What Meta's Instagram signals record about you, and what they don't

Meta publishes examples of what it watches

None of the signals Meta lists for the Instagram feed — the pieces of information behind its guesses — asks a reader to describe themselves, and the list is public. Meta keeps a page on how the feed is ranked, updated 29 June 2026. It names what the system tries to predict about a reader, and lists the signals under each.

Here are three, word for word. One is "how much time you spend on a post". Another counts "how many times you've tapped the reshare button in the last 28 days". The last asks, of a post, "whether you actually viewed it".

Two cautions. The page says these are only some of the significant signals, a sample not an inventory. Several describe the post, not the reader.

Still, every signal about you that I could find on the page has one shape. Each is a record of an action, kept as a count, a stretch of time, or a yes-or-no.

Asking and watching are not the same thing

There are two ways to find out what a person likes. You can ask them to describe themselves, or keep a record of what they do.

Asking gets a description: someone decides what they are like, and says it. Watching gets a record instead. A record is not itself a description, though it can be turned into one. Two studies of Facebook Likes tested how well a record predicts questionnaire answers.

That difference, I think, is how to read what Meta published. Nothing Meta chose to show requires a description of the reader, from anyone. What those signals need is a record of recent behaviour, and the app already holds one.

One limit. Meta's ranking pages cover ranking, not everything Meta holds, and they are the only Meta documents I read.

Facebook Likes beat a friend's guess

In 2015, three researchers found that across their sample, guesses from a program reading only Facebook Likes tracked people's own questionnaire scores more strongly than a friend's ratings did.

They had 86,220 volunteers fill in a personality questionnaire of 100 items. It scores five broad traits, among them how outgoing someone is and how prone they are to negative emotions. Of those volunteers, 70,520 also supplied their Likes, and those pairs trained a program to guess the scores from Likes alone.

For the comparison with people, they took 17,622 volunteers each described by one Facebook friend, on a ten-item form of the same questionnaire. A smaller group had two friends describe them, used to check how far two judges agreed.

The program's guesses and the friends' ratings were scored against what each person said about themselves. The scale is correlation, from minus 1 to 1: 0 means no straight-line relationship, 1 that the two rise together perfectly, minus 1 that one rises as the other falls.

Averaged over the five traits, the headline figures are 0.56 for the program and 0.49 for the friends, a gap of 0.07 on a scale running to 1.

The paper also plots how accuracy rises as a profile carries more Likes, and estimates from it how many Likes it takes to overtake each kind of judge. In its own words, computer models "need only 100 Likes to outperform an average human judge in the present sample" — outperform meaning the guesses tracked differences between volunteers more strongly.

Against benchmarks from a summary of earlier research, the paper estimates the program needs 10 Likes to pass an average work colleague, 70 for a cohabitant or friend, 150 for a family member and 300 for a spouse. Those benchmarks came from that research, not this experiment.

The useful Likes rarely stated the trait

Kosinski, Stillwell and Graepel had asked in 2013 whether Likes alone could predict personal traits. Their dataset held 58,466 US volunteers, about 170 Likes each, though any single trait was measured on fewer volunteers. Among their targets were scores for intelligence and openness to experience, roughly how drawn someone is to new ideas.

For openness, the paper says its "prediction accuracy is close to the test-retest accuracy of a standard personality test": how consistent that questionnaire is when the same person takes it twice.

The strange part is which Likes did the work. The paper says "few users were associated with Likes explicitly revealing their attributes", meaning few had liked a page that stated the trait outright. So "predictions rely on less informative but more popular Likes": widely liked pages, each only weakly tied to the trait. Its own example: no obvious link between Curly Fries and high intelligence.

Three limits. The 2013 and 2015 studies both used volunteers who chose to install a personality app, so neither stands in for people at large. Both used Likes, deliberate acts rather than an unnoticed pause. And neither studied Instagram.

So the finding is limited to those samples, and still striking. A trained model reading only ordinary actions can estimate how someone scores on a questionnaire those actions rarely mention. Whether Instagram predicts anything like this, I do not know.

Where it does ask, it asks about a post

Instagram does ask readers things. Tapping Not Interested is an answer.

But the signal Meta lists is about the post being rated. And on the page for Explore, a Not Interested tap comes back as a signal counting "how many … posts you've clicked not interested in the recent past".

So the question is about a post, and ranking uses a count of presses. Across the pages I read, no signal asks a reader what they are like. Every reader signal on them is a record of something they did.

One thing is easy to miss. The 2015 program did not learn personality from nothing. It learned from the 2015 volunteers' own questionnaire answers, each set paired with that person's Likes. Somebody answered a hundred questions first.