deep learning. Nature 577, 706-710 (2020).
46 Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O.,
Tunyasuvunakool, K., Bates, R., Zidek, A., Potapenko, A., Bridgland, A., Meyer, C., Kohl,
S. A. A., Ballard, A. J., Cowie, A., Romera-Paredes, B., Nikolov, S., Jain, R., Adler, J., Back,
T., Petersen, S., Reiman, D., Clancy, E., Zielinski, M., Steinegger, M., Pacholska, M.,
Berghammer, T., Bodenstein, S., Silver, D., Vinyals, O., Senior, A. W., Kavukcuoglu, K.,

25
 Kohli, P. & Hassabis, D. Highly accurate protein structure prediction with AlphaFold. Nature
596, 583-589 (2021).
47 Baek, M., DiMaio, F., Anishchenko, I., Dauparas, J., Ovchinnikov, S., Lee, G. R., Wang, J.,
Cong, Q., Kinch, L. N., Schaeffer, R. D., Millan, C., Park, H., Adams, C., Glassman, C. R.,
DeGiovanni, A., Pereira, J. H., Rodrigues, A. V., van Dijk, A. A., Ebrecht, A. C., Opperman,
D. J., Sagmeister, T., Buhlheller, C., Pavkov-Keller, T., Rathinaswamy, M. K., Dalwadi, U.,
Yip, C. K., Burke, J. E., Garcia, K. C., Grishin, N. V., Adams, P. D., Read, R. J. & Baker, D.
Accurate prediction of protein structures and interactions using a three-track neural network.
Science 373, 871-876 (2021).
48 Popova, M., Isayev, O. & Tropsha, A. Deep reinforcement learning for de novo drug design.
Science Advances 4, aap7885 (2018).
49 Zhavoronkov, A., Ivanenkov, Y. A., Aliper, A., Veselov, M. S., Aladinskiy, V. A.,
Aladinskaya, A. V., Terentiev, V. A., Polykovskiy, D. A., Kuznetsov, M. D., Asadulaev, A.
& others. Deep learning enables rapid identification of potent DDR1 kinase inhibitors. Nature
Biotechnology 37, 1038-1040 (2019).
50 Schneider, P., Walters, W. P., Plowright, A. T., Sieroka, N., Listgarten, J., Goodnow, R. A.,
Fisher, J., Jansen, J. M., Duca, J. S., Rush, T. S., Zentgraf, M., Hill, J. E., Krutoholow, E.,
Kohler, M., Blaney, J., Funatsu, K., Luebkemann, C. & Schneider, G. Rethinking drug design
in the artificial intelligence era. Nature Reviews Drug Discovery. 19, 353-364 (2020).
51 Sadybekov, A. V. & Katritch, V. Computational approaches streamlining drug discovery.
Nature 616, 673-685 (2023).
52 Iten, R., Metger, T., Wilming, H., Del Rio, L. d. & Renner, R. Discovering physical concepts
with neural networks. Physical Review Letters 124, 010508 (2020).
53 Seif, A., Hafezi, M. & Jarzynski, C. Machine learning the thermodynamic arrow of time.
Nature Physics 17, 105-113 (2021).
54 Wu, T. L. & Tegmark, M. Toward an artificial intelligence physicist for unsupervised learning.
Physical Review E. 100, 033311 (2019).
55 Lu, L., Jin, P. Z., Pang, G. F., Zhang, Z. Q. & Karniadakis, G. E. Learning nonlinear operators
via DeepONet based on the universal approximation theorem of operators. Nature Machine
Intelligence 3, 218-229 (2021).
56 Han, J., Jentzen, A. & Weinan, E. Solving high-dimensional partial differential equations
using deep learning. Proceedings of the National Academy of Sciences, U.S.A. 115, 8505-
8510 (2018).
57 Raayoni, G., Gottlieb, S., Manor, Y., Pisha, G., Harris, Y., Mendlovic, U., Haviv, D., Hadad,
Y. & Kaminer, I. Generating conjectures on fundamental constants with the Ramanujan
Machine. Nature 590, 67-73 (2021).
58 Degrave, J., Felici, F., Buchli, J., Neunert, M., Tracey, B., Carpanese, F., Ewalds, T., Hafner,
R., Abdolmaleki, A., de las Casas, D., Donner, C., Fritz, L., Galperti, C., Huber, A., Keeling,
J., Tsimpoukelli, M., Kay, J., Merle, A., Moret, J. M., Noury, S., Pesamosca, F., Pfau, D.,
Sauter, O., Sommariva, C., Coda, S., Duval, B., Fasoli, A., Kohli, P., Kavukcuoglu, K.,
Hassabis, D. & Riedmiller, M. Magnetic control of tokamak plasmas through deep
reinforcement learning. Nature 602, 414-419 (2022).
59 Tshitoyan, V., Dagdelen, J., Weston, L., Dunn, A., Rong, Z., Kononova, O., Persson, K. A.,
Ceder, G. & Jain, A. Unsupervised word embeddings capture latent knowledge from materials
science literature. Nature 571, 95-98 (2019).
60 Sanchez-Lengeling, B. & Aspuru-Guzik, A. Inverse molecular design using machine learning:
Generative models for matter engineering. Science 361, 360-365 (2018).

26
61 Chen, C., Zuo, Y. X., Ye, W. K., Li, X. G., Deng, Z. & Ong, S. P. A critical review of machine
learning of energy materials. Advanced Energy Materials. 10, 1903242 (2020).
62 Merchant, A., Batzner, S., Schoenholz, S. S., Aykol, M., Cheon, G. & Cubuk, E. D. Scaling
deep learning for materials discovery. Nature 624, 80-85 (2023).
63 Zheng, S., Trott, A., Srinivasa, S., Naik, N., Gruesbeck, M., Parkes, D. C. & Socher, R. The
AI Economist: Taxation policy design via two-level deep multiagent reinforcement learning.
Science Advances 8, eabk2607 (2022).
64 Koster, R., Jan, B., Tacchetti, A., Weinstein, A., Zhu, T. N., Hauser, O., Williams, D.,
Campbell-Gillingham, L., Thacker, P., Botvinick, M. & Summerfield, C. Human-centred
mechanism design with Democratic AI. Nature Human Behaviour. 6, 1398-1407 (2022).
65 Dunjko, V. & Briegel, H. J. Machine learning & artificial intelligence in the quantum domain:
A review of recent progress. Reports on Progress in Physics 81, 074001 (2018).
66 Sturm, B. L., Ben-Tal, O., Monaghan, Ú., Collins, N., Herremans, D., Chew, E., Hadjeres, G.,
Deruty, E. & Pachet, F. Machine learning research that matters for music creation: A case
study. Journal of New Music Research 48, 36-55 (2019).
67 Rajkomar, A., Dean, J. & Kohane, I. Machine learning in medicine. New England Journal of
Medicine. 380, 1347-1358 (2019).
68 Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M. & Sutskever, I.
Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on
Machine Learning Vol. 139 8821-8831 (ICML, 2021).
69 Epstein, Z., Hertzmann, A. & the Investigators of Human Creativity. Art and the science of
generative AI. Science 380, 1110-1111 (2023).
70 Swanson, K., Liu, G., Catacutan, D. B., Arnold, A., Zou, J. & Stokes, J. M. Generative AI for
designing and validating easily synthesizable and structurally novel antibiotics. Nature
Machine Intelligence 6, 338-353 (2024).
71 Crafts, N. Artificial intelligence as a general-purpose technology: An historical perspective.
Oxford Review of Economic Policy. 37, 521-536 (2021).
72 Bloom, N., Hassan, T. A., Kalyani, A., Lerner, J. & Tahoun, A. The diffusion of new
technologies. National Bureau of Economic Research, NBER Working Paper No. w28999
(2021).
73 Caselli, F. & Coleman, W. J. Cross-country technology diffusion: The case of computers.
American Economic Review 91, 328-335 (2001).
74 Comin, D. & Hobijn, B. An exploration of technology diffusion. American Economic Review
100, 2031-2059 (2010).
75 Zenil, H., Tegnér, J., Abrahão, F. S., Lavin, A., Kumar, V., Frey, J. G., Weller, A., Soldatova,
L., Bundy, A. R., Jennings, N. R., Takahashi, K., Hunter, L., Dzeroski, S., Briggs, A., Gregory,
F. D., Gomes, C. P., Rowe, J., Evans, J., Kitano, H. & King, R. The future of fundamental
science led by generative closed-loop artificial intelligence. arXiv preprint arXiv:2307.07522
(2023).
76 Topol, E. J. High-performance medicine: the convergence of human and artificial intelligence.
Nature Medicine. 25, 44-56 (2019).
77 Hidalgo, C. A., Orghian, D., Albo Canals, J., de Almeida, F. & Martín Cantero, N. How
Humans Judge Machines. (The MIT Press, 2021).
78 Raisch, S. & Krakowski, S. Artificial intelligence and management: The automation–
augmentation paradox. Academy of Management Review. 46, 192-210 (2021).
79 Fjelland, R. Why general artificial intelligence will not be realized. Humanities and Social
Sciences Communications 7, 10 (2020).

27
80 Messeri, L. & Crockett, M. J. Artificial intelligence and illusions of understanding in scientific
research. Nature 627, 49-58 (2024).
81 Kanitscheider, I. & Fiete, I. Training recurrent networks to generate hypotheses about how
the brain solves hard navigation problems. In Proceedings of Advances in Neural Information
Processing Systems 4529-4538 (The MIT Press, 2017).
82 Webb, M. The impact of artificial intelligence on the labor market. Social Science Research
Network, SSRN 3482150 (2019).
83 Kogan, L., Papanikolaou, D., Schmidt, L. D. & Seegmiller, B. Technology, vintage-specific
human capital, and labor displacement: Evidence from linking patents with occupations.
National Bureau of Economic Research, NBER Working Paper No. w29552 (2022).
84 Atalay, E., Phongthiengtham, P., Sotelo, S. & Tannenbaum, D. The evolution of work in the
United States. American Economic Journal: Applied Economics 12, 1-34 (2020).
85 Felten, E. W., Raj, M. & Seamans, R. A method to link advances in artificial intelligence to
occupational abilities. AEA Papers and Proceedings 108, 54-57 (2018).
86 Wu, L., Hitt, L. & Lou, B. W. Data analytics, innovation, and firm productivity. Management
Science 66, 2017-2039 (2020).
87 Brynjolfsson, E., Mitchell, T. & Rock, D. What can machines learn, and what does it mean
for occupations and the economy? AEA Papers and Proceedings 108, 43-47 (2018).
88 Wang, D. & Barabási, A.-L. The Science of Science. (Cambridge University Press, 2021).
89 Fortunato, S., Bergstrom, C. T., Börner, K., Evans, J. A., Helbing, D., Milojević, S., Petersen,
A. M., Radicchi, F., Sinatra, R., Uzzi, B., Vespignani, A., Waltman, L., Wang, D. & Barabási,
A.-L. Science of science. Science 359, eaao0185 (2018).
90 Zeng, A., Shen, Z. S., Zhou, J. L., Wu, J. S., Fan, Y., Wang, Y. G. & Stanley, H. E. The
science of science: From the perspective of complex systems. Physics Reports 714, 1-73
(2017).
91 Frank, M. R., Wang, D., Cebrian, M. & Rahwan, I. The evolution of citation graphs in
artificial intelligence research. Nature Machine Intelligence 1, 79-85 (2019).
92 Miao, L. L., Murray, D., Jung, W. S., Lariviere, V., Sugimoto, C. R. & Ahn, Y. Y. The latent
structure of global scientific development. Nature Human Behaviour. 6, 1206-1217 (2022).
93 Liu, L., Jones, B. F., Uzzi, B. & Wang, D. S. Data, measurement and empirical methods in
the science of science. Nature Human Behaviour. 7, 1046-1058 (2023).
94 Sourati, J. & Evans, J. A. Accelerating science with human-aware artificial intelligence.
Nature Human Behaviour. 7, 1682-1696 (2023).
95 Murray, D., Yoon, J., Kojaku, S., Costas, R., Jung, W. S., Milojevic, S. & Ahn, Y. Y.
Unsupervised embedding of trajectories captures the latent structure of scientific migration.
Proceedings of the National Academy of Sciences, U.S.A. 120, e2305414120 (2023).
96 Krenn, M., Buffoni, L., Coutinho, B., Eppel, S., Foster, J. G., Gritsevskiy, A., Lee, H. R., Lu,
Y. C., Moutinho, J. P., Sanjabi, N., Sonthalia, R., Tran, N. M., Valente, F., Xie, Y. X. Y., Yu,
R. S. & Kopp, M. Forecasting the future of artificial intelligence with machine learning-based
link prediction in an exponentially growing knowledge network. Nature Machine Intelligence
5, 1326-1335 (2023).
97 Sinha, A., Shen, Z., Song, Y., Ma, H., Eide, D., Hsu, B. J. & Wang, K. An overview of
Microsoft Academic Service (MAS) and applications. In Proceedings of the 24th
International Conference on World Wide Web 243-246 (WWW, 2015).
98 World Intellectual Property Organization (WIPO). WIPO Technology Trends 2019 –
Artificial Intelligence. (WIPO, 2019).
99 Nivre, J. & Nilsson, J. Pseudo-projective dependency parsing. In Proceedings of the 43rd
Annual Meeting of the Association for Computational Linguistics 99-106 (ACL, 2005).

28
100 Honnibal, M. & Johnson, M. An improved non-monotonic transition system for dependency
parsing. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language
Processing 1373-1378 (ACL, 2015).
101 Benetka, J. R., Krumm, J. & Bennett, P. N. Understanding context for tasks and activities. In
Proceedings of the 2019 Conference on Human Information Interaction and Retrieval 133-
142 (ACM, 2019).
102 Service, R. Science's 2021 Breakthrough of the Year: Protein structures for all,
<https://www.science.org/content/article/breakthrough-2021> (2021).
103 Börner, K., Scrivner, O., Gallant, M., Ma, S., Liu, X., Chewning, K., Wu, L. & Evans, J. A.
Skill discrepancies between research, education, and jobs reveal the critical need to supply
soft skills for the data economy. Proceedings of the National Academy of Sciences, U.S.A.
115, 12630-12637 (2018).
104 Wuchty, S., Jones, B. F. & Uzzi, B. The increasing dominance of teams in production of
knowledge. Science 316, 1036-1039 (2007).
105 Wu, L. F., Wang, D. S. & Evans, J. A. Large teams develop and small teams disrupt science
and technology. Nature 566, 378-382 (2019).
106 Littmann, M., Selig, K., Cohen-Lavi, L., Frank, Y., Honigschmid, P., Kataka, E., Mosch, A.,
Qian, K., Ron, A., Schmid, S., Sorbie, A., Szlak, L., Dagan-Wiener, A., Ben-Tal, N., Niv, M.
Y., Razansky, D., Schuller, B. W., Ankerst, D., Hertz, T. & Rost, B. Validity of machine
learning in biology and medicine increased through collaborations across fields of expertise.
Nature Machine Intelligence 2, 18-24 (2020).
107 Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K. & Galstyan, A. A survey on bias and
fairness in machine learning. ACM Computing Surveys. 54, 1-35 (2021).
108 Young, E., Wajcman, J. & Sprejer, L. Where are the Women? Mapping the gender job gap in
AI. (The Alan Turing Institute, 2021).
109 Xie, Y. & Shauman, K. A. Women in science: Career processes and outcomes. (Harvard
University Press, 2003).
110 Hoppe, T. A., Litovitz, A., Willis, K. A., Meseroll, R. A., Perkins, M. J., Hutchins, B. I., Davis,
A. F., Lauer, M. S., Valantine, H. A., Anderson, J. M. & Santangelo, G. M. Topic choice
contributes to the lower rate of NIH awards to African-American/black scientists. Science
Advances 5, eaaw7238 (2019).
111 Ginther, D. K., Schaffer, W. T., Schnell, J., Masimore, B., Liu, F., Haak, L. L. & Kington, R.
Race, ethnicity, and NIH research awards. Science 333, 1015-1019 (2011).
112 Sugimoto, C. R., Larivière, V., Ni, C. Q., Gingras, Y. & Cronin, B. Global gender disparities
in science. Nature 504, 211-213 (2013).
113 Huang, J. M., Gates, A. J., Sinatra, R. & Barabási, A. L. Historical comparison of gender
inequality in scientific careers across countries and disciplines. Proceedings of the National
Academy of Sciences, U.S.A. 117, 4609-4616 (2020).
114 The National Network for Critical Technology Assessment (NNCTA). Securing America’s
Future: A Framework for Critical Technology Assessment. (NNCTA, 2023).
115 Cachola, I., Lo, K., Cohan, A. & Weld, D. S. TLDR: Extreme summarization of scientific
documents. In Proceedings of the 2020 Conference on Empirical Methods in Natural
Language Processing 4766-4777 (ACL, 2020).
116 Lew, A., Agrawal, M., Sontag, D. & Mansinghka, V. PClean: Bayesian data cleaning at scale
with domain-specific probabilistic programming. In Proceedings of the 24th International
Conference on Artificial Intelligence and Statistics Vol. 130 1927-1935 (JMLR, 2021).

29
117 Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.
S., Bohg, J., Bosselut, A., Brunskill, E. & Brynjolfsson, E. On the opportunities and risks of
foundation models. arXiv preprint arXiv:2108.07258 (2021).
118 Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma,
M., Zhou, D., Metzler, D., Chi, E. H., Hashimoto, T., Vinyals, O., Liang, P., Dean, J. & Fedus,
W. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682 (2022).
119 Moor, M., Banerjee, O., Abad, Z. S. H., Krumholz, H. M., Leskovec, J., Topol, E. J. &
Rajpurkar, P. Foundation models for generalist medical artificial intelligence. Nature 616,
259-265 (2023).
120 Huang, Z., Bianchi, F., Yuksekgonul, M., Montine, T. J. & Zou, J. A visual-language
foundation model for pathology image analysis using medical Twitter. Nature Medicine. 29,
2307-2316 (2023).
121 Goldfarb, A., Taska, B. & Teodoridis, F. Could machine learning be a general purpose
technology? A comparison of emerging technologies using data from online job postings.
Research Policy 52, 104653 (2023).
122 Fawzi, A., Balog, M., Huang, A., Hubert, T., Romera-Paredes, B., Barekatain, M., Novikov,
A., Ruiz, F. J. R., Schrittwieser, J., Swirszcz, G., Silver, D., Hassabis, D. & Kohli, P.
Discovering faster matrix multiplication algorithms with reinforcement learning. Nature 610,
47-53 (2022).
123 Jobin, A., Ienca, M. & Vayena, E. The global landscape of AI ethics guidelines. Nature
Machine Intelligence 1, 389-399 (2019).
124 Arrieta, A. B., Díaz-Rodríguez, N., Del Ser, J., Bennetot, A., Tabik, S., Barbado, A., García,
S., Gil-López, S., Molina, D., Benjamins, R., Chatila, R. & Herrera, F. Explainable Artificial
Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible
AI. Information Fusion. 58, 82-115 (2020).
125 Lenharo, M. An AI revolution is brewing in medicine. What will it look like? Nature 622,
686-688 (2023).
126 Bockting, C. L., van Dis, E. A. M., van Rooij, R., Zuidema, W. & Bollen, J. Living guidelines
for generative AI—why scientists must oversee its use. Nature 622, 693-696 (2023).
127 Schwartz, I. S., Link, K. E., Daneshjou, R. & Cortes-Penfield, N. Black box warning: Large
language models and the future of infectious diseases consultation. Clinical Infectious
Diseases. 78, 860-866 (2024).
128 Ahmadpoor, M. & Jones, B. F. The dual frontier: Patented inventions and prior scientific
advance. Science 357, 583-587 (2017).
129 Mukherjee, S., Romero, D. M., Jones, B. & Uzzi, B. The nearly universal link between the
age of past knowledge and tomorrow's breakthroughs in science and technology: The hotspot.
Science Advances 3, e1601315 (2017).
130 Yin, Y. A., Dong, Y. X., Wang, K. S., Wang, D. S. & Jones, B. F. Public use and public
funding of science. Nature Human Behaviour. 6, 1344-1350 (2022).

30
 Supplementary Information:
Quantifying the Benefit of Artificial Intelligence for Scientific Research
Jian Gao1,2,3, Dashun Wang1,2,3,4*
1
Center for Science of Science and Innovation, Northwestern University, Evanston, IL, 60208, USA
2
Kellogg School of Management, Northwestern University, Evanston, IL, 60208, USA
3
Northwestern Institute on Complex Systems, Northwestern University, Evanston, IL, 60208, USA
4
McCormick School of Engineering, Northwestern University, Evanston, IL, 60208, USA
* Correspondence to: [email protected]

Table of Contents
Supplementary Note 1. Data description ........................................................................................ 2
1.1 Microsoft Academic Graph (MAG) publication data ......................................................................... 2
1.2 United States Patent and Trademark Office (USPTO) patent data..................................................... 3
1.3 Open Syllabus Project (OSP) syllabus data ........................................................................................ 3
1.4 Survey of Doctorate Recipients (SDR) demographic data ................................................................. 3
Supplementary Note 2. Measuring direct AI impact ...................................................................... 5
2.1 Identification of AI-related publications and n-grams........................................................................ 5
2.2 Temporal trends in AI n-grams in disciplines .................................................................................... 6
2.3 Calculation of direct AI impact score ................................................................................................. 6
2.4 Citation impact of papers that mention AI terms................................................................................ 8
Supplementary Note 3. Measuring potential AI impact ................................................................. 9
3.1 Identification of AI-related patents ..................................................................................................... 9
3.2 Inference of AI capabilities and field tasks ........................................................................................ 9
3.3 Calculation of potential AI impact score .......................................................................................... 10
Supplementary Note 4. Unpacking AI’s growing impact ............................................................. 13
4.1 Discipline heterogeneity and field consistency ................................................................................ 13
4.2 Understanding the temporal trends in AI impact .............................................................................. 13
4.3 The relationship between potential and direct AI impact ................................................................. 14
Supplementary Note 5. Measuring AI education levels ............................................................... 16
5.1 Crosswalks from OSP data to MAG data ......................................................................................... 16
5.2 Estimating the levels of AI education ............................................................................................... 17
5.3 Relationships between AI education and AI impact......................................................................... 18
Supplementary Note 6. Measuring collaborations on AI.............................................................. 19
6.1 Identification of AI researchers and domain experts ........................................................................ 19
6.2 Temporal trends in cross-discipline AI collaboration ...................................................................... 19
6.3 Robustness checks on the identification of AI researchers .............................................................. 20
Supplementary Note 7. Measuring demographic disparities ........................................................ 21
7.1 Crosswalks from SDR data to MAG data......................................................................................... 21
7.2 Gender disparity in AI impact .......................................................................................................... 21
7.3 Race disparity in AI impact .............................................................................................................. 21
Supplementary Note 8. Robustness check of main results ........................................................... 23
8.1 Precision-recall analysis of identifying AI research ......................................................................... 23
8.2 Robustness test on the direct AI impact calculation ......................................................................... 24
8.3 Robustness test on the potential AI impact calculation .................................................................... 24
Supplementary Note 9. Career effects of engaging in AI research............................................... 26
9.1 Name-based demographic data inference ......................................................................................... 26
9.2 Effects on citation premium and leading the field ............................................................................ 26
Supplementary Figures ................................................................................................................. 28
Supplementary Tables ................................................................................................................... 41
Supplementary References............................................................................................................ 42
1
Supplementary Note 1. Data description

1.1 Microsoft Academic Graph (MAG) publication data

The publication data are sourced from the Microsoft Academic Graph (MAG) database [1-3],
which is among the largest open-source publication databases in the world. The MAG dataset
contains records of about 200 million documents of various types (e.g., papers and books),
interlinked through citation relationships. For each document, the data include information about
the author, affiliation, publication venue, and research topics. We use the December 2019 version
of the MAG dataset to avoid any overlap with the period of the COVID-19 pandemic, which may
have had a substantial impact on the publication and collaboration landscape, especially in research
that is not related to COVID-19 [4-7]. While the MAG database has been retired as of December
2021, its replacement and successor OpenAlex draws data from MAG’s existing records and other
sources, providing a free, up-to-date version [8, 9]. In addition, there is a publicly available, large-
scale data lake called SciSciNet, which pre-processes the last MAG data snapshot (from December
2021) and provides millions of external linkages to other data sources [10]. At the time of this
writing, a version of the MAG data (from September 2021) is also freely accessible at Zenodo [11].

To focus on scientific publications, we only consider MAG documents that are categorized as
“journal,” “conference,” “book,” or “book chapter.” All other documents are excluded from the
analysis, including those categorized as “patent” and “thesis.” We collect information on the author,
title, abstract, publication year, field of study, and digital object identifier (DOI) for each
publication, as well as the citation relationships among these publications. The MAG “field of
study” taxonomy contains six levels (from L0 to L5), covering more than 700K concepts, which
are not only seeded from Wikipedia pages (in which each page introduces a concept) and the
Unified Medical Language System (UMLS) vocabulary but also automatically discovered from
the MAG literature using some embedding models [12, 13]. Under the “field of study” taxonomy,
publications are categorized into 19 broad disciplines at the L0 level (e.g., “computer science”)
and 292 granular fields at the L1 level (e.g., “machine learning”). Each publication is often
assigned to one L0-level scientific discipline and several L1-level research fields with different
scores. For those belonging to multiple disciplines or fields, we consider the one with the highest
score at each level. Further, we filter documents published during the period from 1960 to 2019,
those with English titles and abstracts, and those that have a “field of study” at both the L0 level

2
and L1 level. When there are multiple versions of the same paper, we only keep the primary record
that is identified by MAG. These data cleaning and filtering processes lead to a subset of 74.6
million publications in the analysis.

1.2 United States Patent and Trademark Office (USPTO) patent data

The patent data are collected from PatentsView [14, 15], a data platform based on bulk data from
the United States Patent and Trademark Office (USPTO) [16]. We use the December 2019 version
of the PatentsView USPTO patent data. For each patent document, the data provide information
about the title, abstract, year granted, technology classes, inventor, and details of the invention.
Each patent is associated with a list of technology classification codes that are based on the
Cooperative Patent Classification (CPC) system. We use these CPC codes, supplemented with a
list of AI-related keywords, to identify AI patents (see Supplementary Note 3.1 for details). We
filter patents that were granted in the period from 1976 to 2019 and those with titles and CPC
codes. This data filtering process leads to a subset of 7.1 million patents in the analysis.

1.3 Open Syllabus Project (OSP) syllabus data

The syllabus data are sourced from the Open Syllabus Project (OSP) [17-19], which is the world’s
first large-scale online database of university course syllabi, covering about nine million English-
language syllabi from 140 countries. Each syllabus document contains the summary text, year,
field name, Classification of Instructional Programs (CIP) code, and a list of scientific references.
As our focus is on recent educational trends, we filter syllabus documents published in the period
from 2000 to 2018, leading to a subset of 4.2 million course syllabi. Both field names and CIP
codes in a syllabus document indicate the educational discipline of the course. Therefore, we
crosswalk these two types of educational disciplines to the research disciplines in MAG (see
Supplementary Note 5.1 for details). We rely on field names in the analysis and use CIP codes as
robustness checks. Each scientific reference contains information about the author, title, journal,
year, and DOI. Most references cited in syllabus documents are research papers and books, which
we match with MAG publication records (see Supplementary Note 5.1 for details).

1.4 Survey of Doctorate Recipients (SDR) demographic data

The demographic data on scientists and researchers are collected from the Survey of Doctorate
Recipients (SDR) [20], which provides demographic information for individuals with a U.S.

3
research doctoral degree in a science, engineering, or health field. We rely on the SDR data for
three primary reasons. First, the data provide reliable survey-based demographic statistics on the
sex, ethnicity, and race of researchers. By comparison, name-based gender and race inference
algorithms often fail to achieve a desirable level of accuracy, as has been shown, for instance, for
author names with East Asian origins [21-23]. Second, the data cover a substantial portion of the
international research workforce. Over 10% of U.S.-trained science, engineering, and health
doctorate holders (about 1.1 million individuals) were asked to participate in the 2017 SDR, and
they live and work throughout the world [24]. Third, the SDR provides statistical data by the field
of doctorate, which is the granularity required to be manually cross-walked to MAG scientific
disciplines (see Supplementary Note 7.1 for details). Our study, therefore, uses summary statistics
from the 2017 SDR data, which were de-identified, covering 815K U.S.-residing employed
doctoral scientists and engineers, by the field of doctorate, sex, and ethnicity/race.

4
Supplementary Note 2. Measuring direct AI impact

2.1 Identification of AI-related publications and n-grams

Identifying AI-related research from large-scale publication databases is a challenging task, as the
definitions of AI are often ambiguous and vary by context [25-33]. Following prior works [34-36],
we identify AI-related publications from the MAG datasets using the five L1-level field categories:
“machine learning,” “artificial intelligence,” “computer vision,” “natural language processing,”
and “pattern recognition.” These five benchmarking AI fields all belong to the L0-level “computer
science” discipline according to the MAG “field of study” formal hierarchy [12]. Publications
within any of these five AI fields are seen as AI-related papers in the analysis. Because MAG used
a topic modeling approach to label the field for each paper, the AI papers we identify using the
five field categories go beyond the explicit mention of these keywords [1-3]. We find that the share
of AI-related papers among all papers published in each discipline has increased for most
disciplines, especially in recent years (Supplementary Figure 1).

In addition, we identify a broader set of papers that mention AI-related terms. Specifically, using
Natural Language Processing (NLP) techniques [37-39], we extract n-grams (bigrams and trigrams)
from the titles and abstracts of AI-related publications that are initially identified by the five MAG
benchmarking AI fields, and we normalize them by lemmatizing and standardizing words. From
these n-grams, we further filter AI n-grams using a list of concepts that belong to the five AI fields
according to the MAG “field of study” formal hierarchy [12]. This process helps filter out n-grams
that are not meaningful for scientific research. Some of the remaining n-grams are still very generic,
so we further hand-curated a list of AI-related n-grams. We apply this n-gram approach to all
publications and identify those that mention at least one AI n-gram as publications that use AI.
This approach allows us to identify AI-related research not only in the five benchmarking AI fields
and the “computer science” discipline but also in other fields and disciplines.

There are several limitations and alternatives to this identification approach. Arguably, papers that
belong to the five benchmarking AI fields in the MAG dataset or mention AI terms in their titles
or abstracts do not necessarily use AI or benefit from AI advances; it is possible that some of these
papers simply discuss AI. It would ideally to scan the full text of the publications with emphasis
on the method sections. However, systematic data on the full text or the method sections are not

5
currently available. While our approach may, therefore, miss some AI-related papers because we
only use the title and abstract of each publication, the criteria we use serve as an imperfect proxy
for AI research. We perform a precision-recall analysis of the approach, finding that it has high
precision and recall in identifying AI-related research (see Supplementary Note 8.1 for details).
We also use two alternative approaches to identify AI research: (1) using a broader set of AI
keywords as MAG field categories and (2) using a set of four arXiv subject categories. We find
that our main results are largely robust (see Supplementary Note 8.1 for details).

2.2 Temporal trends in AI n-grams in disciplines

We apply the n-gram approach to AI papers, identify a list of AI n-grams, and then apply the
approach to all MAG publications in each discipline and field. We find that the landscape of AI
research is evolving dynamically. Specifically, some conventional AI n-grams have become less
common over time, some were dormant for a long period but rose to prominence in recent years,
and some novel AI n-grams emerged only recently (see Figures 1b,c in the main text). At the same
time, we find that AI n-grams are increasingly mentioned in papers in disciplines and fields outside
of AI over time, suggesting that AI is having a progressively broader impact across the sciences.
The increasing trend of AI n-grams in publications from different disciplines is especially evident
in recent years, primarily after 2015. This pattern is observed not only in the Science, Technology,
Engineering, and Math (STEM) or computational disciplines such as physics, medicine, biology,
and chemistry, but also in some social science disciplines such as business, psychology, and
sociology (Supplementary Figure 2).

2.3 Calculation of the direct AI impact score

The dynamic landscape of AI research prompts us to study the use of AI in different fields and the
direct impact of AI (i.e., the relative frequency of AI methods uptake in science) over time. To this
end, we estimate the direct impact of AI by implementing the “AI n-gram framework.” Specifically,
we first extract AI n-grams from the titles and abstracts of AI publications and calculate the
frequency of AI n-grams per paper to approximate cumulative AI advances in research. Formally,
the frequency vector of AI n-grams at year 𝑡 is given by
#
𝐺!"
𝐺#!"
#
= # ,
𝑁!"

6
 #
where 𝐺!" is the vector that summarizes the counts of AI n-grams extracted from AI papers
#
published before year 𝑡, and 𝑁!" is the number of AI papers. Here we use 1960 as the starting year
to cumulate AI n-grams. For papers in each field, we apply the same approach to extract n-grams
(AI n-grams and non-AI n-grams) and calculate their frequency to approximate field development.
Formally, the n-gram frequency vector for the biology discipline at year 𝑡 is given by

𝐺$#
𝐺#$# = ,
𝑁$#

where 𝐺$# is the vector that summarizes the count of n-grams extracted from biology papers
published at year 𝑡, and 𝑁$# is the number of these biology papers. The coordinate of the same AI
n-gram in the biology frequency vector and the AI frequency vector is the same. In other words,
each coordinate of the normalized biology vector 𝐺#$# represents one n-gram, where AI n-grams
have the same coordinates as those in the normalized AI vector 𝐺#!"
#
. Finally, we calculate the direct
AI impact score for biology at year 𝑡 using a dot product measure:

𝑆%# = ) 𝐺#$# ∙ 𝐺#!"

#
,

where the symbol “∑∙” represents the dot product among the biology frequency vector and the AI
frequency vector of the same AI n-grams. In this calculation, only common n-grams in the AI n-
gram vector and the biology n-gram vector are considered, and the same n-gram has the same
coordinate in these two vectors. A larger direct AI impact score indicates that more AI n-grams
are used in the research field and, thus, that it receives a larger direct benefit from AI.

We apply the “AI n-gram framework” to the titles and abstracts of publications between 1960 and
2019. There, we calculate the direct AI impact score for each of the 19 MAG L0-level scientific
disciplines (e.g., “biology”) and year. Applying the same approach, we calculate the direct AI
impact score for each of the 292 MAG L1-level research fields (e.g., “biological system”) and year.
At the scientific discipline level, we find that the direct AI impact score increased over time in the
period 1960-2019, and there has been a notably sharp increase since 2015 across many disciplines,
suggesting the widespread impact of AI on scientific research (Supplementary Figure 3). In
addition to the dot product, we tried two alternative measures to calculate the direct AI impact
score, finding that our main results are largely robust (see Supplementary Note 8.2 for details).

7
2.4 Citation premium of papers that mention AI terms

Based on the curated data on AI n-grams, we explore the extent to which using AI is associated
with a citation premium. Specifically, we calculate the average rate for AI papers (i.e., those that
mention at least one AI n-gram) to be “hit papers,” which we define as being in the top 5% by total
citations in the same field and year. We find that AI papers are more likely to become hit papers
than non-AI papers in each discipline (Supplementary Figure 4a). For a majority of fields, the hit
rate of AI papers is larger than the baseline 5%, and it has a significantly negative correlation (two-
sided Pearson’s correlation test; Pearson’s r = −0.378; P-value < 0.001) with the direct AI impact
score (Supplementary Figure 4b). These results suggest that AI papers in research fields with less
direct AI impact tend to be associated with a larger citation premium on average.

Moreover, we study the broader citation impact of papers that use AI by calculating the share of
citations AI papers receive that are outside-field citations. At the discipline level, we find that AI
papers have a larger share of outside-field citations than non-AI papers (Supplementary Figure 4c).
At the same time, there is substantial heterogeneity across disciplines, as AI papers receive
disproportionately more outside-field citations in biology, geology, and arts than non-AI papers.
At the field level, we find that the ratio of AI to non-AI papers for the share of outside-field
citations is negatively correlated (two-sided Pearson’s correlation test; Pearson’s r = −0.215; P-
value < 0.001) with the direct AI impact score (Supplementary Figure 4d). Taken together, these
results suggest that papers that use AI tend to have a larger and broader citation impact on science
than traditional field papers. Moreover, the citation premium of papers that use AI appears to be
stronger for disciplines with an overall lower propensity to use AI, suggesting that disciplines that
seem distant from AI may reap substantial benefits from using AI to advance their research work.
It should be noted that the analysis does not claim that citations result from the use of AI in the
research. Instead, it highlights the citation premium of papers that mention AI terms in their titles
or abstracts, which is a correlation and does not imply causality.

8
Supplementary Note 3. Measuring potential AI impact

3.1 Identification of AI-related patents

The identification of AI-related patents, like the identification of AI-related papers, is a challenging
task [29, 40-43]. Among the USPTO patents collected from PatentsView, we identify AI-related
patents using a mixed method of keywords and patent CPC codes, as suggested by the World
Intellectual Property Organization (WIPO) [44]. The AI-related keywords (key phrases) and CPC
codes are sourced from the PATENTSCOPE Artificial Intelligence Index [45]. The index provides
AI techniques, AI functional applications, and AI application fields, which have been used to carry
out searches in PATENTSCOPE and other patent databases to identify patent applications related
to AI. Following this line of work, we identify AI-related patents in the USPTO data by matching
CPC codes and searching for key AI phrases in the titles of patents. Altogether, we identify 132K
AI-related patents that were granted between 1976 and 2019. There are, however, other ways to
identify AI-related patents, including the Artificial Intelligence Patent Dataset (AIPD) that results
from a machine learning approach [41] and the AI technological innovation dataset that results
from NLP and machine learning tools for text classification [46], which became available recently.

3.2 Inference of AI capabilities and field tasks

While the explicit mention of AI-related n-grams signals direct AI benefits, it is possible that
research fields could be using AI to perform more of the basic tasks in their scientific research,
suggesting the potential impact of AI on science. To this end, building on the future of work
literature [47-54], we develop the “AI capability–field task framework” to estimate the potential
AI impact on scientific research. This framework relies on the inference of AI capabilities from
AI publications and patents and field tasks from publications.

We infer the capabilities of AI (i.e., the tasks that AI can perform) and the basic tasks of research
fields (i.e., the tasks undertaken by a research field) by applying NLP techniques to analyze
publications and patents. Specifically, we first employ a dependency parsing algorithm (spaCy
dependency parser in the Python library [38, 55]) to extract verb-noun pairs from the titles of AI
publications that are identified by the five MAG AI fields. We then lemmatize verbs and nouns in
these verb-noun pairs, and we group words with similar meanings into one word that represents
the meaning. For example, verbs like “detecting,” “detected,” and “detects” are represented by

9
“detect.” This cleaning and normalization process through lemmatization and standardization
reduces noise in identifying relevant verb-noun pairs from publications. As an illustrative example,
from the AI paper titled, “Using Bayesian networks to analyze medical data” [56], the algorithm
extracts “use network” and “analyze data” as the verb-noun pairs (Supplementary Figure 5).

We apply the verb-noun extraction method to AI publications identified by the five MAG AI fields
and AI patents identified by keywords and CPC codes. The verb-noun pairs extracted from AI
publications and patents together represent the capabilities of AI (see the word cloud of AI verb-
noun pairs cumulative up to 2019 in Supplementary Figure 6a). We apply the same approach to
analyze publications in each field. The verb-noun pairs extracted from a field’s publications
represent the basic tasks of the field (see the word clouds of verb-noun pairs in the biology and
chemistry disciplines in 2019 in Supplementary Figure 6b and Supplementary Figure 6c,
respectively). As some of the verb-noun pairs extracted from publications and patents are very
generic, we inspected and manually curated a list of verbs, nouns, and verb-noun pairs for frequent
words. It is important to note that here we only extract verb-noun pairs from the titles of
publications and patents because they have a much higher signal-to-noise ratio than the other
publication text fields, such as the abstract, and the other patent text fields, such as the abstract,
description, and claims [47].

3.3 Calculation of potential AI impact score

The “AI capability–field task framework” assumes that research fields may potentially benefit
from AI if their basic tasks are aligned with AI capabilities. To this end, we estimate the potential
AI impact score based on the alignment between AI capabilities and field tasks. Specifically, we
first calculate the relative frequency of verb-noun pairs extracted from the titles of AI papers and
AI patents to approximate cumulative AI capabilities for each. Formally, the AI capability
frequency vector for papers at year 𝑡 is given by
#
𝐶!"
𝑃𝑎𝑝𝑒𝑟(𝐶4!"
#
)= # ,
∑ 𝐶!"
#
where 𝐶!" is the vector that summarizes the counts of verb-noun pairs extracted from AI papers
before year 𝑡. We apply the term frequency-inverse document frequency (tf-idf) to discount the
weight of commonly appearing pairs [57]. Here, each vector coordinate is one verb-noun pair, and
the value is its relative frequency. Likewise, we repeat this process for AI patents and calculate the

10
AI capability frequency vector 𝑃𝑎𝑡𝑒𝑛𝑡(𝐶4!"
#
). We further normalize frequency vectors of AI papers
and AI patents to approximate cumulative AI capabilities at year 𝑡. Specifically, we take an
average of common verb-noun pairs in these two frequency vectors, where the same verb-noun
pair shares the same coordinate. Formally, the AI capability frequency vector at year 𝑡 is given by:

7𝑃𝑎𝑝𝑒𝑟8𝐶4!"
#
9 + 𝑃𝑎𝑡𝑒𝑛𝑡8𝐶4!"
#
9;
𝐶4!"
#
= ,
2
where the symbol “+” represents summing two frequencies of the same verb-noun pair. For
simplicity and following prior works [58], here we treat verb-noun pairs extracted from AI papers
and AI patents as equally important in inferring AI capabilities, and we normalize them separately.
In the calculation, we use 1960 as the starting year to cumulate AI capabilities based on AI
publications and AI patents. Because there is no patent data in the period 1960-1975, we only use
𝑃𝑎𝑡𝑒𝑛𝑡(𝐶4!"
#
) to represent the AI capability frequency vector 𝐶4!"
#
for this period.

We then calculate the relative frequency of verb-noun pairs extracted from the titles of papers in a
research field to approximate the current basic tasks of the field. Formally, taking the biology
discipline (or one of its subfields) as an example, the field task frequency vector at year 𝑡 is

𝑇$#
𝑇#$# = ,
∑ 𝑇$#

where 𝑇$# is the vector that summarizes the counts of verb-noun pairs extracted from biology
papers published at year 𝑡. We also apply the tf-idf approach to discount the weight of commonly
appearing pairs in field tasks. Finally, we estimate the potential AI impact score of biology at year
𝑡 by calculating the overlap between its field tasks and cumulative AI capabilities at 𝑡:

∑ 𝑇#$# ∙ 𝐶4!"
#
𝑆&# = ,
∑ 𝐶4!"
#
∙ 𝐶4!"
#

where the symbol “∑∙” represents the dot product among the AI verb-noun vector and the biology
verb-noun vector. Specifically, only common verb-noun pairs in the AI verb-noun vector and the
biology verb-noun vector are considered in the calculation, and the same verb-noun pair has the
same coordinate in these two vectors. The denominator is applied to normalize the AI score for
comparison across time. The normalization assumes that AI can perform all research tasks that are
inferred from AI publications in the five AI fields.

11
A larger potential AI impact score for a field indicates a better alignment between field tasks and
AI capabilities and thus a higher potential for the field to benefit from AI advances. Applying the
“AI capability–field task framework” to all publications between 1960 and 2019, we calculate the
potential AI impact score for each of the 19 disciplines and years. Applying the same approach,
we calculate the potential AI impact score for each of the 292 fields and years. In addition to the
dot product, we tried two alternative measures to calculate the potential AI impact score, finding
that our main results are largely robust (see Supplementary Note 8.3 for details).

12
Supplementary Note 4. Unpacking AI’s growing impact

4.1 Discipline heterogeneity and field consistency

We estimate the direct AI impact using the “AI n-gram framework” and the potential AI impact
using the “AI capability–field task framework” based on publications and patents. At the discipline
level, we find that the differences in the direct AI impact scores span as large as two magnitudes
(Supplementary Figure 7a), suggesting that there is substantial heterogeneity across disciplines.
By comparison, the magnitude of discipline-level differences for the potential AI impact scores is
smaller, despite the rapid increase in the potential AI impact scores of some disciplines, such as
material science and chemistry (Supplementary Figure 7b).

We further explore the within-discipline heterogeneity by plotting the positions of a discipline’s

child fields that are ranked in percentile by their direct and potential AI scores (Supplementary
Figure 8). In the percentiles of 292 fields, the field with the lowest AI score is 0th, and the field
with the highest AI score is 100th. We find that the percentiles of the direct AI impact score and
potential AI impact score for all fields are highly correlated with each other (see Fig. 2e in the
main text), and the two percentiles for the same subfield within each discipline are largely
consistent (Supplementary Figure 8). Moreover, we find that almost every discipline contains
some subfields that are greatly impacted by AI—even those that have relatively small AI impact
scores. For example, “medicine” as a discipline has medium AI impact scores, but it contains some
subfields that have large AI impact scores, such as “nuclear medicine,” “optometry,” and “medical
physics” (Supplementary Table 1). These results suggest that the benefits of AI for scientific
research are pervasive across disciplines and fields.

4.2 Understanding the temporal trends in AI impact

We seek to understand the extent to which the evolving trends in AI impact are associated with
changes in AI capabilities or fields. Here, we calculate a new AI impact score at t2 by holding AI
n-grams and AI capabilities constant at t1. Specifically, we extract AI n-grams from AI-related
papers published before t1 and normalize them to construct the AI n-gram frequency vector as a
measure of cumulative AI capabilities. We use this vector without changing any of its AI n-grams
or their frequency to calculate direct AI impact scores for the subsequent years between t1 and t2.
In other words, only AI terms that existed before t1 are considered; no new terms are added to the

13
AI capability vector after t1, and each AI term’s frequency remains unchanged after t1. In this way,
we can separate the increase in the AI impact score into two parts: the increase due to new AI
capabilities and the increase due to old AI capabilities (Supplementary Figure 9a). Specifically,
the new AI part (X) is the increase in the score compared with the new score at t2, while the old
AI part (Y) is the increase in the new score at t2 compared with the score at t1 (Z).

We start by considering the period from 2015 (t1) to 2019 (t2). For the direct AI impact, we find
that while the increase in 2019 due to old AI is not very significantly correlated with the score in
2015 (Supplementary Figure 9b), the increase in 2019 due to new AI can be largely predicted by
the score in 2015 (Supplementary Figure 9c). This finding suggests that disciplines with a larger
direct AI impact tend to be affected more by new AI capabilities than by old AI capabilities. In
relative terms, the direct and potential AI impact scores are about 20% and 5%, respectively, above
the corresponding new scores in 2019 (Supplementary Figure 7, dashed lines). The sharp increases
in the direct AI impact score are remarkable (over 20%) not only in computer science, but also in
many other disciplines, including physics, medicine, biology, and sociology. Together, the results
suggest that disciplines are disproportionately impacted by new AI advances.

4.3 The relationship between potential and direct AI impact

We study the relationship between potential and direct AI impact. Specifically, we focus on one
snapshot in time and explore the discrepancy between the potential and the direct AI impact of
each field in 2019. To do this, we first fit a regression line and calculate the distance from each
field to the average (i.e., the predicted direct AI impact by its potential). We then rank all fields
according to this distance and construct three field groups (Supplementary Figure 10a): (a)
“higher-direct-impact” fields are those that have higher direct AI impact than potential AI impact;
(b) “higher-potential-impact” fields are those that see higher potential AI impact than direct AI
impact; and (c) “others” fields are those that are not included in the first two groups.

We perform analyses to understand the features associated with the “higher-direct-impact” and
“higher-potential-impact” fields. First, we study research collaboration on AI (see Supplementary
Note 6 for details). Specifically, we calculate the share of each field’s AI papers that are published
collaboratively by domain experts and primary AI authors (i.e., collaborative AI papers). We find
that the share of collaborative AI papers is higher in the “higher-direct-impact” fields than in the

14
“higher-potential-impact” fields (two-sided Student’s t-test; P-value < 0.001; Supplementary
Figure 10b). Second, we consider university education in AI (see Supplementary Note 5 for
details). Specifically, we estimate each field’s AI education level based on its university course
syllabus data. We find that the “higher-direct-impact” fields tend to be associated with a higher AI
education level on average than the “higher-potential-impact” fields (two-sided Student’s t-test;
P-value < 0.001; Supplementary Figure 10c). Note that these analyses are correlational by nature
and hence don’t imply causal relationships. These results nevertheless offer support for the
hypothesis that fewer collaborations with AI researchers and less education in AI may be barriers
to the use of AI in unrealized fields that could have seen larger direct AI benefits.

Next, instead of the snapshot analysis, we examine the “momentum” of fields by studying the
research fields that shifted from being in the “higher-potential-impact” group in 2000 to being in
the “higher-direct-impact” group in 2019 (i.e., “upshift” group), and vice versa (i.e., “downshift”
group). Here again, we find that fields that experienced upward shifts tended to feature more AI
collaborations and higher education levels in AI (two-sided Student’s t-test; P-value < 0.001 for
both AI collaboration and education; Supplementary Figure 11), again consistent with the findings
above. While far from being conclusive due to data limitations, these initial analyses are consistent
with the hypotheses that the barriers to AI adoption may revolve around collaborations with AI
experts and university education in AI. The relationship between the potential and the direct AI
impact of a field is an intriguing research direction that merits substantial further investigation.

15
Supplementary Note 5. Measuring AI education levels

5.1 Crosswalks from OSP data to MAG data

To estimate the levels of AI education in disciplines and correlate them with the AI impact, we
must first align educational disciplines in course syllabi with scientific disciplines in publications
and then identify AI publications referenced by course syllabus documents. Therefore, we begin
by manually cross-walking the Open Syllabus Project (OSP) data to the Microsoft Academic
Graph (MAG) data at the academic discipline level (from education disciplines in OSP to research
disciplines in MAG) and at the scientific publication level (from publications referenced by OSP
syllabi to publications in the MAG dataset), respectively.

First, we crosswalk from OSP educational disciplines to MAG research disciplines. Specifically,
OSP uses two classification systems to categorize syllabus documents [59]. The first system is the
Classification of Instructional Programs (CIP) system, which is a taxonomy of academic
disciplines at institutions of higher education in the United States and Canada [60]. OSP uses the
2010 CIP taxonomy to determine the field that is best associated with each syllabus document.
The CIP codes come in lengths of two-, four- and six-digits, where two-digit codes represent a
discipline, and six-digit codes represent a more granular field of that discipline. For example, the
two-digit CIP code “11” is the discipline for all “Computer and information sciences and support
services” courses, and the six-digit CIP code “11.0104” is the field for all “Informatics” courses.
Each syllabus document is associated with one or more CIP codes at different levels that often
belong to the same two-digit CIP code category. For those with multiple two-digit codes, we only
keep the first for simplicity. We manually crosswalk the two-digit CIP codes covered in the OSP
data (31 codes) to the L0-level disciplines covered in the MAG data (14 disciplines) by matching
the names of these disciplines and selecting the closest match for each CIP code.

Another classification system is the OSP field name classifier, which is trained on a subset of the
CIP taxonomy for describing syllabi [59]. The OSP classification system generally combines CIP
codes within the same two-digit branch of the taxonomy into a field name. For example, CIP code
“45.09” for “International Relations and National Security Studies” and CIP code “45.10” for
“Political Science and Government” are both under CIP code “45” for “Social Sciences” and are
combined into the OSP field named “Political Science.” We manually crosswalk OSP field names

16
(62 names) to MAG disciplines/fields (17 L0-level disciplines and 45 L0/L1-level fields) by
matching field names and selecting the closest match for each OSP field name. Further, we
combine these two classification systems to assign only one MAG L0-level discipline (or field at
the L1 level) to each syllabus document that best describes its academic discipline. We rely on the
OSP field name classification system and supplement it with the CIP classification system when
there are missing values in the CIP codes or inconsistencies in the matched L0-level disciplines.

Second, we match publications cited by OSP syllabus documents to scientific publications in MAG.
Specifically, for publications with DOI information, we use DOI to match them to MAG
publications directly. To reduce false positives, we confirm that the last names of the first author
for each matched publication record are the same. For publications without DOI information,
which are typically books and conference papers, we match them to MAG publications based on
the publication’s title, the last name of the first author, and the publication year. We exactly match
the title and last name but allow a year difference in the publication year. In total, we match about
924K publications cited by syllabus documents published between 2000 and 2018.

5.2 Estimating the levels of AI education

Based on the linkages from OSP to MAG data at the discipline level and the publication level, we
estimate the levels of AI education in each educational discipline. We assume that an educational
discipline’s AI education level is higher if syllabus documents in the discipline cite more
publications related to AI because these citations are a signal that the course teaches more AI
knowledge. For each educational discipline, we calculate the share of citations from syllabus
documents in the discipline to AI publications identified from the MAG data using the five AI
field categories, and we use this measure as the level of AI education for the discipline.

Before we calculate the levels of AI education based on the OSP data, we filter the syllabus
documents to ensure that we are capturing meaningful measures of AI education. First, we use
syllabus documents that cite at least five scientific references to proxy courses taught at the
graduate level and above. The reasoning here is that syllabi with fewer references may more likely
correspond to undergraduate-level courses, and those with more references may correspond to
graduate-level or more research-oriented courses. Students who take these higher-level courses
should have a higher propensity to perform research and thus better represent the student

17
population that benefits from AI education. Second, we only use syllabus documents that were
published in the most recent five-year period, 2014-2018, according to our OSP dataset. As we
find that the impact of AI is most widespread since 2015, measuring AI education levels based on
recent syllabi is more relevant. Alternatively, we could use other criteria to filter the syllabus
documents or estimate AI education levels using other measures, including the share of syllabus
documents that cite at least one AI publication. These results are robust across several analytical
choices (see Supplementary Note 5.3 for details).

5.3 Relationships between AI education and AI impact

Analyzing the relationship between the level of AI education and the impact of AI on scientific
research, we find that the discipline-level correlation decreases, and so does its significance after
excluding the top three computation-intensive disciplines from the analysis (i.e., computer science,
mathematics, and engineering). This observation holds robustness for both the direct AI impact
and the potential AI impact that we estimate based on publication and patent data. Overall, our
analyses convey three sets of findings across several measures. First, we find that the discipline-
level correlation between AI education levels and AI impact scores is insignificant after excluding
the top three disciplines (Supplementary Figure 12a). Second, we compare three samples: syllabi
with one to five references, syllabi with at least five references, and syllabi with at least ten
references. After repeating our analyses for these three samples, we find that consistent with our
intuition, the correlation between AI education and AI impact is weaker for the group with one to
five references than the group with at least five references (Supplementary Figure 12b). We also
find that our results remain robust when using data with at least ten references to identify graduate-
level courses (Supplementary Figure 12c). Third, we find that the results are robust when using
the share of syllabi that cite at least one AI publication to estimate AI education levels
(Supplementary Figure 12d), and the correlation is weaker in the earlier periods of 2008-2013 and
2000-2007 compared to 2014-2018 (Supplementary Figures 12e,f).

18
Supplementary Note 6. Measuring collaborations on AI

6.1 Identification of AI researchers and domain experts

Based on AI papers in disciplines other than computer science, we estimate the level of cross-
discipline collaboration on AI-related research between domain experts and AI researchers. First,
we assign a primary discipline to each researcher. Each paper belongs to one primary discipline in
our data, so we count each author’s publications in each discipline and assign the discipline in
which they published most frequently in the period 1960-2019 to be the author’s primary discipline.
If an author has multiple most frequent disciplines, we randomly select one discipline. Second, we
treat authors whose primary discipline is computer science (CS) as computer scientists and those
with primary disciplines other than CS as domain experts. Although authors who frequently
publish CS papers may specialize in general computer science and not necessarily in AI research,
we use computer scientists as an imperfect proxy for identifying AI researchers and further check
its robustness using an alternative approach (see Supplementary Note 6.3 for details). Third, based
on the composition of authors’ primary disciplines, we categorize each AI paper published in a
discipline other than CS into one of the four co-authorship types: (1) “domain & CS,” which
involves both domain experts and computer scientists; (2) “domain sole,” which involves only
domain experts; (3) “CS sole,” which involves only computer scientists; and (4) “others,” which
involve neither domain experts nor computer scientists (e.g., an AI paper in the geology discipline
published by two physicists who are not domain experts according to our identification method).
Then, we calculate the share of collaborative AI papers (i.e., the “domain & CS” co-authorship
type) for each discipline and correlate it with the AI impact measures.

6.2 Temporal trends in cross-discipline AI collaboration

We analyze cross-discipline collaboration patterns for AI-related papers co-authored by domain

experts and/or computer scientists. Here we consider AI papers published by at least two authors
in the period 1980-2019 and major disciplines that have the largest AI impact scores and the
highest AI education levels. We find that while both the number of collaborative AI papers (i.e.,
“domain & CS”) and the number of domain-only AI papers (i.e., “domain sole”) increase over
time (Supplementary Figures 13a,b), the growth rate of collaborative AI papers is larger than that
of domain-only AI papers (Supplementary Figure 13c). Hence, in terms of the relative share, we

19
see that the share of domain-only AI papers decreases over time across various disciplines
(Supplementary Figure 13d). Meanwhile, there is an increasing share of collaborative AI papers
in major computation-intensive disciplines, especially in recent years (Supplementary Figure 14).
In particular, the share of collaborative AI papers increases rapidly in engineering and mathematics,
suggesting a growing reliance on AI expertise by domain experts or their increasing contribution
to AI research. Together, these results indicate that teamwork and cross-domain collaboration may
be particularly important as AI’s impact on research becomes more pervasive in the future.

6.3 Robustness checks on the identification of AI researchers

To test the robustness of the AI collaboration results, we considered alternative approaches to

identifying AI researchers, as well. An author’s career titles, such as the discipline in which they
obtained a Ph.D. or their academic department, are most informative for identifying primary
disciplines. Unfortunately, the MAG dataset does not include such information for each author,
and we are unable to identify AI researchers by applying AI terms to authors’ titles. However, we
develop an alternative approach to distinguishing primary AI authors based on AI publications in
the MAG data. Specifically, we identify primary AI authors as those who published at least three
AI papers according to the MAG five AI field categories and for whom more than half of their
publications are AI papers—in other words, those who frequently publish AI papers [36]. To
identify a primary discipline for each author, we use our previous approach, which is based on the
discipline in which they publish most frequently (see Supplementary Note 6.2 for detailed
explanations). This new method allows us to identify AI researchers based on their publications
that are directly relevant to AI, which is relatively more precise than the proxy of computer
scientists in general. We show that our main findings are robust with this new method of AI
researcher identification (Supplementary Figure 15).

20
Supplementary Note 7. Measuring demographic disparities

7.1 Crosswalks from SDR data to MAG data

We infer the gender and race/ethnicity compositions of each discipline based on the 2017 Survey
of Doctorate Recipients (SDR) data, which provides summary statistics on demographic
information (see Supplementary Note 1.4). The SDR field of doctorate taxonomy contains field
categories at different levels, including three categories at the top level (“Science,” “Engineering,”
and “Health”), as well as categories at a granular level (e.g., “Economics” and “Physics”). This
taxonomy is aligned with the National Center for Science and Engineering Statistics (NCSES)
Taxonomy of Disciplines (ToD). We manually crosswalk the 31 SDR fields to the 15 MAG
disciplines by comparing the names of these fields and disciplines. We match each SDR field to
the single MAG discipline that best represents its scientific discipline. The SDR data also report
statistics at a more granular field level. For simplicity, we only crosswalk at the discipline level as
we focus on overall trends. Further, we aggregate the SDR data to each matched MAG discipline
to calculate the discipline’s gender and race/ethnicity composition.

7.2 Gender disparity in AI impact

We quantify gender disparity using two measures. First, we calculate the share of women
researchers in each discipline. We find a negative correlation between the share of women
researchers and the AI impact score for both the direct and potential impact (Figs. 4a,b in the main
text). Second, we aggregate AI scores for each gender group by weighting the discipline-level AI
scores using the gender share of disciplines. A larger average AI score for a gender group indicates
that researchers in the group have a greater AI impact and thus benefit more from AI. We find that
women researchers, on average, benefit less from AI (Figs. 4c,d in the main text) than men
researchers. These observations suggest a gender gap across disciplines regarding AI benefit.

7.3 Race disparity in AI impact

The SDR data also report the number of researchers in each race group, including “white,” “Asian,”
and underrepresented minorities (URM). The URM group comprises “Black,” “American Indian
or Alaska Native,” “Hispanic or Latino,” and “Other race.” We quantify racial disparity using two
measures. First, we calculate the share of URM researchers in each discipline. We find a negative

21
correlation between the share of URM researchers and the AI impact score (Figs. 4e,f in the main
text). Second, we aggregate AI scores for each racial group by weighting the discipline-level AI
scores using the race share of disciplines. We find that URM researchers receive the least AI
benefits, especially Black researchers (Figs. 4g,h in the main text). The literature suggests that
workers of color are particularly vulnerable to job loss due to automation and AI [61, 62], and our
results suggest they are also less likely to benefit from AI in scientific research.

22
Supplementary Note 8. Robustness check of main results

8.1 Precision-recall analysis of identifying AI research

We analyzed the precision and recall of our approach in identifying AI research (see
Supplementary Note 2.1 for details). First, we randomly sampled 100 AI papers, defined as those
that mention at least one AI-related n-gram, regardless of the weight or importance of the particular
n-gram(s) in representing AI research. We then manually determine if these papers actually use AI
methods. We find that 14 of the papers fall into one of the MAG five AI fields (“machine learning,”
“artificial intelligence,” “computer vision,” “natural language processing,” and “pattern
recognition”), 69 papers are AI application papers, and 17 papers appear to be false positives that
mention less representative and less common AI n-grams. Therefore, we estimate that the approach
has a precision of around 0.83. Second, we randomly sample 100 AI papers, defined as those that
are identified by one of the four arXiv “Subject Categories”: cs.AI (Computer science: Artificial
Intelligence), cs.LG (Computer science: Machine Learning), cs.NE (Computer science: Neural and
Evolutionary Computing), and stat.ML (Statistics: Machine Learning) [30-33], and we determine
if they are covered by our n-gram approach. We find that 55 papers fall into one of the five MAG
AI fields, 33 papers mention at least one AI n-gram, and 12 papers appear to be false negatives.
We estimate, therefore, that the approach has a recall of around 0.87. Thus, the method has a high
level of precision and recall, and it balances these two metrics well.

We test the robustness of our findings using two alternative approaches to defining AI publications.
In the first approach, we identify the top 20 most frequent AI n-grams (Supplementary Table 2)
and extend the five AI fields to these 20 n-grams across all “field of study” categories, from the
L1 level to the L5 level in MAG (see Supplementary Note 1.1 for details). We treat publications
in any of these 20 fields as AI publications. In the second approach, we identify AI publications
using “Subject Categories” in the arXiv dataset (2023-11 version) [63] which provides the
metadata of each preprint. There, we matched arXiv preprints to MAG publications using the arXiv
Identifier. Following previous works [30-33], we treat preprints in four arXiv “Subject Categories”
as AI publications: cs.AI, cs.LG, cs.NE, and stat.ML. We apply our frameworks to AI publications
defined by these two alternative approaches and find that the main findings are largely robust
(Supplementary Figure 16). The literature also suggests other ways to identify AI publications,
including, for example, topic modeling and embedding-based methods [64, 65].

23
8.2 Robustness test on the direct AI impact calculation

The calculation of the direct AI impact score assumes that AI has a large impact on field X if
important AI techniques are frequently mentioned by papers in the field. It is essential to recognize
that the landscape of AI research evolves rapidly, with some AI methods becoming more popular
and some less popular over time, and this is why we distinguish AI n-grams by their level of
popularity. We use the dot product because it is a simple measure that balances the importance of
AI n-grams and their frequency of use in a field. In the AI vector, the value of each n-gram
coordinate signals its importance in representing AI research. In the field vector, the value of each
n-gram coordinate indicates its frequency of use in the field. The dot product helps to capture both
the importance of AI n-grams and their frequency of use in each field.

We check the robustness of our results using two alternative measures. In the first measure, we
treat each n-gram with the same importance in the AI vector and calculate the sum of the field
vector for its AI n-gram coordinates. Formally, 𝑆%# = ∑ 𝐺#$# ∙ 𝐼!" , where 𝐼!" represents an all-one
vector for AI n-grams. This sum measure captures the total frequency of AI n-grams in a field. In
the second measure, we calculate the weighted average for the field vector using the AI vector as
the weight for AI n-grams. Formally, 𝑆%# = ∑ 𝐺#$# ∙ 𝐺#!"
# ⁄∑ # #
𝐺!" . This weighted average measure
captures the average frequency of AI n-grams in a field. We find that the results of these two
alternative measures are highly correlated with the results of the dot product measure
(Supplementary Figure 17), suggesting the robustness of the main findings.

8.3 Robustness test on the potential AI impact calculation

We validated the “AI capability–field task framework” using several approaches. First, we use the
future of work literature, which has estimated the impact of AI on labor [26, 47-53], as the
foundation for our measurements. Specifically, this literature has used the texts of patents to infer
AI capabilities, used the descriptions of occupations to infer job tasks, and measured the exposure
of jobs to AI using the alignment between AI capabilities and job tasks [47]. We build on this
literature to estimate the potential AI impact on science. Second, we devised two alternative
methods to calculate the potential AI impact score. In the first method, we calculate the fraction of
AI verb-noun pairs in each research field, treating each AI verb-noun pair with the same weight.
Formally, the fraction measure is defined as 𝑆&# = ∑ 𝐶4$# ∙ 𝐼!" , where 𝐶4$# represents the field verb-

24
noun vector, and 𝐼!" represents an all-one vector for the AI verb-noun vector. In the second method,
we calculate the potential AI impact score using each field as the baseline, assuming that field
tasks are well aligned with field capabilities. Formally, the field baseline measure is defined as
𝑆&# = ∑ 𝐶4$# ∙ 𝐶4!"
# ⁄∑ 4 #
𝐶$ ∙ 𝐶4$# . We find that the potential AI scores are highly correlated (Pearson’s
correlation 0.95 and 0.90, respectively), suggesting the robustness of these measurements.

25
Supplementary Note 9. Career effects of engaging in AI research

9.1 Name-based demographic data inference

We study the possibility that engaging in AI research could lead to career-changing experiences
for women and URM researchers. This requires us to perform career-level analyses. To do this,
we employ naming-based inference algorithms (nomquamgender [66, 67] and demographicx [68,
69]) to infer the gender and race of each author in MAG, although the limitations of these
algorithms are noted (see Supplementary Note 1.4 for details). We only consider authors whose
gender can be confidently inferred to be “woman” or “man” for the gender analysis and whose
race can be confidently inferred to be “white,” “Asian,” “Black,” or “Hispanic” for the race
analysis. Based on this group of authors, we then construct a sample of authors who published at
least ten papers from 2000 to 2019 and have at least five years of career history (i.e., the time
difference between their first and last papers). We identify AI researchers as those who published
at least three AI papers as AI researchers and pinpoint the start of their AI engagement by tracing
their first AI paper.

9.2 Effects on citation premium and leading the field

We calculate the average hit rate of papers by an author before and after engaging in AI research,
measuring the fraction of papers that land within the top 5% of papers ranked by total citations in
the same field and year. We find that, on aggregate, the average hit rate of papers by a researcher
exhibits an increase right after engaging in AI research (Supplementary Figure 18a), highlighting
the citation premium of AI engagement. However, when we further unpack these results by gender
and race, we find that the citation premium is more concentrated in the overrepresented groups,
and women and URM researchers appear to benefit less from AI engagement than their
counterparts (Supplementary Figures 18b,c). This opens up a new question for further research.

We also examined whether women or URM authors lead the way in their fields in collaborating
with AI researchers. Here we assign a primary discipline to each author based on the discipline in
which the individual publishes most frequently and treat those with a primary discipline of
computer science as AI researchers (see Supplementary Note 6.1 for detailed explanations). We
identify the top 5% of researchers in each field in terms of 1) the number of papers and 2) their
number of hit papers as those who lead the way in their respective fields. We find that while

26
researchers who collaborate with AI researchers indeed tend to lead the way in their respective
fields (Supplementary Figures 19a,d), the effect is more salient for men and white researchers (see
Supplementary Figures 19b,e for gender and Supplementary Figures 19c,f for race). The results
are robust when we use alternative methods to identify AI researchers, as mentioned above (see
Supplementary Note 6.3 for details). Together, these results provide some preliminary evidence
that women and URM researchers may see fewer benefits from AI advances.

27
Supplementary Figures

Supplementary Figure 1. The share of AI papers in each discipline. Each panel corresponds to
a discipline, and disciplines are ordered by their share of AI papers in 2019, from the largest share
(upper left) to the smallest share (lower right). The lower-right panel compares different disciplines
in the same plot with the same y-axis.

28
Supplementary Figure 2. AI n-grams and temporal trends in each discipline. Top 5 AI n-
grams per paper for a few major disciplines. These AI n-grams are extracted from papers published
in each discipline in the period 2000-2019, and they are colored according to their relative rankings
in AI papers. The color code is the same as the one in Fig. 1c of the main text.

29
Supplementary Figure 3. Temporal trends in the direct AI impact score for each scientific
discipline. The results cover the period between 1960 and 2019. Each panel corresponds to a
discipline, and disciplines are ordered by their direct AI impact scores in 2019, from largest impact
(upper left) to smallest impact (lower right). The panel in the lower-right corner compares the
direct AI impact score across different disciplines. All curves are under the same y-axis, and the
color of each curve corresponds to the one in each panel.

30
Supplementary Figure 4. Citation impact of papers that mention AI-related terms. (a) The
rate of AI papers (i.e., those that mention at least one AI-related term) being hit papers and the
same rate for non-AI papers (i.e., those that do not mention any AI-related terms) in each research
discipline. Hit papers are defined as the top 5% of papers ranked by total citations in the same field
and same year. Disciplines are presented in ascending order based on their direct AI impact scores.
The horizontal line marks the 5% hit rate. (b) The negative correlation between the direct AI
impact score and the hit paper rate for AI papers in each field. The correlation was determined
using a two-sided Pearson’s correlation test. Linear fit (centre line) of the data with 95%
confidence intervals (error bands) is shown. The horizontal dashed line marks the 5% hit rate. (c)
The share of all citations that both AI and non-AI papers receive in each discipline that are outside-
discipline citations. Disciplines are presented in descending order of the share for non-AI papers.
(d) The negative correlation between the direct AI impact score and the ratio of the share of
outside-field citations in AI papers to the share of outside-field citations in non-AI papers. The
correlation was determined using a two-sided Pearson’s correlation test. Linear fit (centre line) of
the data with 95% confidence intervals (error bands) is shown. The horizontal dashed line marks
1, the value at which AI papers receive the same share of outside-field citations as non-AI papers.

31
 Text of Paper Title: Using Bayesian Networks to Analyze Medical Data

Dependency Parsing

Term Extraction

Verb-Noun Pairs: (using, networks) (analyze, data)

Lemmatization & Grouping

Results: use network analyze data

Supplementary Figure 5. Illustration of the extraction verb-noun pairs from publications.

The analysis uses the dependency parsing algorithm embedded in spaCy, an open-source Python
library for advanced NLP applications. The spaCy dependency parser is used to identify the “dobj”
(direct object) relationship among words in the title of a publication, where verb-noun pairs are
extracted by pairing up verb and noun words under the “dobj” relationship. Then, lemmatization
and standardization are used to clean verbs and nouns, and words with similar meanings are
grouped into one word. This process produces standardized verb-noun pairs.

Supplementary Figure 6. Word cloud of verb-noun pairs extracted from publications and
patents. (a) Verb-noun pairs extracted from the titles of AI-related papers and patents in the period
of 1960-2019. (b) Verb-noun pairs extracted from the titles of biology publications in 2019. (c)
Verb-noun pairs extracted from the titles of chemistry papers in 2019.

32
Supplementary Figure 7. Trends in the direct and potential AI impact. (a) Temporal trends in
the direct AI impact score. The dashed line shows the new score calculated using cumulative AI
n-grams fixed in 2015 and the current AI n-grams. Percentage changes in the scores are shown.
Disciplines are sorted by their direct AI impact score in 2019. The lower-left panel shows all results
colored by disciplines according to each panel. (b) Temporal trends in the potential AI impact
score. The dashed line shows the new score calculated using cumulative AI capabilities fixed in
2015 and the current field tasks. The order and color of each discipline are the same as panel (a).

33
Supplementary Figure 8. Alignment between direct and potential AI impact despite the
within-discipline heterogeneity. The plot shows the percentile of each discipline’s subfields
among all 292 MAG fields. The legend is presented in the bottom-right panel. Fields are ordered
by the direct AI scores in the upper row and the potential AI scores in the lower row.

Supplementary Figure 9. Understanding the increase in AI impact. (a) Illustration of the

decomposition of the increase in AI impact scores. Z is the score at t1, Y+Z is the new score in t2,
and X+Y+Z is the score at t2. The increase in AI impact from t1 to t2 consists of X (the component
that results from new AI capabilities; in green) and Y (the component that results from old AI
capabilities; in red). (b) The correlation between the direct AI impact score in 2015 (Z) and the
increase in 2019 due to old AI capabilities (Y). The correlation was determined using a two-sided
Pearson’s correlation test. Linear fit (centre line) of the data with 95% confidence intervals (error
bands) is shown. (c) The correlation between the direct AI impact score in 2015 (Z) and the
increase in 2019 due to new AI capabilities (X). The correlation was determined using a two-sided
Pearson’s correlation test. Linear fit (centre line) of the data with 95% confidence intervals (error
bands) is shown. All correlations were determined using a two-sided Pearson’s correlation test.

34
Supplementary Figure 10. Understanding the potential vs. direct impact of AI in research
fields. (a) Categorizing fields in the “potential-direct” AI impact plane in 2019. The correlation
was determined using a two-sided Pearson’s correlation test. Linear fit (centre line) of the data
with 95% confidence intervals (error bands) is shown. (b) The differences in the collaboration on
AI between the two groups. (c) The differences in the education in AI between the two groups.
The differences were tested using a two-sided Student’s t-test.

Supplementary Figure 11. Categorizing fields using their “momentum” from 2000 to 2019.
Research fields that shifted from “higher-potential” in 2000 to “higher-direct” in 2019 are in the
“upshift” group; research fields that shifted from “higher-direct” in 2000 to “higher-potential” in
2019 are in the “downshift” group. (a) The differences in the collaboration on AI between the two
groups. (b) The differences in the education in AI between the two groups. The differences were
tested using a two-sided Student’s t-test.

35
Supplementary Figure 12. Correlations between AI education and AI impact. The level of AI
education is calculated using syllabus documents that cite (a) one to five scientific references,
which is a proxy for undergraduate-level courses; (b) at least five scientific references, which is a
proxy for graduate-level courses; (c) at least ten references, which is a proxy for more research-
oriented courses. (d) The level of AI education is calculated by the share of syllabus documents
that cited at least one AI publication. Only syllabus documents from 2014 to 2018 are used for the
calculations in the panels (a-d). (e) The level of AI education is calculated using syllabus
documents with at least five references from 2008 to 2013. (f) The level of AI education is
calculated using syllabus documents with at least five references from 2000 to 2007. Throughout
all panels, the correlation was determined using a two-sided Pearson’s correlation test. Throughout
all panels, linear fits (centre lines) of the data with 95% confidence intervals (error bands) are
shown, and all correlations were determined using a two-sided Pearson’s correlation test.

36
Supplementary Figure 13. The growth in AI publications by different co-authorship types
from 1990 to 2019 in each discipline. (a) Collaborative AI papers, i.e., “domain & CS.” (b)
Domain AI papers, i.e., “domain sole.” (c) The slope of a linear fit for AI papers. The number of
“domain & CS” papers has, on average, a larger slope than “domain-only” papers. Scatters
represent disciplines, bar plots show mean values, and error bars represent standard errors for the
disciplines (n=10). (d) Temporal changes in the share of domain-only AI papers in disciplines.

37
Supplementary Figure 14. Temporal trends in cross-discipline AI collaboration. The share of
collaborative AI papers co-authored by domain experts and computer scientists (i.e., “domain &
CS”) in each discipline. Only AI-related papers published outside of the computer science
discipline in the period 1980-2019 are involved in the calculation. Only top disciplines are shown.

Supplementary Figure 15. Robustness test of results for identifying primary AI authors. (a)
The treemap chart shows the fraction of AI papers. (b) The positive correlation between AI impact
and the share of collaborative AI papers. The correlation was determined using a two-sided
Pearson’s correlation test. Linear fit (centre lines) of the data with 95% confidence intervals (error
bands) is shown. (c) The increasing trends in the share of collaborative AI papers in major
scientific disciplines.

38
Supplementary Figure 16. Robustness tests of the result using two alternative definitions of
AI-related papers. (a) Identify AI research using the top 20 most frequent AI n-grams. (b)
Identify AI research using four arXiv subject categories. The direct AI impact scores of research
disciplines under different definitions are highly correlated with each other. Throughout all panels,
the correlations were determined using a two-sided Pearson’s correlation test, and linear fits
(centre lines) of the data with 95% confidence intervals (error bands) are shown.

Supplementary Figure 17. The high correlation between direct AI impact scores. The scores
for disciplines are calculated using different measures. (a) Comparing the scores based on the dot
product measure and the sum measure. (b) Comparing the scores based on the dot product measure
and the weighted average measure. (c) Comparing the percentiles of raw values based on the dot
product measure and the sum measure. (d) Comparing the percentiles of raw values based on the

39
dot product measure and the weighted average measure. Throughout all panels, the correlation
determined using a two-sided Pearson’s correlation test, and linear fits (centre lines) of the data
with 95% confidence intervals (error bands) are shown.

Supplementary Figure 18. The average hit rate of papers published by authors before and
after engaging in AI research. (a) The results for authors across all demographic groups. (b) The
results for men or women authors. (c) The results for white, Asian, or URM authors. Here only AI
researchers are considered. The x-axis is the time difference between the publication year and the
starting year of each author’s AI engagement.

Supplementary Figure 19. The probability of authors leading the way in their field is based
on two metrics. (a-c) Their number of papers. (d-f) Their number of hit papers. Panels (a) and (d)
show the results for all demographic groups. Panels (b) and (e) show the results for gender, and
Panels (c) and (f) show the results for race.

40
Supplementary Tables

Supplementary Table 1. Alignment between direct and potential AI impact. The top three
subfields within each discipline by the direct AI impact score and the potential AI impact score.

No. Discipline Top 3 subfields (direct AI impact) Top 3 subfields (potential AI impact)
pattern recognition, machine learning, computer pattern recognition, computer vision, machine
1 computer science
vision learning
2 mathematics algorithm, mathematical optimization, statistics algorithm, statistics, computational science
computer engineering, industrial engineering, computer engineering, control engineering,
3 engineering
control engineering embedded system
4 geology remote sensing, soil science, mineralogy remote sensing, geodesy, petroleum engineering
5 physics meteorology, acoustics, medical physics medical physics, acoustics, meteorology
6 geography remote sensing, cartography, meteorology remote sensing, geodesy, cartography
environmental soil science, water resource management, environmental engineering, soil science, water
7
science environmental engineering resource management
cognitive science, cognitive psychology, mathematics education, cognitive psychology,
8 psychology
communication applied psychology
mineralogy, chromatography, environmental
9 chemistry mineralogy, analytical chemistry, food science
chemistry
econometrics, management science, actuarial operations management, econometrics,
10 economics
science environmental economics
biological system, computational biology, biological system, computational biology, animal
11 biology
neuroscience science
12 medicine nuclear medicine, optometry, medical physics nuclear medicine, medical physics, radiology
optoelectronics, nanotechnology, composite
13 materials science composite material, metallurgy, optoelectronics
material
risk analysis, actuarial science, process process management, risk analysis, actuarial
14 business
management science
15 sociology communication, demography, regional science regional science, communication, demography
16 philosophy linguistics, humanities, epistemology linguistics, humanities, epistemology
17 political science public relations, law, public administration public relations, public administration, law
18 art visual arts, humanities, aesthetics humanities, visual arts, aesthetics
19 history archaeology, genealogy, ethnology archaeology, genealogy, ethnology

Supplementary Table 2. The top 20 most frequent AI n-grams based on AI publication in

MAG. These AI n-grams can be matched to the MAG “field of study” categories from the L1 level
to the L5 level.
No. AI n-grams (MAG field categories) No. AI n-grams (MAG field categories)
1 machine learning 11 recurrent neural network
2 convolutional neural network 12 decision tree
3 deep learning 13 reinforcement learning
4 support vector machine 14 supervised learning
5 deep neural networks 15 pattern recognition
6 artificial intelligence 16 natural language processing
7 computer vision 17 genetic algorithm
8 random forest 18 machine translation
9 artificial neural network 19 word embedding
10 generative adversarial network 20 extreme learning machine

41
Supplementary References
1. Sinha, A., et al., An overview of Microsoft Academic Service (MAS) and applications, in
Proceedings of the 24th International Conference on World Wide Web. 2015, WWW: Florence,
Italy. p. 243-246.
2. Wang, K.S., et al., A review of microsoft academic services for science of science studies.
Frontiers in Big Data, 2019. 2.
3. Wang, K.S., et al., Microsoft Academic Graph: When experts are not enough. Quantitative
Science Studies, 2020. 1(1): p. 396-413.
4. Palayew, A., et al., Pandemic publishing poses a new COVID-19 challenge. Nature Human
Behaviour, 2020. 4(7): p. 666-669.
5. Viglione, G., Are women publishing less during the pandemic? Here's what the data say.
Nature, 2020. 581(7809): p. 365-366.
6. Raynaud, M., et al., Impact of the COVID-19 pandemic on publication dynamics and non-
COVID-19 research production. BMC Medical Research Methodology, 2021. 21(1): p. 255.
7. Gao, J., et al., Potentially long-lasting effects of the pandemic on scientists. Nature
Communications, 2021. 12(1).
8. Singh Chawla, D., Massive open index of scholarly papers launches. Nature, 2022.
9. Priem, J., H. Piwowar, and R. Orr, OpenAlex: A fully-open index of scholarly works, authors,
venues, institutions, and concepts. arXiv preprint arXiv:2205.01833, 2022.
10. Lin, Z., et al., SciSciNet: A large-scale open data lake for the science of science research.
Scientific Data, 2023.
11. Microsoft Academic. Microsoft Academic Graph. 2021; Available from:
https://zenodo.org/record/6511057.
12. Microsoft Academic. Expanding Concept Understanding in Microsoft Academic Graph. 2020;
Available from: https://www.microsoft.com/en-
us/research/project/academic/articles/expanding-concept-understanding-in-microsoft-
academic-graph/.
13. Shen, Z., H. Ma, and K. Wang, A web-scale system for scientific knowledge exploration, in
Proceedings of Association for Computational Linguistics. 2018: Melbourne, VIC. p. 87-92.
14. PatentsView. USPTO PatentsView. 2019; Available from: http://www.patentsview.org.
15. Toole, A., C. Jones, and S. Madhavan, Patentsview: An open data platform to advance science
and technology policy. 2021, USPTO Economic Working Paper No. 2021-1.
16. USPTO. USPTO Research datasets. 2020; Available from: https://www.uspto.gov/ip-
policy/economic-research/research-datasets.
17. Nowogrodzki, A., Mining the Secrets of College Syllabuses. Nature, 2016. 539(7627): p. 125-
126.
18.Chau, H., et al., Connecting higher education to workplace activities and earnings. 2023. 18(3):
p. e0282323.
19. Börner, K., et al., Skill discrepancies between research, education, and jobs reveal the critical
need to supply soft skills for the data economy. Proceedings of the National Academy of
Sciences, U.S.A., 2018. 115(50): p. 12630-12637.
20. National Center for Science and Engineering Statistics. Survey of Doctorate Recipients. 2019;
Available from: https://www.nsf.gov/statistics/srvydoctoratework/.
21. Huang, J., et al., Historical comparison of gender inequality in scientific careers across
countries and disciplines. Proceedings of the National Academy of Sciences, U.S.A., 2020.
117(9): p. 4609-4616.

42
22. Kozlowski, D., et al., Avoiding bias when inferring race using name-based approaches. Plos
One, 2022. 17(3).
23. Santamaria, L. and H. Mihaljevic, Comparison and benchmark of name-to-gender inference
services. Peerj Computer Science, 2018.
24. NCSES. 2021 Survey of Doctorate Recipients. 2021; Available from: https://www.nsfsdr.org/.
25. Bloom, N., et al., The diffusion of disruptive technologies. NBER Working Paper w28999,
2021.
26. Acemoglu, D., et al., AI and jobs: Evidence from online vacancies. NBER Working Paper
w28257, 2020.
27. Bianchini, S., M. Muller, and P. Pelletier, Artificial intelligence in science: An emerging
general method of invention. Research Policy, 2022. 51(10).
28. Littmann, M., et al., Validity of machine learning in biology and medicine increased through
collaborations across fields of expertise. Nature Machine Intelligence, 2020. 2(1): p. 18-24.
29. Cockburn, I.M., R. Henderson, and S. Stern, The impact of artificial intelligence on innovation:
An exploratory analysis, in The Economics of Artificial Intelligence: An Agenda. 2018,
University of Chicago Press: Chicago, IL. p. 115-146.
30. Klinger, J., J. Mateos-Garcia, and K. Stathoulopoulos, A narrowing of AI research? arXiv
preprint arXiv:2009.10385, 2020.
31. Tang, X., et al., The pace of artificial intelligence innovations: Speed, talent, and trial-and-
error. Journal of Informetrics, 2020. 14(4): p. 101094.
32. Krenn, M., et al., Forecasting the future of artificial intelligence with machine learning-based
link prediction in an exponentially growing knowledge network. Nature Machine Intelligence,
2023. 5: p. 1326-1335.
33. Klinger, J., J. Mateos-Garcia, and K. Stathoulopoulos, Deep learning, deep change? Mapping
the development of the Artificial Intelligence General Purpose Technology. arXiv preprint
arXiv:1808.06355, 2018.
34. Frank, M.R., et al., The evolution of citation graphs in artificial intelligence research. Nature
Machine Intelligence, 2019. 1(2): p. 79-85.
35. AlShebli, B., et al., China and the US produce more impactful AI research when collaborating
together. arXiv preprint arXiv:2304.11123, 2023.
36. AlShebli, B., et al., Beijing's central role in global artificial intelligence research. Scientific
Reports, 2022. 12(1).
37. Nivre, J. and J. Nilsson, Pseudo-projective dependency, in Proc. of the 43rd Annual Meeting
of the Association for Computational Linguistics (ACL). 2005, ACL. p. 99-106.
38. Honnibal, M. and M. Johnson, An improved non-monotonic transition system for dependency
parsing, in Proceedings of the 2015 Conference on Empirical Methods in Natural Language
Processing. 2015, Association for Computational Linguistics. p. 1373-1378.
39. Benetka, J.R., J. Krumm, and P.N. Bennett, Understanding context for tasks and activities, in
Proceedings of the 2019 Conference on Human Information Interaction and Retrieval. 2019,
ACM. p. 133-142.
40. Fujii, H. and S. Managi, Trends and priority shifts in artificial intelligence technology
invention: A global patent analysis. Economic Analysis and Policy, 2018. 58: p. 60-69.
41. Giczy, A.V., N.A. Pairolero, and A.A. Toole, Identifying artificial intelligence (AI) invention:
A novel AI patent dataset. Journal of Technology Transfer, 2022. 47(2): p. 476-505.

43
42. Santos, R.S. and L.L. Qin, Risk Capital and Emerging Technologies: Innovation and
Investment Patterns Based on Artificial Intelligence Patent Data Analysis. Journal of Risk and
Financial Management, 2019. 12(4).
43. Verendel, V., Tracking artificial intelligence in climate inventions with patent data. Nature
Climate Change, 2023. 13: p. 40-47.
44. WIPO, PATENTSCOPE Artificial Intelligence Index. 2020.
45. WIPO, WIPO Technology Trends 2019 – Artificial Intelligence. 2019, World Intellectual
Property Organization: Geneva.
46. Miric, M., N. Jia, and K. Huang, Using supervised machine learning for large-scale
classification in management research: The case for identifying artificial intelligence patents.
Strategic Management Journal, 2023. 44(2): p. 491-519.
47. Webb, M., The impact of artificial intelligence on the labor market. SSRN 3482150, 2019.
48. Felten, E.W., M. Raj, and R. Seamans, A method to link advances in artificial intelligence to
occupational abilities. AEA Papers and Proceedings, 2018. 108: p. 54-57.
49. Felten, E., M. Raj, and R. Seamans, Occupational, industry, and geographic exposure to
artificial intelligence: A novel dataset and its potential uses. Strategic Management Journal,
2021. 42(12): p. 2195-2217.
50. Brynjolfsson, E. and T. Mitchell, What can machine learning do? Workforce implications.
Science, 2017. 358(6370): p. 1530-1534.
51. Brynjolfsson, E., T. Mitchell, and D. Rock, What can machines learn, and what does it mean
for occupations and the economy? AEA Papers and Proceedings, 2018. 108: p. 43-47.
52. Kogan, L., et al., Technology, vintage-specific human capital, and labor displacement:
Evidence from linking patents with occupations. 2022, NBER Working Paper No. w29552.
53. Frank, M., Y.Y. Ahn, and E. Moro, AI exposure predicts unemployment risk, in arXiv preprint
arXiv:2308.02624. 2023.
54. Atalay, E., et al., The evolution of work in the United States. American Economic Journal:
Applied Economics, 2020. 12(2): p. 1-34.
55. Kiperwasser, E. and Y. Goldberg, Simple and accurate dependency parsing using bidirectional
LSTM feature representations. Transactions of the Association for Computational Linguistics,
2016. 4: p. 313-327.
56. Kim, I.C. and Y.G. Jung, Using Bayesian networks to analyze medical data, in Proceedings of
the Third International Workshop on Machine Learning and Data Mining in Pattern
Recognition. 2003: Leipzig, Germany. p. 317-327.
57. Leskovec, J., A. Rajaraman, and J.D. Ullman, Mining of massive data sets. 2020, New York,
NY: Cambridge University Press.
58. Wu, L., L. Hitt, and B.W. Lou, Data analytics, innovation, and firm productivity. Management
Science, 2020. 66(5): p. 2017-2039.
59. OSP. Open Syllabus Dataset Documentation. 2020; Available from:
https://docs.opensyllabus.org/index.html.
60. Malitz, G.S., A classification of instructional programs (CIP). 1987, Washington, D.C.: Center
for Education Statistics, Office of Educational Research and Improvement, US Department of
Education.
61. Cook, K., et al., The future of work in black america. 2019, McKinsey Insights.
62. Autor, D., The faltering escalator of urban opportunity, in Securing Our Economic Futue.
2020, The Aspen Institute: Washington, DC, USA. p. 108-136.

44
63. Cornell University Library. arXiv Dataset: arXiv dataset and metadata of 1.7M+ scholarly
papers across STEM. 2023; Available from: https://www.kaggle.com/datasets/Cornell-
University/arxiv/data.
64. Sourati, J. and J. Evans, Complementary artificial intelligence designed to augment human
discovery. arXiv preprint arXiv:2207.00902, 2022.
65. Baruffaldi, S., et al., Identifying and measuring developments in artificial intelligence: Making
the impossible possible, in OECD Science, Technology and Industry Working Papers No.
2020/05. 2020.
66.Van Buskirk, I. Name-Based Gender Classification. 2021; Available from: https://osf.io/tz38q/.
67. Van Buskirk, I., A. Clauset, and D.B. Larremore, An open-source cultural consensus approach
to name-based gender classification, in Proceedings of the International AAAI Conference on
Web and Social Media. 2023. p. 866-877.
68. Liang, L. and D.E. Acuna. demographicx: A Python package for estimating gender and
ethnicity using deep learning transformers. 2021; Available from:
https://doi.org/10.5281/zenodo.4898367.
69. Ke, Q., et al., A dataset of mentorship in bioscience with semantic and demographic
estimations. Scientific Data, 2022. 9(1): p. 467.

45