Reading minds-controlling minds

Though it iscertainly true that every thought has a neural correlate – to say that isjust to say that thoughts are realized in brains, so the claim should... Unless and until we develop

Much of the interest and anxiety provoked by new and developingneuroscientific technologies is centered around two issues: theextent to which these technologies might allow their users to readthe thoughts of people, and (as if that prospect was not disturbingenough) the extent to which these technologies might actually beused to control people Some commentators believe that one or both

of these issues are pressing, in the sense that the relevant gies will soon be available; some even believe that these technologiesalready exist In this chapter, we will ask how worried we should be.Are these technologies imminent? And if they are, are they asthreatening as they appear?

technolo-m i n d r e a d i n g a n d technolo-m i n d c o n t r o l l i n g

There has been a great deal of interest in the possibility of ‘‘brainreading’’ as a lie detection technology The problems with existing liedetectors are well known: they produce high rates both of falsepositives and of false negatives, and they can be ‘‘beaten’’ by peoplewho deliberately heighten their responses to control questions,which are used to establish a baseline for comparison In its overview

of current lie-detection techniques, the US National ResearchCouncil concluded that there is ‘‘little basis for the expectation that

a polygraph test could have extremely high accuracy’’ (NationalResearch Council 2003: 212) The reasons for this conclusion aremany: because the responses measured are not uniquely involved indeception, because they include responses that can be deliberatelycontrolled and because the technology is difficult to implement in

the real-world The authors conclude that further research andinvestment can be expected to produce only modest gains in accu-racy For these kinds of reasons, conventional polygraph results areinadmissible as evidence in most jurisdictions.

It is recognition of the severe limitations of current liedetection technologies that is responsible, in important part, for thecurrent interest in lie detectors that directly ‘‘read’’ the mind.Polygraph machines are sensitive to changes in somatic states such

as skin conductance (changes in the resistance of the skin to tricity, which is a good indicator of increased sweating), heart rateand blood pressure Now, the problem with these measures is that,

elec-at best, they are only correlelec-ated with deliberelec-ate deception They areindications of increased nervousness, and can be manipulated.Moreover, even with a naı¨ve subject who does not know how tomanipulate these responses, the correlation between the responsesand deliberate deception is far from perfect; hence the false positives(where these responses peak but the subject is not lying), and thefalse negatives (where the subject is lying, but their somaticresponses do not reflect this fact) Proponents of neurologicallybased lie detection argue, apparently plausibly, that such technolo-gies would not be subject to these limitations It might be possible

to fool a system that measures only stand-ins for lies, but it is notpossible to fool a system capable of honing in on the lies them-selves Of course, lies are not directly detectable, but the technologymight be capable of doing the next best thing Thoughts are realizedneurologically; in the jargon of cognitive science, they have neuralcorrelates The correlation between the lie and its neural correlate

is, by definition, perfect Hence, if the lie detection technologycan hone in on the neural correlates of lies, it will not give falsepositives or false negatives, and it will not be able to be fooled.Brains do not lie

The problem with this line of thought is obvious Though it iscertainly true that every thought has a neural correlate – to say that isjust to say that thoughts are realized in brains, so the claim should

be uncontroversial to anyone who does not believe in substancedualism – it does not follow that for every type of thought there is adistinct neural correlate Indeed, the latter claim is rather implau-sible, given that there are endless ways in which thoughts might becategorized (‘‘concerning animals with fewer than four legs,’’ ‘‘con-cerning either my maiden aunt or a ming vase,’’ ‘‘concerning an oddnumber of elephants’’ – the categories are limited only by our inge-nuity) But the idea that we can detect lies by honing in on theirneural correlates requires that to the thought-type ‘‘deliberatedeception’’ there corresponds one, or at any rate a tractable number,

of distinct neural correlates Unless and until we develop the nology to translate neural correlates into thoughts (a prospect weshall consider later in this chapter), we can only hope to detect lies

tech-if we discover that there is something distinctive about deliberatedeception; something we can look for in brain scans

So reliable neurological detection depends upon there beingsomething neurologically distinctive about deliberate deception Isthere? There are various proposals for such distinctive correlates,either of the entire class of deliberate deception, or of various smallersegments of that class Perhaps the best-known technique of neuro-logical lie detection, ‘‘brain fingerprinting,’’ targets one possibledistinctive correlate Brain fingerprinting does not aim to detect anyand all deliberate deception; instead, it is aimed at detecting guiltyknowledge The technology uses memory and encoding relatedmultifaceted electroencephalographic response (MERMER), whichcombines electroencephalography (EEG) data from several sites onthe scalp Electrodes are attached to the subject’s scalp at severalsites, and brain activity is measured while the subject is askedquestions or shown pictures If the subject recognizes a picture or

is familiar with the content of the question, they are supposed toexhibit a characteristic response, the P300 wave (so called because

it occurs 300 milliseconds after the stimulus) If they do not havethe relevant knowledge, the amplitude of the P300 wave will besignificantly smaller

MERMER and brain fingerprinting are the brainchildren ofLawrence Farwell; Farwell has aggressively commercialized thetechnology, through his company Brain Fingerprinting Laboratories.Farwell claims that brain fingerprinting has several advantages overpolygraphy Because the P300 wave is involuntary, it does not dependupon the cooperation of the person being tested (with one exception:the subject must sit still during testing) This makes it especiallysuitable for testing people who might be planning a crime, forinstance suspected terrorists The tester might show them bomb-making equipment, and examine their EEG Most importantly, Far-well claims that brain fingerprinting cannot be manipulated.Whereas clever criminals can beat conventional lie-detection tests,they cannot control the involuntary P300 response In one test of thetechnology, subjects were instructed to attempt to conceal theknowledge being probed; nevertheless the guilty knowledge wasdetected There were no false positives, false negatives or inde-terminate cases (Farwell and Smith2001).

Farwell and colleagues claim these kinds of results are a tacular vindication of brain fingerprinting However, there are anumber of problems with the technology First, the method of ana-lysis used by Farwell is proprietary and undisclosed; for that reasonthere cannot be independent testing of its validity What few teststhere are for brain fingerprinting come exclusively from Farwell’slaboratory Independent testing is, of course, the gold standard ofgood science; in its absence, we are entitled to a high degree ofscepticism (Wolpe et al 2005) Moreover, even assuming that Far-well’s tests have been scrupulously conducted, and his own invest-ment in the technique has had no effect in biasing his results(consciously or unconsciously), his sample sizes are too small toyield a great deal of confidence

spec-Moreover, there are difficulties concerning the ecologicalvalidity of the technology; that is, the extent to which the findingscan be generalized to the world outside the laboratory It requirescarefully controlled situations; in particular, it requires that the

tester possess information that they know will be available only tothe perpetrator of the crime Sometimes such information will beknown, but probably only in the minority of cases (Tancredi2004).Suppose the subject exhibits a P300 to images of a terrorist trainingcamp Should we conclude that they have trained as a terrorist, orthat they watch CNN? In fact, in the only field study to date on thetechnology, it performed at around chance accuracy (Miyake et al.

1993)

Finally, it appears that Farwell may well be wrong in thinkingthat P300 detection methods cannot be beaten The P300 wave is aresponse that is produced when the stimulus is meaningful to thesubject; it can therefore be manufactured by any method that makesirrelevant probes (used to set the baseline for comparison of waveamplitude) meaningful to them Apparently, this is not very difficult.Rosenfeld and colleagues (Rosenfeld et al.2004) instructed subjects

to perform a variety of covert acts in response to irrelevant stimuli:wiggling toes, pressing the fingers of the left hand onto their legs andimagining the experimenter slapping them in the face These coun-termeasures were sufficient to defeat Farwell’s six-probe paradigm.Reaction-time data remained a significant predictor of guilt on a one-probe variety of the test, but this is of little comfort to proponents ofMERMER-based guilty knowledge tests; the six-probe test is neces-sary to avoid too great a rate of false positives as a consequence ofsubjects’ finding the probe meaningful for coincidental reasons

It is apparent that the P300 test has not, so far, lived up to thehype However, it may well improve, as the hardware gets moresophisticated, the algorithms that interpret the data are refined andthe experimenters find better ways to probe guilty knowledge.Moreover, the P300 test is only one of several deceit-detectiontechnologies currently under investigation, some of which are alsoaimed at detecting lies by reading brains Langleben and colleagues(Langleben et al 2002) used fMRI to scan the brains of subjectsengaged in intentional deceit; they discovered that areas of theanterior cingulate cortex and the superior frontal gyrus were more

active during deception than when subjects responded truthfully.Once again, there are questions concerning the ecological validity ofthe technique: how well will it generalize from the controlledlaboratory with willing subjects to the outside world where condi-tions are uncontrolled and subjects are uncooperative? Even in thelaboratory, the accuracy of the test is not all that high: it showed abetween-group difference between deceivers and the truthful, but theeffect is not great enough to identify individual deceivers with a highdegree of confidence In addition, technical limitations of fMRI – itsrelative lack of spatial and temporal resolution – will probably need

to be overcome before the test has an acceptable degree of reliability.Once again, however, the technology and the testing methods can

be expected to improve

What does the foreseeable future hold? I think it is safe to claimthat the kind of mind reading technology which is most feared,which can scan the brains of subjects and reveal intimate detailsabout their thoughts, without their knowing that they are under themental microscope, is (at least) a long way off The most promisingmethods of mind reading require that we build up a set of data on anindividual subject: we need to establish a baseline for responses weknow to be truthful, against which to compare the probes of interest(see Illes et al.2006 for review) Conditions must be carefully con-trolled and the subject (relatively) cooperative Moreover, neitherEEG equipment nor, especially, fMRI equipment, is anywhere nearportable or concealable We can expect to see mind-reading tech-nology that is of some help in detecting deception in the laboratory,and that can therefore be used in the kinds of situations in whichpolygraphy is employed today, long before we see covert surveillance

of thoughts – if indeed that ever turns out to be possible

What about laboratory tests for mental states and dispositionsother than lies? Once again, work is proceeding along several fronts.Many studies have shown brain alterations associated with chronicschizophrenia; the possibility therefore exists that the disease could

be diagnosed on the basis of brains scans (Farah and Wolpe 2004)

Some researchers claim to have discovered identifiable neuralcorrelates of normal personality traits The work of Canli and col-leagues (2001; 2002) is the best known and most interesting in thisvein They found that extraversion was correlated with particularkinds of responses to images with positive emotional qualities,whereas neuroticism was associated with differences in responses toimages with negative emotional content Phelps and colleagues(2000) used fMRI to study the responses of white subjects to photo-graphs of black faces They found a correlation between the degree ofactivation of the amygdala and negative evaluation of blacks Weshall return to this study shortly.

What are the prospects for a genuine mind-reading machine –one that is capable of interpreting brain states more generally? Out-side of personality traits, neuroscientists have had some success indetecting the neural correlates of the orientation of lines to which asubject is attending (Kamitani and Tong2005); when subjects viewed

a visually ambiguous figure, the researchers were able, on the basis offMRI data, to determine how the subject was resolving the ambi-guity Once again, the data was interpretable only after an initial

‘‘training’’ run was used to establish a baseline Quiroga et al (2005)claim to have been able to isolate neural correlates of a much widerrange of representations In their small study, they apparentlyshowed that representations of a single person, building, or a singleclass of objects – e.g., cartoons from The Simpsons – are encoded insuch a way that, no matter how they are presented, they activatespecific neurons Thus, one of their subjects had a neuron thatresponded preferentially to pictures of Jennifer Aniston, no matterwhat angle the picture was taken from, and relatively little to pic-tures of other people, famous or non-famous Another subject had aneuron that responded preferentially to pictures of (what he took tobe) the Sydney Opera House, as well as to the words ‘‘Sydney Opera,’’but not to other buildings or people

This study used recordings of single neurons, rather than fMRI

or EEG techniques It might seem to provide the basis for a much

more powerful mind-reading technique, since the range of thoughtsthat it could detect is far wider than other techniques However,there are good reasons to suggest that it will not result in a mind-reading machine anytime soon First, the technology is invasive,requiring electrodes to be implanted deep in the brain (Quiroga andcolleagues were only able to carry out the experiment because theyhad a pool of intractable epileptics, who required surgery, to draw on.Electrodes are implanted in the brains of such subjects to locate thefoci of seizures, to allow the surgeon to locate the precise area forintervention) Second, once again the technique required training andcooperation Even if we could get single neuron recordings fromsubjects, we could not interpret them unless we already had a set ofdata showing correlations between the firing of the relevant neuronsand particular mental states Third, the authors themselves cautionthat the fact that they were able, in the short time they had available,

to discover pictures to which the individual neurons respondedsuggests that the neurons probably respond to other images as well Ifthe Jennifer Aniston neuron responded only to pictures of Jen, itwould be nothing short of a miracle that the researchers had beenable to hit upon its precise stimulus But if the neurons respond

to many different images – and perhaps to sensations and abstractthoughts as well – then even the possession of single-cell recordingsplus a set of correlation data will not be sufficient to tell us what thesubject is thinking We may have to conclude that he is thinkingabout Jennifer Aniston, or parliamentary democracy, or Friday NightFootball, or a pain in his toe, or something else for which we don’thave any data yet

The development of a general mind-reading technology, able toread the thoughts of people even in the absence of preliminarytraining and the establishment of a baseline, is possible only if there

is a great deal of commonality in the neural correlates of mentalstates across persons That is, it will only be possible to construct adevice to read the thoughts of anyone – whether for the purposes ofdetecting potential terrorists or selling them cola – if it is possible to

construct some kind of translation manual, which details thecorrelations between particular brain states and particular thoughts.

If my thought that ‘‘elephants are gray’’ has neural correlates whichare very different from your thought that ‘‘elephants are gray’’, thenconstructing the translation manual will be difficult or impossible(impossible if there are few commonalities across the population;difficult if the differences are tractable – for instance, if there areidentifiable groups, between which the neural correlates of thoughtsdiffer markedly, but within which there is a great deal of common-ality – just as there is a great deal of commonality within, but notbetween, the vocabularies of different languages)

Moreover, even the construction of a mind-reading machinereliably able to read the thoughts of a single person, upon whom themachine has been trained, depends upon our thoughts having stableneural correlates across time Perhaps my thought that ‘‘elephantsare gray’’ today has a very similar neural realization to the samethought, in my head at least, tomorrow, but perhaps next week, ornext month, or next year, it will be quite different

There is already evidence for some kinds of commonalitieswithin and across subjects The method used by Kamitani and Tong(2005) to detect the orientation of lines to which a subject isattending uses data from extensive testing of the visual systems ofmonkeys; from this data, we know that orientation is represented inthe early stages of visual processing in ways that are consistentacross primates However, the degree of consistency is not sufficient

to underpin the development of a mind-reading machine: an initialset of data is necessary to make the neural activity meaningful.Thanks to the data on the primate visual system, we know where tolook for orientation-tuned neurons, but gathering data on individualsubjects remains indispensable for applying the technique

It’s not difficult, however, to think of ways in which this datacould be gathered unobtrusively; that is, without the subjects’ beingaware that it is taking place We could flash lines in such a waythat subjects had their attention attracted to them, and use this

information to hone in on the relevant neurons Perhaps we could,but so what? This is, after all, a relatively uninteresting mental state:

it is difficult to imagine a realistic scenario in which knowing theorientation of lines to which subjects are attending is importantenough to justify the massive investment necessary to justifydesigning and building a machine able to provide such information.Moreover, even if such a machine were in existence, we would havelittle reason to worry about it There are good reasons to worry about

a loss of privacy, but few to worry about the loss of this kind ofprivacy in particular Apart from a very few, very peculiar, situations,none of us will be worried if others are capable of determining theorientation of lines we are attending to

But mightn’t the results be capable of generalization; that is,mightn’t we expect to be able gradually to expand the range of mentalstates we are capable of detecting? The answer depends, once again,upon the degree of commonality in the neural correlates of thoughtsacross subjects and within subjects across time The visual system,especially the early visual system, is relatively (the emphasis is verynecessary here) simple, and the relationship between its contents andwhat it represents relatively easy to decode It may turn out, for all

we know at the moment, that more complex and abstract thoughtshave neural correlates that are far less stable across time and acrosssubjects In recent work, Haynes and colleagues were able to identifywhether a subject had chosen to add or to subtract two numbers withseventy percent reliability (Haynes et al in press) Arithmeticaloperations are at a greater level of abstraction than visual perception,but this is still a long way off from complex conceptual thought.Indeed, such operations, and all the other kinds of mental states that

we have hitherto been able to detect via fMRI, may be subserved bybrain modules, innate brain structures dedicated to specific tasks If

an important cognitive task regularly confronted our distant tors in the environment of evolutionary adaptation, then a modulemay have developed to perform it efficiently Since brain modulesare (typically) discrete entities, and perform a regular function in a

ances-predictable way, we can expect that it will prove relatively easy todecode their processes, in such a way as to be able to predict theiroutput But our more abstract beliefs are unlikely to be processed bymodules Instead, they are far more likely to be handled by domain-general mechanisms In general, the higher the level of abstraction of

a thought, and the more it pertains to matters which were likely to bevariable in (or entirely absent from) the environment of evolutionaryadaptation, the lower the probability that belief acquisition andretention will be handled by modular mechanisms: since theseconditions will usually be satisfied by the kinds of beliefs we arelikely to be concerned to keep private, we have little reason to thinkthat past successes at detecting neural correlates predict disturbingkinds of future successes in these areas

These considerations suggest that detecting neural correlates ofmore abstract thoughts will prove far more difficult than detectingthe neural correlates of the kinds of thoughts processed by modules.Moreover, it may be that domain-general thoughts have neural cor-relates that are far more varied across subjects, and perhaps evenacross time within the brain of a single subject It might, that is, turnout that the neurons involved in my thought that ‘‘elephants aregray’’ are located in quite different regions of the brain to thoseinvolved in the very same thought in your mind; it might even turnout that across time the location of the neurons involved in thatthought in me shifts The degree of neuroplasticity – the ability of thebrain to reorganize itself – is much lower in adults than in children,but we now know that new neural cells are produced throughout thelifespan Cell death is compensated for, at least to some extent, andnew learning forms new neural connections It would be surprising,then, if the neural correlates of particular thoughts were entirelyunaltered over the lifespan There is evidence, indeed, that preciselythe same mental state never occurs twice, at least for some classes ofmental state The act of recalling a past event seems subtly to alterthe memory, so that when the event is recalled a second time, it isthe remade memory that is recalled Older memories are also altered

by the context of recall, and by newer memories since laid down(Schacter 1996) It may be that a very large number of our mentalstates, especially our complex thoughts which represent the world asbeing a certain way, are reshaped by our causal history: since ourbrains develop as a result of the interaction of endogenous resourceswith the environment, my precise history of learning and perceptionwill shape the morphology of my brain (Rose2005) Which neuronsactivate when I represent Jennifer Aniston will depend on when

I learned about her; in what context, both internal and external Thiscontext will certainly differ from person to person

Useful, or threatening, mind-reading machines are not going

to be built soon: not, at least, mind-reading machines capable ofdetecting the details of complex thoughts Machines that detect, say,emotional arousal are another matter: given that the structures foremotional processing are relatively discrete, and relatively invariantacross individuals, such machines are an in-principle possibility.Whether these machines would be more useful for lie detection thanexisting polygraph technology is, however, a moot point: they prob-ably could be beaten in the same kinds of ways After all, people dohave some indirect control over their thoughts: we can learn tocontrol our emotional arousal, at least temporarily, and in at leastsome situations (meditators are proficient at this skill)

The most immediate ethical problem arising from these newtechniques of measuring the neural correlates of mental states stemsfrom the dangers of premature adoption (Wolpe et al.2005) The aura

of prestige and objectivity which surrounds science generally is haps even stronger in relation to the science of the mind at its cuttingedge When neuroscience is applied to produce what are apparentlypictures of thoughts, especially when the technologies are sur-rounded by hype (some of it, though not the bulk, originating fromreputable scientists) these pictures are apt to be given a weight theymay not deserve Reporting on Farwell’s P300 technique, the Dis-covery Channel described it as providing ‘‘an infallible witness’’.ABC news reported that the technology read the mind of subjects too