A Facebook Like for Curly Fries Helped Predict a Reasoning Score

A food-page Like carried an unrelated clue

A Facebook Like for Curly Fries helped predict a score on a reasoning test.

A study asked volunteers to share old Facebook Likes, quiz scores, and profile labels. Researchers used one group to find patterns that linked the records. They then tested those patterns on people whose data had not helped build them.

The Curly Fries page did not contain a reasoning score. Its value came from the wider mix of pages liked by many people.

Enough Likes could beat a familiar human judge

A later study examined the Big Five, five broad traits about how people tend to think, feel, and act.

The researchers compared Like-based scores with each person's own quiz answers. They also compared those answers with a Facebook friend's short rating. Across the group, the Like model matched the self-reports more closely than the friend ratings did.

The model needed enough Likes to build a useful pattern. Based on an earlier review of human ratings, the authors estimated that about 70 Likes could beat a friend or housemate. In the study itself, about 100 Likes beat the average human judge.

The two Like counts came from different comparison groups. Neither number said how certain the model could be about one person's traits. The volunteers had also chosen to join a trait app, so the sample did not stand for everyone.

Indirect Likes worked because groups differed

The earlier study also used Likes to predict profile labels, including sexual orientation. Those labels came from gender and Facebook's “Interested in” field. The model tried to recover a label stored elsewhere. It did not discover a fact that nobody had stated.

Most useful Likes were indirect. The paper named Britney Spears and Desperate Housewives among pages tied to that result. Neither page named the trait. The clue came from how the wider Like patterns differed across groups.

Curly Fries did not become a clue because the food page described reasoning. It became a clue because the people who liked it also formed other patterns. The signal came from the crowd around the click.

Instagram uses behavior patterns for a different task

Meta's public “system cards” describe how Instagram ranks Feed, suggested posts, and Explore. Sample signals include viewing time, shares, hides, profile visits, past actions, and features of each post.

The systems use those signals to predict an action that may come next. Past viewing time can help predict more viewing time. A past share can help predict another share. No trait quiz is needed for those ranking tasks.

The bridge to the Facebook studies is an inference, not proof of one shared system. Both start with records of behavior and use a group pattern to answer a new question. The questions are different: the Facebook papers predicted separate scores and labels, while Instagram's cards describe predicted actions used for ranking.

The cards are only a sample of ranking inputs and outputs. Nothing cited here shows that Instagram predicts the traits in the old Facebook papers.

The useful surprise is smaller than mind reading. A simple act can gain a second meaning when it is compared with the acts of a crowd.