Part 38 · Module 5: Evaluation, Sentiment and Topics

Incident: Counting happy words is not measuring happiness

43 min read·14 Oct 2026

Module 5 · Incident file · About 30 minutes

What happened

The study ran for one week (amber), was published two years later with very small effects (blue), and the journal then raised concern about consent (red).

In June 2014 Adam Kramer, Jamie Guillory and Jeffrey Hancock published "Experimental evidence of massive-scale emotional contagion through social networks" in PNAS, the Proceedings of the National Academy of Sciences (doi.org/10.1073/pnas.1320040111).

For one week in January 2012, Facebook showed 689,003 users fewer positive or fewer negative posts in their News Feed. It judged whether a post was positive or negative by counting words from a sentiment word list, LIWC. Then it measured the sentiment of what those users posted next, the same way. The effects were very small.

The users had not been asked. The journal later published an editorial expression of concern about consent. Two questions from this module sit inside the case: does a word count measure what people feel, and is a very small difference worth acting on? A third is older than NLP: may you change what people see in order to measure them?

What happened

The rest of this part opens with an account