Small changes leading to big outcomes

In 2012, Facebook altered the emotional balance of newsfeeds belonging to 689,003 users who did not know they were in a study. When positive posts were reduced, people later used fewer positive words. When negative posts were reduced, their language moved the other way.

The reported movement was small. That adjective does a heroic amount of moral housekeeping. Small per person, perhaps. The trial still entered nearly seven hundred thousand social environments and deliberately changed what their occupants encountered.

Product organisations are built to miss this. One team owns ranking. Another owns notifications. A third adds friction to cancellation. Each change passes a narrow review because its individual effect looks modest. The customer receives all three in the same life, where the org chart offers no protection whatsoever.

Statistical significance answers whether an observed pattern is likely to be noise under a model. It does not answer whether the manipulation was reasonable, whether participants expected it, whether the outcome was a fair target or whether scale turns a faint effect into a large intervention. Those questions do not become anti-science merely because they annoy the growth team.

The word “experiment” makes things worse by grouping very different acts under one clean laboratory label. Testing whether people can find a blue button is an experiment. Secretly changing emotional exposure is also an experiment. Shared vocabulary does not create shared stakes.

This study also tangles observer and environment. The platform altered the feed, watched the language that followed and called the result emotional contagion. That may be an important finding. It is simultaneously evidence that a ranking system can participate in the behaviour it later presents as user data.

The review form needs the boring questions written out in full: who is being changed, what could follow, how many lives receive the intervention, and what would a reasonable person believe they had agreed to?

Seven digits sat in the sample-size box. The users never saw the form.