The experiment
For one week in January 2012, Facebook changed what 689,003 of its users saw. For some, the amount of positive emotional content in the News Feed was reduced; for others, the negative. Nobody was told.
The researchers then measured the emotional language in those users' own subsequent posts. The finding, published in the Proceedings of the National Academy of Sciences on 17 June 2014 under the title Experimental evidence of massive-scale emotional contagion through social networks, was that the manipulation worked in the expected direction: cut the positive content and people wrote slightly fewer positive words and slightly more negative ones, and the reverse when negative content was cut.
The authors were Adam D. I. Kramer of Facebook, with Jamie E. Guillory and Jeffrey T. Hancock.
Facebook’s Menlo Park headquarters. The experiment was run from here on 689,003 users who were not told. Wikimedia Commons — LPS.1, CC0
How big the effect was
This is the number everything else depends on, and it is almost never quoted.
The reported effect sizes ran from a Cohen's d of 0.001 to 0.02. The upper end of that range means that removing a substantial share of the emotional content from someone's feed for a week shifted their own emotional word use by two hundredths of a standard deviation.
Kramer's own public defence put it in plainer units: over the week of the experiment, affected users produced on average about one fewer emotional word per thousand.
The effects were statistically significant, which is what a sample of 689,003 buys you. They were also, as critics pointed out at the time, tiny — and there are further doubts about whether they measured emotion at all, since no real-world assessment of anyone's mood was made and the text-analysis tool used was not designed for very short posts.
What happened next
The reaction was immediate and it came from inside the discipline rather than from outside it.
PNAS issued an Editorial Expression of Concern about two weeks after publication. Its editor-in-chief, Inder M. Verma, acknowledged that Facebook's collection of the data may have involved practices not fully consistent with the principles of obtaining informed consent and allowing participants to opt out.
The consent question had an awkward institutional answer. Cornell's institutional review board had determined that its researchers were not directly engaged in human subjects research, because they had no contact with the subjects and worked with data Facebook had already gathered — so no review was required. The authors' position was that the work was consistent with Facebook's Data Use Policy, which every user accepts on creating an account, and that this constituted informed consent.
That argument is the heart of the matter. It holds that agreeing to a terms-of-service document is the same thing as consenting to be an experimental subject. The Electronic Privacy Information Center filed a complaint with the Federal Trade Commission over it.
What the theory makes of it
The conspiracy reading treats the published study as the visible edge of something larger: that Facebook was establishing whether it could induce depression, despair or passivity across a population, and that the paper is the small acknowledged fragment of an unacknowledged programme.
The effect size is what that reading runs into. A manipulation that shifts emotional word use by one word in a thousand is not a demonstration that moods can be driven at will; it is close to the floor of what a sample that large can detect at all. If this experiment were the proof-of-concept for population-scale emotional control, it would be a proof that the control is extremely weak.
No evidence has been produced of a further programme. What exists is one published week-long study, its data, its authors and a journal's expression of concern.
Why it did not need to be a conspiracy to matter
The thing that makes this case useful is that the documented version is already the troubling one, and does not improve by being enlarged.
A company ran a psychological experiment on 689,003 people who had not agreed to take part in an experiment, on the theory that a terms-of-service document they had clicked through years earlier constituted consent. A university's ethics board decided it had no jurisdiction because its researchers never touched the people involved. The journal published it and then said, two weeks later, that it was concerned.
None of that required the effect to be large. The objection is to the structure — that the infrastructure people live inside can be adjusted to see what happens to them, under rules written by the people doing the adjusting — and that objection is fully supported by the paper as published. Reaching past it for a secret depression programme trades a documented problem for an undocumented one.