Influence
Part 9  The Marketplace of Attention
Chapter 261 of 360

689,003 Subjects: The Emotional Contagion Experiment

For one week in January 2012, Facebook altered the news feeds of 689,003 users.

For one group, posts containing positive emotional words were reduced in frequency. For another, posts containing negative words were reduced. The researchers then measured the emotional content of what those users subsequently posted themselves.

The effect was small and it was there. Users who saw fewer positive posts produced fewer positive words and slightly more negative ones, and vice versa. The paper — by Adam Kramer of Facebook with Jamie Guillory and Jeffrey Hancock — was published in the Proceedings of the National Academy of Sciences on 17 June 2014, titled Experimental evidence of massive-scale emotional contagion through social networks.

The reaction was not about the effect size. It was about the fact that it had been done.

None of the 689,003 people had been asked. The paper's ethics statement asserted that the research was consistent with Facebook's Data Use Policy, to which all users agree on account creation. PNAS published an editorial expression of concern noting that the collection may not have conformed to the principles of obtaining informed consent, while declining to retract.

Two things are worth separating.

The finding itself is a demonstration of emotional contagion — the process Hatfield, Cacioppo and Rapson described, in which emotional states transfer between people — operating through text at scale. The mechanism is straightforward: feed composition sets the affective baseline against which a person evaluates their own circumstances, and mood-congruent processing means an established mood shapes what is subsequently noticed, recalled and expressed.

The more important point is not scientific. It is that the study is not evidence that a platform can affect a population's mood. It is evidence that doing so is a routine product decision.

Every change to a ranking algorithm is an experiment of this kind conducted without publication. The reason this one attracted attention is that it was written up. A/B testing of feed composition, at populations in the hundreds of thousands, is the ordinary practice of the industry, and the only unusual feature of the emotional contagion study is that the outcome measured was mood rather than engagement.

Which is the observation to carry into the next six chapters. The instruments described there are not scandals. They are the normal operation of systems that optimize something, and the something is chosen by a small number of people.

The case

Adam Kramer, Jamie Guillory and Jeffrey Hancock’s study ‘Experimental evidence of massive-scale emotional contagion through social networks’, published in PNAS on 17 June 2014, in which Facebook manipulated the emotional valence of 689,003 users' news feeds; PNAS later issued an editorial expression of concern.

The mechanism

The study demonstrated that altering the ratio of positive to negative posts changed users' own posting valence — emotional contagion in the sense described by Elaine Hatfield and John Cacioppo, operating at platform scale without awareness or consent. Mechanistically, feed composition sets the affective baseline against which users evaluate their own lives, engaging mood-congruent processing (Gordon Bower). The finding matters less as a discovery than as proof that a private firm can move the mood of a population as a routine product decision.

What this chapter covers

  1. A platform experiments on its own users
  2. Feed valence resets the affective baseline
  3. One week in January 2012, published June 2014
  4. Users' expressed emotion shifts measurably
  5. Publication triggers global ethics backlash

Cross-ref: Defense — Auditing Your Information Diet