Meta · Emotional Contagion (2014)

In June 2014, three researchers — Adam D. I. Kramer (Facebook), Jamie E. Guillory (UC San Francisco), and Jeffrey T. Hancock (Cornell) — published in PNAS a paper reporting the results of a one-week experiment conducted on Facebook in January 2012 across 689 003 randomly selected users. Two conditions: the platform's algorithm down-weighted positively-valenced posts in one cohort's news feed, negatively-valenced in the other. The users were not informed. The users were not asked for consent. The result, statistically significant at high confidence, was that users in the positivity-suppressed condition produced posts more negative than baseline; users in the negativity-suppressed condition produced posts more positive. The effect size was modest — about 0.1 standard deviations — but operationally significant at scale: in a population of nearly 700 000, the manipulation produced affective-state changes in tens of thousands of subjects.

The experiment instantiated, openly and at scale, the operation the Desire Engine names. The platform's recommendation algorithm — the same algorithm that organized News Feed in the routine course of operation — was the instrument. The experiment introduced no new technology; it held existing parameters at experimental values for a controlled period. The Desire Engine that was deployed for the experiment had been deployed, at non-experimental parameters, on the same users for several years before and after. The experimental week was a moment when the routine was held at a chosen value in place of the optimization layer's default; it differed from the surrounding routine only in that.

The Meta experiment is what the apparatus's Desire Engine looked like in a distributed, low-resolution deployment — affect modulated at the level of a single content variable, on a one-week schedule, with the results published in a peer-reviewed venue. The chair concentrates the same operation into furniture: the modulation is continuous where the experiment's was occasional, the resolution sub-second where its was weekly, the publication a brochure where its was PNAS. The mechanism is identical. Only the discretion has changed.

The most useful experimental result, in the long run, may have been the regulatory non-response. There was no regulatory action against Facebook. There was no class action by the affected users. There was no industry-wide reckoning. The experiment's publication had no measurable effect on the subsequent deployment of similar mechanisms by Facebook, by other platforms, or by the industry as a whole. The paper became, over the next decade, a citation — never a precedent. The state had every opportunity to act; the state declined. Everything that has been built since has built on that freedom.