Sunday, July 6, 2014

Welcome to the World of Technochicracy! Revisiting the FB Debacle


"I have a joly wo, a lusty sorwe" - Chaucer, Troilus and Criseyde

A recent research paper published in PNAS and written by Adam D. I. Kramer - a member of Facebook's Core Data Team - and two information scientists from Cornell, made the news and caused some outcry last week. The authors admitted to having manipulated the news feed of some 700,000 thousand Facebook users back in January 2012; two experiments were conducted in parallel: in one, posts with "positive emotional content" had a lower chance of cropping up in a user's feed, and in the other experiment, many posts with "negative emotional content" were omitted from the feed. 

The theory goes something like this: it is well attested that emotional contagion occurs in face-to-face situations. Without knowing it, we scan other people's faces for clues on their emotional status. Once we've gauged that, mirror neurons in our brain fire away, and we adapt our own facial expressions accordingly. A sort of unconscious mimicry. Negativity breeds negativity, and happiness makes the world go round. Previous studies have focused on nonverbal cues, but what about verbal ones? What about situations where the people are miles apart? If a friend posts a negative status update on Facebook, will I catch the negativity bug and post a negative one myself? Or rather, given a large enough sample ("N=689,003" in this case) is there a statistically significant correlation between the emotions of group members exposed to positive and negative posts?

And the answer, sez Kramer et al., is yes! The connection is tenuous, like "gold to airy thinness beat" (incidentally not a phrase used in the paper), but it is there; about one in a thousand posts exhibited "emotional contagion." Given the scale of Facebook, however, "this would have corresponded to hundreds of thousands of emotion expression in status updates per day."

Relying on a vague "research" clause in the Facebook User Policy, the authors conducted an exercise in manipulation with hundreds of thousands of users. This makes a mockery of the idea of informed consent - a mockery more egregious than the false pretenses used by Stanley Milgram in his harrowing 1961 experiment on authority and obedience. While Milgram's subjects were in the dark about the real purpose of the experiments, at least they knew that they took part in one. We are veering dangerously close to mind control land here. In fact, in one of the first books written on the psychology of brain washing, William Sargent explains how:
"Various beliefs can be implanted in many people after brain function has been sufficiently disturbed by accidentally or deliberately induced fear, anger, or excitement. Of the results caused by such disturbances, the most common one is temporarily impaired judgement and heightened suggestibility. Its various group manifestations are sometimes classed under the heading of 'herd instinct'" (William Sargent, Battle for the Mind: A Physiology of Conversion and Brain-washing, 1957)
Now, Dr. Sargent was a behaviorist (white coat, stopwatch in hand), and the point is not that his analysis is especially lucid (it isn't). No, what is eerie here is the similarity between his Pavlovian notions of human behavior and the underpinnings of the Facebook study: deliberately induced feelings? Heightened sensibility? Herd instinct? Do we hear a bell ringing in the distance?

Here is what the study authors have to say about their methodology:
"Posts were determined to be positive or negative if they contained at least one positive or negative word, as defined by Linguistic Inquiry and Word Count Software (LIWC2007) (9) word counting system, which correlates with self-report and physiological measures of well-being, and has been used in prior research on emotional expression. LIWC (7, 8, 10). LIWC was adapted to run on [...] the News Feed Filtering system, such that no text was seen by the researchers. As such, it was consistent with Facebook's Data Use Policy, to which all users agree prior to creating an account on Facebook, constituting informed consent for this research."
In the wake of this debacle, most people have focused on the second part - the dilution and slippery slopification of "informed consent"; there are excellent pieces on the responsibility of the social scientist (and less-than-stellar apologies), but what really rubs me the wrong way has more to do with the first part, and what I have previously called technochicracy. It is the lingering suspicion that the stunning "big data" vista (complete with the cloud services floating overhead) is a set-piece propped up - like a Potemkin village in the midst of Silicon Valley - in front of a 1950s or 1960s landscape of dumb terminals and behaviorist labs.

This experiment was framed as a groundbreaking study in emotional contagion; thanks to big data crunching and state-of-the-art software, an effect only previously observed in face-to-face interaction in an artificial milieu could now be studied on a massive scale with humans, so to speak, in their natural habitat. And yet, the underlying methodology is so hopelessly crude as to bring to mind Pavlov's experiments on conditioned reflexes in dogs. In the words of his American disciple B. F. Skinner, "Pavlov's attention was directed mostly to the glandular part of this total response, because it could be measured by measuring the flow of saliva." A stimulus (a bell ringing before food is served, an upbeat status message) triggers a physiological response that can be adequately measured in a beaker or by said Linguistic Inquiry and Word Count Software. Conceptually speaking, there is little difference between measuring the flow of saliva and counting positive/negative words in the flow of big data.

In fact, the idea of trigger words in status messages correlating with emotional well-being, at least on the aggregate level, owes more to 19th century positivism than any "cutting edge" science. It makes a mockery of the human condition. Who better to knock down 19th century underpinnings than a 19th century poet? When accused of mixing gravity and levity in Don Juan, Lord Byron sent the following letter to his publisher to answer the critic:
"His metaphor is, that 'we are never scorched and drenched at the same time'. Blessings on his experience! Ask him these questions about 'scorching and drenching'. Did he never play at Cricket, or walk a mile in hot weather? Did he never spill a dish of tea over his testicles in handing the cup to his charmer, to the great shame of his nankeen breeches? [...] Did he never tumble into a river or lake, fishing, and sit in his wet cloathes in the boat, or on the bank, afterwards 'scorched and drenched', like a true sportsman? 'Oh for breath to utter!' - but make him my compliments; he is a clever fellow for all that - a very clever fellow." (Byron’s Letters and Journals VI:207)
If we think of Don Juan as a status update, the reason it is so difficult to parse is (and Byron makes this abundantly clear) that human emotion cannot be reduced to a discrete number on a happiness scale. Even in the extreme, seemingly most clear-cut cases, our sadness or happiness is seldom unalloyed:

"Every cloud has a silver lining..." 

"Ay, in the very temple of Delight / Veil'd Melancholy has her sovran shrine." 

Etc.

But what about the statistically significant (albeit minuscule) effect Kramer et al. observed? There are, in fact, many ways of accounting for it that have nothing to do with emotion whatsoever. I might, for example, tag a friend and quote what s/he says in my status update (the system is far too crude to take quotation into account.) There is also the possibility that a few stray words I have recently read linger in my mind, winding their way into my next status message.

On a more general level, Chomsky's critique of Skinner's language behaviorism bears repeating. One of his points is that the causality between stimulus and response can only be "studied" by multiplying categories or properties of the stimulus-object until it loses all and any pretense of objectivity:
"Consider first Skinner's use of the notions stimulus and response. In Behavior of Organisms he commits himself to the narrow definitions for these terms. [...] Evidently, stimuli and responses, so defined, have not been shown to figure very widely in ordinary human behavior. We can, in the face of presently available evidence, continue to maintain the lawfulness of the relation between stimulus and response only by depriving them of their objective character. A typical example of stimulus control for Skinner would be the response to a piece of music with the utterance Mozart or to a painting with the response Dutch. These responses are asserted to be "under the control of extremely subtle properties" of the physical object or event. Suppose instead of saying Dutch we had said  Clashes with the wallpaper, I thought you liked abstract work, Never saw it before, Tilted, Hanging too low, Beautiful, Hideous, Remember our camping trip last summer? or whatever else might come into our minds when looking at a picture (in Skinnerian translation, whatever other responses exist in sufficient strength). Skinner could only say that each of these responses is under the control of some other stimulus property of the physical object. If we look at a red chair and say red, the response is under the control of the stimulus redness; if we say chair, it is under the control of the collection of properties (for Skinner, the object) chairness, and similarly for any other response. This device is as simple as it is empty." (Noam Chomsky, "A Review of B. F. Skinner's Verbal Behavior", 1967)
An even better example (pace Chomsky) of the "suppose instead" refutation would be Proust's musings on the "petite phrase" he once heard in a sonata:
"When, after that first evening at the Verdurins’, he had had the little phrase played over to him again, and had had sought to disentangle from his confused impressions how it was that, like a perfume or a caress, it swept over and enveloped him, he had observed that it was to the closeness of the intervals between the five notes which composed it and to the constant repetition of two of them that was due that impression of a frigid and withdrawn sweetness; but in reality he know that he was basing this conclusion not upon the phrase itself, but merely upon certain equivalents, substituted (for his mind’s convenience) for the mysterious entity of which he had become aware"
In much the same way as the sensuous effects and enveloping power of the musical phrase are not part and parcel of  its "stimulus properties," every status update containing the word "happy" (or a synonym) is not a property intrinsic to the word "happy." Proust's tentative explanation involving "certain equivalences" is a marvelous debunking of the causality assumption. For all its duh!-simplicity, the "lingering" hypothesis (mine and Proust's) is far too complex to be accounted for by Skinnerian behaviorism (whether of the 1950s stamp or dressed-up in technobabble.) Emotional contagion does exist, but it does so in a complex interplay of facial expression, mimicry and thought - both conscious and unconscious. It cannot be reduced to a stimulus-response scheme.

In fact, when we peak through the cool Matrix-curtain of big data, we find a faded Polaroid from the past - Brylcreemed men in white coats studying lab rats or dogs cowering in cages - men whose theories and methodologies were already debunked by the time horn-rimmed glasses went out of fashion... Welcome to the world of Technochicracy!