Experiment · Oct to Nov 2026
aas
show me your friends…
Show me your friends and I’ll tell you who you are. aas asks the same of the data a person leaves behind, what they write, where their phone goes, how they sleep, what they play and listen to, and keeps only what holds up on people the models never saw.
Where it stands
aas started in October as a question about music: why does a recommendation that lands feel like mind reading? In November it widened to the rest of a digital life and tried to read personality from it. Those readings were honest and mostly vague. A score that puts you near the middle of a scale fits almost anyone, the way a horoscope does.
So the work turned to something that can be checked. From one half of a person's listening history, the engine writes a few sentences that are true of them and rare in everyone else. Then a blind judge gets that person's other half next to four strangers' and has to pick which one the sentences describe. Guessing gets one in five. The sentences get about seven in ten.
Being recognisable is not the same as meaning something, though. The newest round turns those cold measurements into takes: a few lines about your own listening that should hit home, like the artist you only play in December or the album you wore out in a week and never touched again. Most of this round went into finding out what real data was lying about, and then into what makes a true line land.
The newest round is open below. Every earlier round is folded underneath, newest first, exactly as it was written at the time. Open any of them to see the work behind it.
Latest · Round five · 9 Nov 2026
Takes that hit home
From measurements to takes
Round four showed that three sentences can pick a person out of a lineup. But read one of them out loud, “a typical listening day runs 18 hours or more from first play to last”, and nobody smiles. It is true and rare, and it says nothing about what the music was doing in that person's life.
So the output changed. Instead of measurements, aas now gives a handful of takes, five to seven short lines of five kinds: a moment (a long silence, a week where everything was one album), a relationship with an artist (the one you come back to after every break), two sides of you (who you are at the weekend), a comfort (the song that keeps returning), and a change (the artist who was everything and then was gone).
The work is split in two so the words can't make things up. First the engine reads the whole history and finds facts, each with its numbers, its dates and how unusual it is, both against the person's own past and against 357 other listeners. Then a language model writes the takes from those facts and nothing else, and a checker throws out any take whose numbers, dates or names are not in the facts underneath it.
First contact with real histories
Until now the takes had only been tried on an invented listener, with moments planted where the engine could find them. This round ran them on 43 real people: 13 who published their full Spotify history themselves, and 30 ListenBrainz listeners drawn at random before anyone looked at them.
The first thing real data does is lie. About a third of the facts the engine chose were not about the person at all, they were about the data. A rain-sounds track left on all night read as heartbreak. A history that starts where the download starts read as “the day you discovered” an artist. A four-year hole where someone used a different app read as a four-year silence. Clocks with no time zone put evening listening after midnight, and the same artist spelled two ways became two artists, one of whom you had “abandoned”.
Each of those now has a rule: loops, sleep and background audio and sync dumps are set aside, night takes only run when the clock can be trusted, the edges of a record are treated as edges, and breaks longer than a year are taken as gaps. Each rule was written against the real case that broke it, and has a test.
| Chosen facts, 43 people | Before | After |
|---|---|---|
| about the data rather than the person | 60 of 192 | 3 of 184 |
| takes the checker had to drop | 17 | 9 |
True is not the same as landing
Cleaning the data made the takes true. It did not make them good. Forty takes, drawn at random, were read by hand and sorted into three piles: specific and true and likely to land, true but flat, and wrong. Before the fixes it was 11, 23 and 6. After, 9, 26 and 5. Taking out the false drama took some of the drama with it.
A blind judge was also given a summary of one person's listening and two sets of takes, theirs and a stranger's, and asked which belonged to them. Chance is half. With the numbers left in, it picked right 77% of the time across 108 tries. With the numbers taken out it got 58%, barely above chance. The facts carry the person. The words, mostly, do not yet.
What lands
Reading all 175 takes side by side, the good ones have something in common. They have a shape in time with edges. Something comes back on a rhythm: an artist played every December and at no other time, five years running. Something starts and stops: an album played through nineteen tracks deep over six days, then once more, ever. Something differs inside the same person: 27% of songs skipped in the morning, 46% in the evening. Or someone is who you come back to: four of your ten breaks from music ended with the same artist.
The flat ones are levels with no edges. “You have never gone eight weeks without this artist” sounds personal and is true of 87% of listeners. A song that turns up most months is true of 92%. Twenty plays of a new artist in one day is true of 94%. Those three kinds alone were 68 of the 175 takes.
So every fact now carries its shape, and the engine prefers the ones with edges. Three new finders look for them directly: albums worn out and dropped, seasons that come back, and morning against evening (only where the clock can be trusted). The season finder is the cautionary one. The first version fired for half of everyone, and it fired just as often when each person's months were shuffled at random, so it was finding chance. Tightened, it fires for 2 of the 43 and for none of 172 shuffled copies.
| Chosen facts | Before | After |
|---|---|---|
| with a shape that lands | 24% | 39% |
| steady levels | 53% | 40% |
| people with nothing to say about them | 4 | 2 |
That 39% says what the engine now prefers. It does not say the written takes land, which no judge and no hand count can settle.
What it does not know yet
Whether a take hits home is the listener's call. Every number on this page comes from strangers whose histories are public, read by a judge and by hand. The next round is people reading their own: a small invite-only app that shows each take with a plain “is this you?” underneath, and a lineup where a person picks their own takes out of three with the names blacked out. It is built and waiting for its first listeners.
A few people will get nothing. Two of the 43 have no fact strong enough to say, and the engine says so rather than padding with the flat kind.
Run 9 Nov 2026 on 13 Spotify extended histories published by their owners and 30 ListenBrainz listeners drawn with a fixed seed; population figures from 357 public ListenBrainz listeners, numbers only. Takes written by Claude Sonnet 5.5 from the facts alone; about $8 of model calls across the round. No artist, track or listener appears here; the examples are described, not quoted. Code: github.com/ubmids/aas.
Earlier rounds
Newest first. Each one is how the work stood on that day, failures included.
Round four · 6 Nov 2026
Choosing the sentences together
Picking the three sentences as a team. It missed its own 75% bar on 352 new listeners (68.8%), but beat round three’s rule on the same people by 7.2 points, the same gain it showed before.
+ open− close
Round four · 6 Nov 2026
Choosing the sentences together
Picking the three sentences as a team. It missed its own 75% bar on 352 new listeners (68.8%), but beat round three’s rule on the same people by 7.2 points, the same gain it showed before.
Three sentences that work as a team
Round three chose each of its three sentences on its own merits: rare, comfortably cleared, from a measure that holds steady. That leaves two things on the table. Sentences repeat each other. Someone whose listening days run long often also comes back for many sessions a day, so a second line about sessions mostly says the first line again. And people at the edge drift back toward the middle. A listener who is extreme on some measure in one stretch of their history is usually less extreme in the next, so a rare sentence about them is often no longer true a few months later.
Round four chooses the three together. It starts with none and adds, one at a time, the sentence that most raises the chance a blind judge picks the right person out of five. That chance is worked out from how likely each sentence is to still be true later, with the drift toward the middle built in, and from how many of the decoys it would also fit. A sentence that says what an earlier one already said adds almost nothing, so it is passed over.
It was built and tuned on the 636 listeners from round three's two sets, all of whom had now been seen. There it scored 75.2% in lineups of five, against 68.2% for round three's rule on the same lineups. The bar for a third set of people was then written down and pushed before any of them were measured: at least 75%, with the bottom of the 95% range above round three's confirmed 67.5%.
The bar failed. On 352 new listeners, none of them in either earlier set, three sentences chosen together picked the right person 68.8% of the time, with a 95% range of 65.5% to 72.1%. Both conditions miss.
The improvement itself held. Round three's rule, run on the same 352 people and the same lineups, scored 61.5%, and choosing together beat it by 7.2 points (4.2 to 10.4), the same gain of about seven points it showed on the 636. Against all 352 at once, three sentences put the right person first 13.6% of the time, where round three's three manage 8.4%.
The lesson is about the crowd. Round three's rule scored 67.5% on its own new listeners and 61.5% on these, so a whole new crowd can simply be harder, here because fewer of its people were still like themselves from one half of their history to the next. A bar pinned to one crowd's number mixes the method with the crowd. The fair comparison is head to head on the same people, and on that one the new rule won twice.
The full write-up is a paper: Who Are You?
The price of readability
A sentence on a card can say how long your listening days run. It cannot say which artists you play. That limit has a cost, and the third set measures it. Each person's first half is compared with all 352 second halves. The artist list puts their own second half on top 94.0% of the time. All 57 printable measures at once do it 42.3% of the time, and the three sentences a card prints do it 13.6% of the time, against a chance of 0.3%.
How it fades
The same matching can be run with a gap between the two stretches of history. On the 261 listeners with enough history for a year's gap, the same people at every gap, two 26 week windows were matched with nothing, one, three, six and twelve months between them. The sentences are scored in lineups of five; the 57 measures and the artist list are scored on finding one person among all 261.
| Gap | 3 sentences | 57 measures | Artists |
|---|---|---|---|
| none | 76.7% | 64.0% | 96.9% |
| 1 month | 74.5% | 59.4% | 97.7% |
| 3 months | 71.2% | 47.9% | 96.9% |
| 6 months | 71.0% | 42.9% | 94.6% |
| 12 months | 61.7% | 29.5% | 90.4% |
Everything fades. Over a year the 57 measures fall from 64.0% to 29.5%, more than half gone, and three sentences in a lineup fall from 76.7% to 61.7%. The artist list goes from 96.9% to 90.4%. How someone listens drifts much faster than what they listen to.
Numbers from the aas engine's identity write-up, 6 Nov 2026, run once on the third set after its rule and bar were committed and pushed. The 57 measures here use a distance learned from how each person's measures move between quarters of their history.
Round three · 5 Nov 2026
The part that works
Instead of placing people on a scale, three sentences true of them and rare in others. They picked the right listener out of five 67.5% of the time on 333 new people.
+ open− close
Round three · 5 Nov 2026
The part that works
Instead of placing people on a scale, three sentences true of them and rare in others. They picked the right listener out of five 67.5% of the time on 333 new people.
Why trait scores read like horoscopes
A personality score places you on a scale shared with everybody else. When the reading is weak, most people land near the middle, and the sentence that comes back (“you are about as open to new experience as most people”) fits nearly anyone who reads it. That is how a horoscope works. It feels personal because it is true, and it is true because it is true of almost everyone. Round two's trait scores were honest, and for most people they were exactly that kind of sentence.
Round three turns the question around. Instead of placing you on a scale, it looks for things that are true of you and false of nearly everyone else, written so you can check them against your own memory. Each comes with a count of how many other listeners it also fits, so you can see how rare it is.
Statements true of you and rare in others
The engine measures 57 things about a listening history: how much of it goes to artists you met in the last month, how long a listening day runs, how often you come back for another session, how much of a season is the same songs again, and so on. None of them names an artist or a song, and none touches health, religion, sexuality or politics. For each measure it has a few round thresholds and a sentence for each, such as “you play at least 2,000 different artists in a typical season.” A statement qualifies for you when it is true of you and of only a small share of everyone else. The ones shown are those you clear most comfortably, from measures that hold steady over time, and that the fewest others share.
These are from three of the 333 new listeners, drawn at random before anyone looked at their cards. They are written as the engine prints them.
Your listening days are long: a typical day runs 18 hours or more from first play to last. Yours: 19.5 hours. True of 5 of the 332 other listeners (1.5%).
You meet at least 400 artists you have never played before every month. Yours: 608 new artists a month. True of 3 of the 332 other listeners (0.9%).
No more than 10% of your plays go to artists that were in your top 50 four to six months before. Yours: 6.1%. True of 6 of the 332 other listeners (1.8%).
On a listening day you come back for 5 or more separate sessions. Yours: 7.0 sessions a day. True of 16 of the 332 other listeners (4.8%).
The second and third are the same listener. Listeners are kept under salted hashes and no artist, track or username appears on a card. These four happen to be rarer than most: across all the cards, about two lines in five fit a tenth of other listeners or fewer, and the rest fit between a tenth and a fifth.
The lineup test
A rare sentence is only worth something if it really belongs to you. The test for that is a lineup. Take one person's listening and cut it in half by time. Make three statements from the first half only. Then put that person's second half next to the second halves of four strangers, and ask a blind judge which of the five the statements describe. The judge simply counts how many of the three statements are true of each and picks the highest. Guessing gets it right one time in five.
The rule was written down and committed before the test was run: at least 60% right, with the bottom of its 95% range well above chance and a shuffle test that chance could not match. Round one, on 304 ListenBrainz listeners with statements true of a tenth of others or fewer, got 54.9%. Far above chance, and short of the bar, so it was reported as a miss.
The miss had a plain cause. Too many people had fewer than three statements to give, and a person with nothing to say scores chance. So round two added 23 measures, 57 in all, let a statement fit up to a fifth of other listeners instead of a tenth, and preferred statements the person clears comfortably from measures that hold steady over time. On round one's listeners that reached 69.7%.
That number cannot be trusted on its own, because every change was chosen by looking at those same people, and a rule tuned on a group always flatters itself on that group. So the new rule was frozen and committed while 333 new listeners were still downloading, people nobody had looked at and none of them in round one. The script that scores them refuses to run a second time.
On the new listeners, three statements picked the right person 67.5% of the time, with a 95% range of 63.8% to 71.0%. It clears the bar, and it lost about two points moving from the people it was tuned on to people it had never seen. Matching on all 57 measures at once, instead of three sentences, finds the right person 89.5% of the time (87.1% to 91.7%), so the two halves of a listening history are very alike and three sentences carry about two thirds of that likeness. Generic statements, true of the person and of most others, get 31.7%: a few horoscope lines narrow a lineup a little, and only a little.
A game you can check
Statements describe the past. The other half of this stage is predictions a person can check for themselves in the weeks that follow. Three games were set in advance, each scored person by person on held-out weeks against a strong simple rule, and each needed to win for at least 70% of people.
The one that passes is the simplest to say. You have just heard three new artists, one session each. Which one will you keep? After that first session the engine names one. On the 9 Spotify listeners with enough new artists it was right 69% of the time against 37% for picking the most popular, and it beat popularity for all 9. On 149 ListenBrainz listeners it was right 70% of the time against 46%, and it beat popularity for 132 of them. Chance is one in three.
There is a catch, found after the run. A much plainer rule does nearly as well: pick the one you played most in its first session. On Spotify that rule also gets 69%, the same as the engine. On ListenBrainz it gets 63%, and the engine still beats it for 104 of the 149. So the prediction holds, and most of what it knows is how much you played each artist on day one.
The other two games missed. Calling which of your next 10 tracks you will skip gets 7.9 right against 7.5 for the better simple rule, and it wins for 9 of 12 Spotify people where the bar needed 11. Naming next week's top five artists barely beats your last four weeks: it wins for 56% of ListenBrainz listeners and 5 of 12 Spotify people.
What to keep in mind
- The pass came partly from loosening what rare means, from a tenth of listeners to a fifth. That change was made before the new listeners were seen, and every card prints how many others each line fits.
- The new listeners were found the same way as the old ones, by following ListenBrainz's similar-users links. They are new people from the same kind of population, mostly keen listeners who log everything. Other kinds of listener are untested.
- Some accounts are odd. Among the rarest cards are one with music on 1.4% of days over 14 years and one where 99% of plays fall in the same four hours. They passed every cleaning rule and may be logging habits more than listening habits.
- No card has been shown to the person it describes. Whether a true, rare statement feels like it describes you is still untested.
Numbers from the aas engine's statements write-up and its prediction game write-up, 5 Nov 2026, each with its rule committed before the run it scores.
Round two · 5 Nov 2026
v1, reading one person
One person’s own files in, a personality reading out, with how sure it is. Honest, and too vague to be useful: about r 0.16.
+ open− close
Round two · 5 Nov 2026
v1, reading one person
One person’s own files in, a personality reading out, with how sure it is. Honest, and too vague to be useful: about r 0.16.
From research data to your own files
Round one asked whether personality shows in each kind of data, across thousands of strangers. Round two asks the question a person would ask: given my own data, what can you say about me, and how sure are you? v1 is the engine as it stood on 5 November 2026. It takes the files Reddit, X, Google and Spotify hand over when you ask for a copy of your data, reads them on your own machine, and returns each answer with a range and a stated accuracy. Where the data can't support a number, it says so and gives the reason.
There are three readers. The first reads the Big Five from writing. It is a bag of words, a model that looks only at which words someone uses and how often, trained on 2,467 student essays whose writers also took the questionnaire. The second reads the four Myers-Briggs letters and is trained on the 8,675 forum users from round one. The third measures how someone listens to music, using the dials from the listening work further down, and makes no personality claim, since no link from listening to the Big Five has been tested. Searches are left out, because a two-to-four-word query looks nothing like the sentences the models learned from.
Accuracy depends on the kind of writing
A model trained on essays has learned how students write when asked to write freely for twenty minutes. Tweets, Reddit comments and talking to a camera are different. So the readers were tested on whole datasets of writing they never saw, and each one states the accuracy measured on the kind of writing it is given.
The drop is steep. On student essays it never saw, the Big Five reader reaches AUC 0.68 for openness and 0.57 to 0.59 for the other four. On YouTube video transcripts, where strangers rated the speakers, it averages 0.59. On tweets from 294 people who rated themselves (PAN 2015) it averages 0.54, and only neuroticism clears chance on its own, at 0.57. On lines from a sitcom it is at chance. Reddit has not been measured at all, so it borrows the tweet numbers.
Openness, the strongest trait on essays, does not travel. On every other kind of writing its range includes a coin flip. So if less than half of a person's words are essay-like, the reader gives no openness number and says why. In practice that refuses openness for anyone who writes mostly on Reddit or X.
The letter reader has a harder problem. No Myers-Briggs-labelled writing from outside the forum is openly available, so its accuracy off the forum has to be inferred, from how its letters line up with the Big Five on the student essays. That gives AUC 0.61 to 0.62 per letter, keeping between a third and a little over half of its strength on the forum. Spot checks on tweets came out lower still, 0.48 to 0.58.
Every Big Five answer comes with an 80% range, a claim that the person's own questionnaire score falls inside it four times in five. When accuracy is close to chance, a range that keeps that claim is nearly as wide as the one that fits anybody, and the answer sits close to average. For most people's writing, that is what v1 gives today.
The AI reader
The obvious next question is whether a large language model reads people better. It was never trained on any of these datasets, so it has no student-essay habits to lose. Claude Haiku 4.5 was given up to about 1,500 tokens of each person's text, with type and trait words stripped out, and asked to rate them as their questionnaire would. Nothing was trained. The test was written down before any rating came back: 150 people from each of three datasets, the same text for every reader, and a pass meant beating the bag of words by at least 0.03, with the 95% range of that lead clear of zero.
It passed. On the two kinds of writing the trained model never saw, tweets and video transcripts, Haiku averaged AUC 0.618 against 0.558, a lead of 0.060 with a 95% range of 0.025 to 0.093. It led on both: 0.679 against 0.607 on the videos, and 0.557 against 0.508 on tweets, where both were weak. A small model run locally, Qwen3 at 1.7 billion parameters, answered 3.0 on every trait in 15 of 16 valid replies before the run stopped on a disk limit, so it gave no verdict.
Haiku can't ship. It runs on Anthropic's servers, and aas promises that a person's text never leaves their machine. So it serves as a benchmark, the mark a reader on the laptop should aim for. There is also a caveat. The three datasets were public long before Haiku was trained, so it may have seen them. None of its 450 written reasons names a dataset or a score, which is suggestive and proves nothing.
Teaching a small model on the laptop
Distillation is a way to copy a big model's judgement into a small one. The big model, the teacher, rates a large pile of other people's writing once. A small model, the student, is then trained to give the same ratings from the same text. Only the student ships, and it runs on the person's own computer, so their writing goes nowhere. The teacher is paid for once, during training, and never sees a user.
Here the teacher was Haiku with the same prompt as before, and the pile was 2,000 forum users from round one, their types thrown away unread. Nobody from the test sets was in it. Haiku rated all 2,000 for $3.35. The student is a pair of simple linear models, one over word counts and one over sentence meanings, averaged together. It is 0.6 MB on disk and matches about 40% of the variation in Haiku's ratings of forum users it had not seen.
The pass mark was fixed before any rating came back. On the same 150 people per set, the student had to keep at least half of Haiku's lead over the bag of words: 0.588 on tweets and videos, and 0.596 over all three sets with the essays included. It scored 0.586 and 0.583. It failed both, the first by 0.002. On the essays it lost to a bag of words that had been trained on essays.
A wider look is kinder. On every test person, 294 tweeters and 404 video speakers rather than 150 of each, the student beat the bag of words by 0.030, with a 95% range of 0.010 to 0.051, clear of zero. Almost all of that came from the videos; on tweets the two were level. The rule was written for the 150-person comparison, so this does not change the verdict.
Why not adjust the student and run the test again? Every rerun against the same test sets is another look at them, and enough looks will find a pass by luck. The student's numbers stay as the facts of this one test. It is saved, labelled as failed, and not used by default. The next attempt needs a new pool with more kinds of writing than one typology forum, judged by the same frozen rule. It was also trained on Claude's outputs, and Anthropic's terms on that have to be checked before any commercial use.
Putting it together
Someone who brings Reddit comments, tweets and a blog gets readings from both text readers, and some of them speak to the same trait. The I or E letter bears on extraversion, for example. Fusion combines them into one number per trait. Normally the weights would be learned from people who have every source and a questionnaire, and no such set exists yet. So v1 uses a Bayesian model that needs no joint training data. Each source is weighted by its own tested accuracy, discounted for new people, and two readers of the same text are assumed to share some of their mistakes.
With no real people to check it on, it was checked in simulation: 20 rounds of 3,000 invented people, using the accuracies the readers state. The fused r comes to about 0.16 on average: 0.196 for openness, 0.146 for conscientiousness, 0.172 for extraversion, 0.121 for agreeableness and 0.148 for neuroticism. The 80% ranges hold 80% of the time on every trait, so they mean what they say.
In short, combining barely beats the best single source today. Taking each trait's best source on its own gives 0.153. A second source adds a little to conscientiousness and agreeableness and nothing measurable to extraversion. Openness rests entirely on the inferred N or S letter, and neuroticism entirely on the Big Five reader. The ranges are 2.54 wide on the z scale, against 2.56 for the range that fits anyone, so the fused number moves a person only a little from average. If the failed student replaced the bag of words for non-essay writing, the same simulation would average 0.184, with extraversion and neuroticism up and conscientiousness down. That swap has not been made.
What it returns for one person
To show the output, one person was invented in full. Every post, comment, search, artist and play was written or generated for the example, and no real person's data went in. They are a hospital pharmacist who works shifts, rides a bike, bakes bread and fixes old bicycles. They bring 30 Reddit comments, 22 posts on X, one blog post and five searches, 1,598 words without the searches, and a year of listening, 9,425 plays.
| Trait | Big Five reader | Stated AUC | Fused | Fused 80% range |
|---|---|---|---|---|
| openness | refused | −0.279 | −1.58 to +1.02 | |
| conscientiousness | +0.114 | 0.55 | +0.036 | −1.23 to +1.30 |
| extraversion | −0.223 | 0.55 | −0.281 | −1.56 to +1.00 |
| agreeableness | −0.121 | 0.55 | −0.116 | −1.38 to +1.15 |
| neuroticism | −0.074 | 0.57 | −0.056 | −1.32 to +1.21 |
Values are z scores against the questionnaire's norms, so 0 is average and about plus or minus 1 covers two people in three. Every value sits near 0 and every range spans most of the scale. For someone who writes mostly comments and tweets that is the right answer: the readers state an accuracy close to chance for this mix of writing, and the output follows. Each line also explains itself in words. Three of them, as the engine prints them:
openness: not shown to work on this kind of writing, so no number is given. … it has only been shown to work on essay-like long prose, which is 22% of your words (at least 50% needed).
Chance you are I rather than E: 66% (80% range 64% to 68% if your texts were resampled). … Untested off the forum … the chance above is pulled toward 50% to match.
Explore against stay: -2.36 (0 is the typical listener, about +/-1 covers two in three people), so you lean staying with old favourites.
The person and everything they wrote, searched and played are invented. Excerpts are cut from the engine's full output for this person, with … marking the cuts.
The listening gives the clearest picture: 83% of their plays went to artists already in their top 50 four to six months before, more than 99% of the 304 reference listeners. Given only their first five posts on X, 91 words, and two weeks of listening, the letter reader and the music reader decline and say why, and the Big Five ranges grow to the width that fits anybody.
What v1 is, and what it is not
v1 is an engine that reads a person's own exports on their own machine, gives each answer a range and the accuracy measured for that kind of data, and declines where it has nothing to go on. Every accuracy a reader states comes from a test on people its model never saw, and the fused reader marks its own as untested.
It cannot yet stand in for a personality test. No reader has been checked against ordinary people's own questionnaires; every accuracy here comes from public research datasets. On the writing most people produce, comments and tweets, it is close to chance, so for most people it says little beyond average. Fusion has never met a real person with several sources. Phones, watches and games have no reader, because nothing there survived at 46 to 149 people. Music is described and never turned into personality.
What would move it
- PANDORA, about 1,600 Reddit authors with Big Five scores. The loader and the retraining path exist and the access request has not been made. It would give the first measured number for Reddit. The bar for the main study is r 0.20 on Reddit authors the model never saw, for at least three traits, with tweets not below about 0.12.
- A varied pool for the student. More kinds of writing than one typology forum, rated by the teacher the same way for a few dollars, and tested under the same frozen rule.
- A real study. People who bring their own exports and take the questionnaire, so the stated accuracies, the ranges and the fusion are checked on people instead of in simulation. Its hypotheses are registered, and its power analysis recommends waiting for a better text model first.
v1 is honest about how little it knows, and that honesty has a cost. Trait scores this weak read like a horoscope when they are handed to one person, because a sentence that fits most people tells nobody anything about themselves. So the next stage turned to statements that are specific to one person, true of them and false of nearly everyone else, and to predictions a person can check on their own. That stage has now started, and round three below is its first result: a blind judge picks the right person from three statements two times in three, and one of three prediction games passes. What comes after it is a test with real people. Shown their own reading next to a few belonging to strangers, can people pick out which one is theirs?
source →Numbers from the aas engine as tagged v1.0 on 5 Nov 2026, from its maths write-up and the test documents it cites. Essays from Pennebaker and King (1999); tweets from the PAN 2015 author profiling task; video transcripts from Biel and Gatica-Perez (2013).
Round one · 1 Nov 2026
Personality from a digital world
Writing, phones, watches, games and how people answer a test, read one at a time. Writing reads thinking against feeling at AUC 0.83; the rest says little.
+ open− close
Round one · 1 Nov 2026
Personality from a digital world
Writing, phones, watches, games and how people answer a test, read one at a time. Writing reads thinking against feeling at AUC 0.83; the rest says little.
The idea
There is an old line that goes “show me your friends and I’ll tell you who you are.” The people around you say something about you. These days most of a life also leaves a trace: the things you write, the places your phone goes, when you fall asleep, what you play, what you play on repeat. aas asks how much of a person can be read from those traces, and how sure anyone can be about it.
Psychologists have been asking a version of this for over a decade. In 2013 Kosinski, Stillwell and Graepel showed that the pages people liked on Facebook predicted their personality, and in 2015 Youyou, Kosinski and Stillwell found that with about 300 likes a model judged someone's personality better than their spouse did. Those studies used one platform and a lot of people. aas wants to see what each part of a digital life adds, and to keep the claims small enough to check.
It started in October as a narrower question about music: why does a recommendation that lands feel like mind reading? That work is further down the page. This round widens it from one kind of data to several, and from taste to personality.
What “who you are” means here
A claim like “you are introverted” is only worth making if it can be wrong. So aas uses a measure of personality that people fill in themselves, and treats their own answers as the thing to predict. The one psychologists trust most is the Big Five, which places everyone somewhere along five scales:
- openness, how drawn you are to new ideas, art and experiences
- conscientiousness, how organised and dependable you are
- extraversion, how much you seek out people and stimulation
- agreeableness, how warm and cooperative you are with others
- neuroticism, how strongly you feel stress, worry and low moods
Nobody is one or the other on any of these. They are positions on a line, and most people sit near the middle. Some datasets only have Myers-Briggs, the four-letter types like INTJ, which are less reliable but split into four yes-or-no questions that can still be tested.
Each dataset in this round comes with people who took one of these tests and agreed to share their data for research. The model sees their data, guesses where they sit, and is marked against what they actually said about themselves.
How to read the numbers
The most important rule is that a model is only ever marked on people it has never seen. Anything can memorise the people it was trained on. The work splits the people into five groups, learns from four, predicts the fifth, and rotates until everyone has been predicted by a model that never met them.
Two scores appear on this page. When the answer is a position on a scale, the score is a correlation, written r. At 0 the guesses have nothing to do with the truth, at 1 they line up perfectly. For reading personality from digital traces, published work usually lands around 0.3 to 0.4. For scale, in the 2015 study a work colleague judged someone at about 0.27 and a friend at about 0.45. When the answer is yes or no, the score is AUC. Take one real introvert and one real extravert at random, and AUC is how often the model ranks the introvert as more introverted. 0.5 is a coin flip and 1 is never wrong.
Small studies throw up good-looking numbers by luck, so every score also goes through a luck check. The personality answers are shuffled so they belong to the wrong people, the whole thing is rerun 200 times, and a result only counts if real answers beat shuffled ones clearly, with the bar raised for every extra trait tested at the same time. Most of what follows does not pass. Those results stay on the page anyway, since a method that never says no can't be trusted when it says yes.
Round one in three numbers
AUC for thinking against feeling, read from 8,675 people’s forum posts with every type word taken out.
of AUC that type words were worth on judging against perceiving. People on a type forum write their own type.
AUC for whether a new artist will stick, read after one session, from the listening work that came first.
Writing turned out to be where personality shows. Along the way a leak in the data was caught that would have flattered the model by up to 0.14. The sections below go through each part of a life in turn.
The data
No public dataset holds one person's whole digital world next to their test. Each covers one part of a life, so round one reads them one at a time.
| Part of a life | Dataset | People | Test |
|---|---|---|---|
| what you write | PersonalityCafe forum posts | 8,675 | Myers-Briggs |
| where your phone goes | StudentLife (Dartmouth) | 46 | Big Five |
| what your watch sees | LifeSnaps (Fitbit) | 52 | Big Five |
| what you play | Steam profiles (Lin and Li) | 149 | Big Five |
| how you answer | Open Psychometrics IPIP | 597,365 | Big Five |
How it was scored, in detail
- Every dataset is turned into the same two tables, one row per person: their test scores, and numbers describing their behaviour, each computed from their own data alone.
- Five-fold cross-validation, repeated five times, so every prediction comes from a model that never saw that person.
- The labels are shuffled and the whole thing refit 200 times, to see how often noise does as well. Holm's adjustment across the traits tested together.
- With no signal, cross-validation scores below zero (around −0.1 to −0.2 at these sizes), so a score is read against what shuffled labels get.
- Features are few and chosen before any result, about 15 per sensor dataset, since 50 people and 200 features can explain anything.
Writing
PersonalityCafe is a forum about Myers-Briggs types, and 90% of its users write their own type somewhere in their last 50 posts. A model can read “as an INTJ” and look brilliant. Counting type names, and nothing else, reaches AUC 0.87 to 0.90. So every type name, partial type, function code and link comes out before anything is scored, and the scores are kept for both versions.
Once the words are gone, writing still says something. Thinking against feeling reads at 0.83, introvert against extravert and intuition against sensing at 0.70 to 0.73, judging against perceiving at 0.64. Plain word counts with logistic regression do better than a small sentence-embedding model (0.88 for thinking and feeling), so most of the signal is in what people talk about. Of the readable measures, thinkers write fewer “I” words, fewer joy words and exclamation marks, longer words, and swear more.
People on a type forum may also write to their type, so the labels are weaker than a proper test. A Reddit set with Big Five scores (PANDORA, 1,600 people) is the next text dataset, and its loader is written.
Phones, watches and games
From the phones: how far someone roams, how much of the day they spend in their own dorm, minutes of conversation the microphone picked up, Bluetooth devices nearby, how regular their days are, when they sleep and how much that moves. From the watches: steps, active minutes, fitness, sleep, resting heart rate, and whether they are a morning or evening person. From Steam: level, badges, games, screenshots, reviews, groups and friends, logged. Self-reported mood was left out of the watch data because it is close to a neuroticism question.
Extraversion came out near 0.24 in both sensor sets, from different devices and different features, a little under the 0.37 Stachl and colleagues found on 624 phones in 2020. With about 50 people the interval runs from 0 to 0.5, so neither set can confirm it or rule it out. One single link did survive its adjustment: students with irregular bedtimes scored higher on openness (rho 0.59, 44 students). A later test that pooled both sensor sets with a matching measure of rest timing did not repeat it (0.16 on the watches, the only independent check), so it stays unconfirmed. Steam counts carried nothing, and gender and age alone did better than the whole profile.
How fast someone answers a personality test, without what they answered, predicts all five traits on 20,000 people. Faster answerers score higher on openness and neuroticism. It is under 2% of the variance, the kind of result a large sample makes significant and nothing else makes useful.
The ruler
The test is what every prediction is marked against, so it gets checked too, on the answers of 597,365 people after duplicates, skipped questions, rushed answers and straight lines are removed. All five scales reach Cronbach's alpha 0.80 to 0.90. A factor analysis that is never told which question belongs where finds five factors, and all 50 questions land on their own trait. A 20-question version, chosen on half the people and tested on the other half, agrees with the full 50 at 0.91 to 0.95, which caps how well anything can score against it.
A pre-registered rerun then pooled the two sensor sets, 92 people, and extraversion failed its own rule there too. The closest single link, people whose activity is centred later in the day being more extraverted, reached rho 0.30 and missed the bar (adjusted p 0.053). So round two went where the signal was, to writing.
Before this · Oct 2026
Listening histories
Where it started. Whether a new artist will stick can be read after one session (AUC 0.792); forecasting new favourites mostly cannot.
+ open− close
Before this · Oct 2026
Listening histories
Where it started. Whether a new artist will stick can be read after one session (AUC 0.792); forecasting new favourites mostly cannot.
Why it started with music
A recommendation that lands feels like it knows you. aas started as a question about how much a listening history actually says, and which of it can be checked. Every claim on this page is a prediction scored against what the person went on to do.
The histories
Two sets. Thirteen Spotify extended streaming histories that their owners published on GitHub, which carry skips, how each track started, shuffle and device. And 343 public ListenBrainz histories, found by following its similar-users links out from ten long-standing accounts, of which 304 have enough listening around the cut to be scored: a median of 4.1 years and 59,000 listens each. They lean towards self-hosters and Last.fm veterans, so they are not typical listeners. Nothing is redistributed and nobody is named; ListenBrainz users are kept under a salted hash.
How it was tested
- Each person's last 8 weeks are held out, and anything tuned is tuned on the 8 weeks before.
- A model that learns across people never sees the person it scores.
- 95% intervals by resampling people, a sign test per comparison, Holm's adjustment across each round.
- It survives only if the adjusted p is under 0.05, more people win than lose, and the interval sits above zero.
The forecast that does not work
Ask which unheard artist someone will come to love in the next 8 weeks, where a new love means 10 or more full plays on at least 3 days. A two-stage ranker proposes artists from recent favourites' ListenBrainz similar-artist lists and the taste map, then scores them with what the lists say, popularity and how this person usually finds music. With 13 people it never beat popularity in a way that held. With 293 it does: 4.7% of new loves land in its top 50, against 3.1% for the lists alone and 2.4% for popularity, and stage one reaches 18%. It beats popularity and is still a guess.
Most new loves are on no recent favourite's list at the moment the window opens, so most people have none in any top 50. New loves are better read looking back than forward.
The forecast that does
Narrow the question and ask it later. Someone has just heard an artist for the first time. Will it stick?
At the end of that first session a logistic regression reads what happened in it (plays, different tracks, whether a track was played again, whether the listener came back to the artist after someone else, how long the session ran), the similar-artist lists, the taste map, popularity, and how often this listener's new artists usually stick. On the 304 ListenBrainz listeners it reaches 0.792 [0.784, 0.800] against 0.618 for popularity, better for 283 of 303 people, and 0.811 on the held-out weeks for the 112 people with enough new artists there. Only 5.2% of new artists stick. None of the 15,382 rows changed when re-read at the cut.
What carries it is the session; the first session alone reaches 0.736. On the Spotify histories, taking the first session out of the model costs 0.051 of AUC; taking out the lists costs 0.009 and how the track started 0.006, and the map, popularity and time of day change nothing measurable. Skips, play lengths and start reasons, the columns ListenBrainz lacks, add 0.014 after the first session (not after adjustment) and nothing after 7 days. Where an artist came from adds a little on top of how the first listen went.
Three from the held-out weeks (Spotify, full model)
- Subsonica: came on after Fugazi, 4 plays all past 30 seconds, a favourite lists it. 66%. Stuck by day 2.
- Cutting Crew: picked, played 10 times in the session. 95%. Did not stick.
- Penguin Piano: came on by itself, 2 plays. 2%. Stuck by day 13.
The door
Looking back works where looking forward fails. For each new love, aas reads the facts at its first play and compares them with artists of similar standing that were not loved. A recent favourite already listed it on ListenBrainz 5.3x as often as never-played artists named on as many lists [4.9, 5.6], higher for all 302 people. Against artists the same person tried once and dropped it is 1.87x.
There is a leak to check. ListenBrainz builds those lists from its own users' sessions, and these people are its users. If their sessions drove the door, the heaviest listeners would show it most. They show a little more: 5.6x for the heaviest third against 4.5x for the lightest, which still holds for 101 of 101. The door is real; part of its size may not be.
On the Spotify histories, which say how a track started, “came on after an artist that lists it” is 2.40x as common for loves as for dropped artists. Without start reasons it reads 1.42x, under the 1.5 bar a door has to clear.
Two dials
Six traits can be read without skips: the share of plays going to artists new this month, how spread out someone's artists are on the map, how much they return to old favourites, how fast their top 10 turns over, and two measures of when they listen. Each is measured on the two halves of a person's history; across 304 people the halves agree at r 0.72 to 0.88, every interval above 0.5.
A factor analysis keeps two factors: eigenvalues 3.10 and 1.86 against 1.27 and 1.15 for shuffled data, together explaining 83%. The first dial is explore against stay, and it holds: new share, spread and turnover load at −0.79 to −0.90, old friends at +0.94, and none of them touch the second (|r| ≤ 0.12). The second is how regular your listening time is. The hour itself needs a time zone, which this data lacks, and the dial rests on two measures that correlate at −0.86, so it is close to one quantity measured twice. What it does show is that how fixed your hours are has nothing to do with how much you explore.
What held and what did not
| Question | Model | Baseline | People | Verdict |
|---|---|---|---|---|
| will it stick, after the first session | 0.792 [0.784, 0.800] | base rate 0.466 | 301/2 | survives |
| the same, against popularity | 0.792 | popularity 0.618 | 283/20 | survives |
| the same, held-out weeks | 0.811 [0.783, 0.838] | popularity 0.607 | 104/8 | survives |
| will it stick, after 7 days | 0.884 [0.879, 0.890] | first-week plays 0.860 | 257/46 | survives |
| the same, held-out weeks | 0.901 | first-week plays 0.878 | 41/30 | weak |
| next 8 weeks of artists, recent over all-time | 0.706 [0.677, 0.736] | all-time 0.604 | 218/72 | survives |
| new loves in the top 50, ranker | 4.7% pooled | popularity 2.4% | 132/24 | survives |
| the same, held-out weeks | 3.6% pooled | popularity 1.5% | 22/6 | weak |
| eras, era-aware forecast | 0.706 | decay 0.706 | 67/76 | no difference |
| eras, current era only | 0.688 | decay 0.706 | 107/192 | worse |
Eras read well as seasons of a life, but an era-aware forecast ties plain decay, and using only the current era is worse: a story, not a forecast. The listening clock beats a flat week by 0.456 bits per play, read from UTC times, so it measures how predictable someone's timing is; whether the hour itself predicts is untested here.
Some things need what only the Spotify histories carry, and rest on 13 people: skip prediction (AUC 0.848 against 0.617, all 13), the selves a person listens as, and the game of picking a real track out of decoys.
The listening work guesses no personality. Its traits are measured, each with the plays behind it, and no sensitive trait is modelled anywhere in aas.
Listening histories and similar-artist lists from ListenBrainz's public API; Spotify extended histories as published by their owners. The figure uses the fitted weights of the ListenBrainz model, exported from the code that produced every number on this page.