MIND

READERS
BRAIN-SCANNING MACHINES MAY SOON BE CAPABLE OF DISCERNING RUDIMENTARY
THOUGHTS AND SEPARATING FACT FROM FICTION BY PHILIP ROSS

IMAGINE A WORLD YOU COULD TRUST—REALLY TRUST— verse with minds trapped inside paralyzed bodies,
where truth was transparent and juries, police, expose to analysis the suppressed fears and desires
locksmiths and gossip columnists were largely of the stormy unconscious, even observe the in-
overthrown. Human society would be orderly, sights and errors by which a student moves to-
boring and as alien as an anthill. ward the solution of a math problem.
This is the promise and the threat of a machine The idea of looking directly at brain activity to
that could read minds. The hoary polygraph has tell truth from falsehood dates back roughly 20
never filled the bill. It measures not thoughts but years, to when J. Peter Rosenfeld of Northwestern
only the indirect physiological consequences of University observed an interesting feature in the
thoughts—blood pressure and respiration, among electroencephalograph, or EEG, a chart of the
others—that hint that a subject may be lying. The brain’s electrical signals as detected on the surface
result, critics charge, is false positives—an honest of the skull. The P300 wave had already been
answer misjudged as a lie—and false negatives—a known to be evoked by oddball cues, such as hear-
lie misjudged as the truth. The courts have long ing one’s name mentioned in a list of other words.
ruled polygraph findings inadmissible as evidence. Rosenfeld found that lying elicited it, too. He is
Just last October the National Research Council now mapping the P300 wave across the scalp to
damned the device as a “blunt instrument,” of lit- get enough spatial resolution to improve the sen-
tle use in ferreting out criminals, spies and terrorists. sitivity of the test.
Greek philosopher Diogenes walked with a The next step appears to have been articulated
lamp, in search of an honest man. Yet why shine for the first time by the often prophetic science
your lamp into someone’s face when you can look columnist David Jones, a.k.a. Daedalus, who
at the very brain? There you might do better than wrote in 1996 that “a modern magnetic-resonance
merely tell truth from lies. You might also con- brain scanner should be a perfect lie detector. . . .

74 SCIENTIFIC AMERICAN SEPTEMBER 2003

COPYRIGHT 2003 SCIENTIFIC AMERICAN, INC.
 Telling the truth should activate just one site in the The two brain images from each category were av-
brain.... Telling a lie should activate two sites: one eraged and compared.
holding the lie and the other holding the truth that It turned out that all areas activated during truth
it is masking.” telling were also triggered during lying but that a
Five years later Daniel Langleben of the Univer- number of areas were active particularly during ly-
sity of Pennsylvania and his colleagues used func- ing. “That suggests that the default position is truth,
tional magnetic resonance imaging (fMRI) to scru- and deception is some sort of process you perform
MELISSA SZALKOWSKI

tinize the brains of subjects acting out a question- on truth,” Langleben remarks. He notes that sever-
and-answer series. Under certain conditions, the al areas activated more during lying— including the
subject would tell a string of falsehoods in such a anterior cingulate cortex and part of the left pre-
manner as to mimic lying; in other conditions, the frontal cortex—are associated with suppression of re-
subject would utter a string of truthful statements. sponse, as when the brain decides to go with one of

www.sciam.com SCIENTIFIC AMERICAN 75

COPYRIGHT 2003 SCIENTIFIC AMERICAN, INC.
 ANTERIOR
CINGULATE
CORTEX

LEFT
PREFRONTAL
CORTEX
A graphical
rendition of a brain
slice highlights
two areas that are
more active when
a subject falsely
denies possessing
the five of clubs
than when the
person tells
the truth.

two conflicting responses and must therefore inhib- polygraphs have almost everything to do with it. In-
it the other one [see illustration above]. deed, polygraphs are often used to instill fear as
According to this theory of “cognitive load,” ac- much as to detect it (like the “fear-o-sensor” that a
tor Sean Connery, when asked his name during the dog waves over a stranger in one of Gary Larson’s
filming of a movie, cannot help but flash to the words cartoons). Second, brain imaging follows a phe-
“Sean Connery”; it is only with a modicum of effort nomenon that is much closer to thought, in the train
that he chokes off that response to say instead, of events, than are the pulse, skin conductance, res-
“Bond. James Bond.” So far two other fMRI groups piratory rate and so on—“output that is 10 times re-
have published similar research; more have written moved from what’s happening in the brain,” Lan-
papers now wending their way to publication. gleben says.
None of these groups have yet claimed much Even fMRI does not sample the neurons them-
power in catching a particular hostile witness in a selves, though, but just the oxygen in the nearby
particular lie. “As a practical method, this thing is bloodstream. More precisely, it measures the ratio
not even in the proof-of-concept stage,” Langleben of oxygenated to deoxygenated blood. The ma-
admits. “In April [2004] we will take the next step chine can pinpoint metabolic activity at good res-
and try to determine the size of the truth-versus-lie olution, of about four millimeters in diameter, yet
effect at a given spot in the brain.” He expects to use it is relatively slow, tracking activity occurring for
a larger sample, 60 to 90 subjects, and to create sit- two seconds or so. That’s not really fast enough to
catch a thought.
To capture that level of complexity would re-
BRAIN IMAGING follows quire recording a signal lasting for mere millisec-
onds, providing a snapshot of, say, calcium ions in
processes that are MUCH CLOSER the neurons themselves. To detect it, however,
would require magnets several times as powerful as
TO THOUGHT than the pulse, skin even Langleben’s four-tesla unit. No such magnets
big enough for humans exist, and, for safety rea-
conductance and respiratory rate sons, none are likely be approved for that purpose.
“I can tell you there won’t be human studies in 20-
LUCY READING ( left); JAMES SALZANO ( right)

measured by A POLYGRAPH. tesla machines,” asserts Marcus E. Raichle, an

fMRI researcher at Washington University. “It can
stimulate the vestibular system, making you feel
uations closer to real-life deception—perhaps a pok- dizzy; it can heat up the brain, manipulating the
er game. (It might be a little hard to simulate, very thing you’re supposed to study.”
though, inside a churning, claustrophobia-inducing Another approach to get good resolution in both
MRI machine.) space and time might come by combining fMRI with
In principle, brain imaging is better than a poly- EEG. One might measure both things at the same
graph, he argues, for two reasons. First, it seems to time, or correlate a lie-detecting component of fMRI
have nothing to do with general anxiety, whereas with a given aspect of EEG. “If we did that, we could

76 SCIENTIFIC AMERICAN SEPTEMBER 2003

COPYRIGHT 2003 SCIENTIFIC AMERICAN, INC.
discard fMRI and use the EEG signal, and it would more sophisticated tool. In principle, law-enforce-
be 10 times cheaper,” Langleben says. ment officers might use the combination technolo-
Although it may be too hard for today’s brain gy to tell not only that a bank robber is lying but
scanners to trap a hostile witness or a cheating that the loot is stashed in the garage.
spouse, they may well suffice to divine certain, sim- A brain decoder that worked on all brains still
ple thoughts of a willing communicant, leading to might not allow for telepathy on the order of a Vul-
a more general form of mind reading. Unlike lie de- can mind meld in the Star Trek series, which enabled
tection, the task, of course, is made simpler when universal translation. An English sentence, beamed
the subject cooperates with the testing. Already into the mind of a non-English speaker, might seem
monkeys with electrodes implanted in motor-con- gibberish. Even if the receiving (or eavesdropping)
trol areas of their brains have been taught, through person spoke the same language, he might be puz-
biofeedback techniques, to convey neural impuls- zled by the idiomatic dialect in which a mind con-
es over an Internet connection to manipulate a ro- verses with itself, with all its coded entries, abbrevi-
botic arm [see “Controlling Robots with the Mind,” ations and emotional associations.
by Miguel A. L. Nicolelis and John K. Chapin; Sci- Concocting near-perfect lie detection may, none-
entific American, October 2002]. Niels Birbaumer
of the University of Tübingen in Germany has re-
ported a degree of success in using biofeedback to fMRI can determine which
train patients immobilized by nerve damage to vary
their brain waves and so to spell out sentences on of 12 SIMPLE CATEGORIES
a computer screen.
But true mind reading must do better, by catch- a subject is contemplating with
ing a word or concept exactly as it forms itself in the
brain. Marcel A. Just of Carnegie Mellon Universi- 80 TO 90 PERCENT accuracy.
ty claims he has done just that with fMRI, by limit-
ing the concepts to a small number and keeping
them very simple— carpentry tools, for instance, or theless, be much easier than making a sophisticated
kinds of dwellings. “We have 12 categories and can thought reader— and almost as dangerous to men-
determine which of the 12 the subjects are thinking tal privacy. Indeed, it would not be necessary to em-
of with 80 to 90 percent accuracy,” he explains. He ploy such a machine—the threat of its use would ex-
is even better at distinguishing brains reading a clear ercise a powerful deterrent force.
sentence from those reading an ambiguous one or As Daedalus concluded, “Like the atom bomb,
imagining a verb as opposed to a noun. it is best reserved as a sort of ultimate social weap-
Just’s colleague Tom Mitchell, a computer sci- on. If widely deployed outside the courtroom, it
entist, has devised a means to classify the complex would make social life quite impossible.”
brain images that their experiments produce. He an-
alyzes them with neural networks, a type of software Philip Ross writes on science and technology
that can tune itself to improve its ability to distin- from New York City. His work has also appeared
guish patterns. “If isolated words can be identified in Acumen Journal of Sciences, IEEE Spectrum,
with some degree of accuracy, it ought to be possi- Forbes, and the New York Times.
ble to do even better with entire sentences,” Mitch-
ell says. That is because sentence structure con- MORE TO E XPLORE
strains the possibilities that the neural network must Event-Related Potentials in the Detection of Deception, Malingering, and False
consider. “If you know that a sentence has two Memories. J. Peter Rosenfeld in Handbook of Polygraph Testing. Edited by
Murray Kleiner. Academic Press, 2001. Preprint available at
words, then one must be a verb, the other a noun. www.psych.northwestern.edu/psych/people/faculty/rosenfeld/NewFiles/P300%
“One experiment I would love to do is to find 20and%20ERP%207-99.pdf
words that produce the most distinguishable brain Brain Activity during Simulated Deception: An Event-Related Functional Magnetic
activity,” he adds. Such words might serve as the Resonance Study. D. D. Langleben, L. Schroeder, J. A. Maldjian, R. C. Gur, S. McDonald,
J. D. Ragland, C. P. O’Brien and A. R. Childress in Neuroimage, Vol. 15, No. 3,
building blocks for a neural interface, much as par- pages 727–732; March 2002. Available at
ticularly discriminable English words were favored www.uphs.upenn.edu/trc/conditioning/neuroimage15–2002.pdf
in the early, limited-vocabulary protocols of voice- The Polygraph and Lie Detection. Board on Behavioral, Cognitive, and Sensory
processing software. Sciences and Education (BCSSE), Committee on National Statistics (CNSTAT).
National Academies Press, 2003. Available at
Should this concept-recognition system work www.nap.edu/books/0309084369/html
with even minimal reliability, it might be coupled Marcel Just’s Laboratory Center for Cognitive Brain Imaging of CMU Web site is
with lie-detecting fMRI software to produce a much available at http://coglab.psy.cmu.edu/index–main.html

www.sciam.com SCIENTIFIC AMERICAN 77

COPYRIGHT 2003 SCIENTIFIC AMERICAN, INC.