Tags
Algorithm, Apply Magic Sauce, Asher Peres, Brexit, Cambridge Analytica, Carl Brigham, Christopher Wylie, Crispin Wright, David Mermin, David Stillwell, Dr Hugh Morrison Queen's University Belfast, Dr Youyou Wu, Jerome Kagan, latent traits, ludwig Wittgenstein, Malcolm Gladwell, Michal Kosinski, microtargeting, Newsweek, Nina Burleigh, Observer, President Donald Trump, Ross & Nisbitt, Stanford University, Thore Graepel, University of Cambridge
Dr Hugh Morrison (The Queen’s University of Belfast [retired])
Despite the fact that research carried out at the University of Cambridge’s Psychometrics Centre is at the very heart of the Cambridge Analytica debate, this research seems to have been subject to little or no serious scrutiny. This is puzzling given that an analysis of the Apply Magic Sauce algorithm, developed at the University, could settle, once and for all, whether tools developed at Cambridge could have influenced the Brexit or Trump votes. The impression has been created that the Cambridge academics, armed only with an individual’s Facebook “likes,” could somehow use this Apply Magic Sauce algorithm to peer into the mind of that individual.
The Cambridge academics (erroneously) portray psychological constructs such as personality and intelligence as “inner”, “private” traits, somehow hidden in mind. It is claimed, however, that provided one has access to an individual’s Facebook likes, one can use the Apply Magic Sauce algorithm to represent numerically that individual’s five-trait personality profile, together with his or her intelligence. It is easy to see how the title of a paper by Michal Kosinski, David Stillwell & Thore Graepel (2013)[1], published in PNAS – “Private traits and attributes are predictable from digital records of human behaviour” – gives the impression that tools developed at Cambridge can somehow “reveal” an individual’s personality/intelligence, given his or her digital footprint. It will be argued in this essay that no scientific basis exists for this claim.
Nina Burleigh, writing in Newsweek (18.06.2017), paints a picture of a dystopian future in which algorithms can be used to infer psychological profiles by stealth: “Big Data, artificial intelligence and algorithms designed and manipulated by strategists like the folks at Cambridge have turned our world into a Panopticon, the 19th century circular prison designed so that guards, without moving, could observe every inmate every minute of every day.” The Cambridge Analytica whistleblower, Cristopher Wylie, goes further in the Observer newspaper of 18.03.2018. According to Wylie, because “personality traits could be a precursor to political behaviour,” (p. 10) then the tools developed at Cambridge could represent a “psychological warfare mind**** tool” (p. 9) capable of disrupting the democratic process itself.
This brief essay asks the important question, “is there a shred of scientific truth in the claim that one can exploit the link between inner mental states (personality traits, intelligence etc.) and voting intention, in order to nudge an individual’s behaviour in the polling booth?” The unequivocal response is “No,” because the Cambridge academics’ claim to have developed an algorithm capable of inferring personality from Facebook likes has no scientific basis.
The Apply Magic Sauce algorithm claims to predict (and quantify) personality traits. This cannot be so for a very simple reason: such traits do not exist. This we know from the extensive rule-following literature (see, for example, Crispin Wright’s 2001 book, “Rails to Infinity”[2]). Dynamic intrinsic attributes are the preserve of Newtonian mechanics, and are not available to psychologists. Personality is a property of the interaction between person and measuring tool rather than an intrinsic attribute of a person; it is a joint property of the individual and the instrument used to measure personality.
One of the towering figures of 20th century thought, Ludwig Wittgenstein (1958, p. 143) dismissed unequivocally the notion of inner mental traits posited by the Cambridge team: “There is a kind of disease of thinking which always looks for (and finds) what would be called a mental state from which all our acts spring as from a reservoir.”[3] One cannot capture an individual’s personality in a number because personality isn’t a property of the individual. Moreover, it is inconceivable that the Cambridge academics were not aware that their interpretation of personality as a quantifiable trait was entirely at odds with serious scholarship in their own discipline.
Ross and Nisbett’s classic textbook The Person and the Situation[4] is a staple of undergraduate psychology. This book argues that personality is so entangled with the context in which it is expressed that it is meaningless to conceive of personality as some free-standing quantifiable inner state. One of psychology’s most respected thinkers, Jerome Kagan, has identified this tendency for some psychologists to picture psychological attributes as traits hidden in mind as undermining their profession.
On page xvii of Psychology’s Ghosts: The Crisis in the Profession and the Way Back[5], Kagan “concludes that ‘agents in a context’ should replace the current, restricted focus on stable properties of individuals that, like their eye colour, are presumably available for expression in all settings.” Professor Kagan points out that measurement in psychology is not a matter of checking up on traits which already exist in mind. Kagan (1998, p. 16)[6] writes: “Most investigators who study ‘anxiety’ or ‘fear’ use answers on a standard questionnaire or responses to an interviewer to decide which of their subjects are anxious or fearful. A smaller number of scientists asks close friends or relatives of each subject to evaluate how anxious that person is. A still smaller group measures the heart rate, blood pressure, galvanic skin response, or salivary cortisol level of their subjects. Unfortunately, these three sources of information rarely agree.”
Kagan is making the point that in order to communicate unambiguously (the hallmark of science) one cannot omit the measuring instrument. Psychological predicates only have definite properties relative to a specified measuring tool. One cannot attribute a definite value to a psychological attribute construed as a property of a person; this attribute only has definite properties relative to a particular instrument. Kagan (1998, p. 77) cautions: “Modern physicists appreciate that light can behave as a wave or a particle depending on the method of measurement. But some contemporary psychologists write as if that maxim did not apply to consciousness, intelligence or fear.” This poses a fundamental problem for the Cambridge project because personalities predicted from Facebook likes involve no interaction with an appropriate measuring tool. In these circumstances ascribing a definite personality makes no sense.
In Malcolm Gladwell’s popular text The Tipping Point,[7] he depicts the incoherence which is the consequence of divorcing psychological predicates from the contexts in which they are expressed, and treating them as traits. Indeed, psychology has a name for the error which afflicts the Cambridge algorithm: “The Fundamental Attribution Error.” According to the Fundamental Attribution Error, behaviour is not determined by some theoretical mental trait within the individual; rather, behaviour can only be understood by examining the interaction between the individual and the measuring instrument. Finally, even if one were to accept that psychological predicates can be represented as latent traits, such models break down at the level of individual (see Borsboom, Mellenbergh & van Heerden, 2003, p. 217)[8], presenting intractable problems for any “personalisation machine” capable of “microtargeting.”
Turning to the claim that one can measure intelligence from Facebook likes, one finds the Cambridge algorithm at odds with one of the central tenets of quantum mechanics (see below). The reader will be aware of the decades-old debate: do intelligence tests measure some inner, hidden mental state which psychologists call “intelligence”, or do they merely measure the ability of the testee to answer a series of questions on something psychologists call an “intelligence test”?
Psychology was aware – as far back as the 1930s – of the profound difficulties associated with interpreting intelligence as a trait: “He [Carl Brigham] recognized that a test score could not be reified as an entity inside a person’s head: Most psychologists working in the test field have been guilty of a naming fallacy which easily enables them to slide mysteriously from the score in the test to the hypothetical faculty suggested by the name of the test. Thus, they speak of sensory discrimination, perception, memory, intelligence, and the like while reference is to a certain objective test situation” (Gould, 1996, p. 262)[9].
The distinguished American physicist, David Mermin (1993, p. 1)[10] notes: “When you measure IQ are you learning something about an inherent quality of a person called “intelligence,” or are you merely acquiring information about how the person responds to something you have fancifully called an IQ test? Until the advent of quantum theory in 1925 physicists were above such concerns. But since then, with the discovery that experiments at the atomic level necessarily disturb the objects of investigation, precisely such reservations have been built into the foundations of physics.” In modern physics, momentum is not an intrinsic property of an electron; rather, it is a property of the electron’s interaction with the measuring instrument.
The introduction to Werner Heisenberg’s 1989 book[11] contains a single sentence with profound implications for the Apply Magic Sauce algorithm: “[T]he reality is in the observations [interactions], not in the electron.” The lesson for Apply Magic Sauce is clear: one can only speak meaningfully about an individual’s intelligence when that person interacts with an intelligence test. Intelligence is not an inner trait and the Cambridge algorithm can no more predict intelligence than it can predict personality.
The Cambridge academics’ central claim for their algorithm is that it can predict an individual’s personality without requiring that person to take a personality test. Alas, removing the measuring instrument actually deprives any references to personality or intelligence of their very meaning. Those who feared that academics at Cambridge had actually developed an algorithm which could derive an individual’s intimate traits from his or her digital footprint, without requiring that person to take the relevant test, can take comfort from the words of the physicist Asher Peres: “Unperformed experiments have no results.”[12] In summary, it seems inconceivable that an algorithm with the conceptual difficulties of Apply Magic Sauce could be deployed to nudge voting intention in any predetermined direction.
drhmorrison@gmail.com
[1] Kosinski, M., Stillwell, D., & Graepel, T. (2013). Private traits and attributes are predictable from digital records of human behaviour. Proceedings of the National Academy of Sciences of the United States of America, 110 (15), 5802-5805.
[2] Wright, C. (2001). Rails to infinity. Cambridge, MA: Harvard University Press.
[3] Wittgenstein, L. (1958). The blue and brown books. Oxford: Blackwell.
[4] Ross, L, & Nisbett, R.E. (1991). The person and the situation. Philadelphia: Temple University Press.
[5] Kagan, J. (2012). Psychology’s ghosts. New Haven: Yale University Press.
[6] Kagan, J. (1998). Three seductive ideas. Cambridge, MA: Harvard University Press.
[7] Gladwell, M. (2000). The tipping point. Boston: Little, Brown & Company.
[8] Borsboom, D., Mellenbergh, G.J., & van Heerden, J. (2003). The theoretical status of latent variables. Psychological Review, 110 (2), 203-219.
[9] Gould, S.J. (1996). The mismeasure of man. London: Penguin Books.
[10] Mermin, D. (1993). Lecture given at the British Association Annual Science Festival. London: British Association for the Advancement of Science.
[11] Heisenberg, W. (1989). Physics and philosophy. London: Penguin Books.
[12] Peres, A. (1978). Unperformed experiments have no results. American Journal of Physics, 46, 745-747.