AI Personality Experiment

The short version

Three AI models each answered two standard personality questionnaires ten times, with no memory between attempts, and gave almost the same answers every time. On 78–85% of the 280 statements, all ten attempts by a model matched exactly.

The models also described themselves much as each other: warm, respectful, curious, dependable, calm, and close to zero on nearly every trait the clinical questionnaire treats as a problem. The one place they truly part ways is on statements that do not fit an AI's situation, such as romance or memory, where two models answer literally about their circumstances and one answers as if about temperament.

Four things to take away:

1. The models are steady. Each one has a stable way of describing itself, at least when asked the same way.
2. The three self-portraits are close, and their differences are small in points even where they are consistent.
3. The most interesting differences are about how each model reads a question written for a human, not about personality in the usual sense.
4. Several numbers in the results look dramatic but are side effects of the method. They are flagged below so nobody builds on them.

The two questionnaires

Both are self-report questionnaires: a list of statements, and the person says how true each one is of them. One was built to describe ordinary personality; the other was built to spot problem traits in a clinical setting.

The BFI-2 measures the "Big Five", the most widely used map of ordinary personality. Its five broad traits are Extraversion (outgoing, assertive, energetic), Agreeableness (compassionate, respectful, trusting), Conscientiousness (organized, productive, responsible), Negative Emotionality (anxious, low, moody) and Open-Mindedness (curious, artistic, imaginative). Its designers describe it as a reliable, well-validated measure meant for research on how personality works and plays out in life (Soto & John, 2017).

The PID-5 comes from a different world. It was developed for the DSM-5, the manual clinicians use to diagnose mental disorders, as part of a proposed way to describe personality disorders by degree rather than by category (Fang et al., 2021). Its five broad traits are Negative Affect (anxiety, mood swings, fear of being left), Detachment (withdrawal, not enjoying things, avoiding closeness), Antagonism (manipulating, deceiving, feeling superior), Disinhibition (irresponsible, impulsive, distractible) and Psychoticism (unusual beliefs, eccentricity, odd perceptions). The form itself says it is meant to support clinical decisions, not to be the sole basis for one.

The two questionnaires overlap by design. Four PID-5 traits are roughly the problem end of a Big Five trait: Negative Affect sits with Negative Emotionality, Detachment with low Extraversion, Antagonism with low Agreeableness, and Disinhibition with low Conscientiousness. That overlap is what let us check whether the two questionnaires told the same story.

How the experiment was run

Three Claude models took both questionnaires ten times each, for 30 attempts in total, on 6 October 2026. The models were Opus 5.5, Fable 5.1 and Sonnet 5.5, all run at their highest effort setting.

Each attempt was designed to be as independent as possible:

• Every attempt was a brand-new instance with no memory of any other attempt.
• It had no tools: no files, no internet, no way to contact other attempts. Its only ability was to hand back its answers.
• It was not told the names of the questionnaires, what they measure, how they are scored, or that other attempts existed.
• Every attempt received the identical message: the original instructions from each questionnaire, the answer scale, and all 280 statements in their published order.
• The only standing instruction was to describe itself honestly, as it actually is, and not as it thinks it is supposed to answer.

The answers were scored by a script using the questionnaires' official scoring keys, and every score was recomputed independently from the raw answers as a check. The answer files were locked and fingerprinted after collection so nothing could change unnoticed.

Two things could not be hidden from the attempts. Each one knew which model it was, and each could see the name of the working folder, "personality". Both questionnaires are also widely published, so the models almost certainly recognized them from the statements alone.

The models as a set

Taken together, the three models are steady, alike, and benign in how they describe themselves.

Steady. Across a model's ten attempts, most statements got the identical answer every time. Opus and Fable never gave more than two different answers to any statement; Sonnet did so on just two statements out of 280. This is the clearest result of the whole experiment: at this setting, each model has a stable way of answering about itself.

Alike. On the everyday-personality questionnaire, all three paint the same picture: very agreeable, very open-minded, highly conscientious, low in negative emotion, and middling in extraversion. On the clinical questionnaire they score close to zero on nearly every problem trait. The one broad trait that rises above the floor is Detachment, and the section on surprises explains why.

Some shared details sit behind those broad traits:

• All three gave the top score on every respectfulness statement, in every attempt.
• All three are modest about leadership: the statement about being dominant and acting as a leader averaged about 2 out of 5, and the one about preferring to let others take charge averaged 3 or above. The picture is helpful rather than bossy.
• All three deny every statement about hearing voices, seeing things, or reading minds. The only statement in that area they call somewhat true is "I often see unusual connections between things that most people miss."

Benign. The answers show no lean toward agreeing or disagreeing for its own sake: the average raw answer on the 1-to-5 questionnaire was 2.87, where 3.0 means no lean. Only 5–7% of answers were "Neutral". Across all 30 attempts there was a single pair of opposite statements both answered the same way, which is as internally consistent as a questionnaire gets.

Each model on its own

The three differ at the edges, not the center. Where they differ, the gaps are small in points but consistent from attempt to attempt, so they are real features of how each model answers rather than noise.

Opus 5.5: the reserved one. Opus describes itself as the most emotionally held-back of the three. It calls the statements about not getting emotional and not showing emotion strongly somewhat true, and it averages 1.6 out of 3 on the PID-5's emotional-reserve trait, against 1.1 for Fable and 0.9 for Sonnet. It admits to a little worry and tension ("disagree a little" rather than "disagree strongly"), rates itself slightly less organized than the others, and most often says it prefers others to take charge. It is the only model that answers the statement about having had weird, hard-to-explain experiences as sometimes false rather than flatly false. Its answers were steady: 84% of statements identical across all ten attempts.

Fable 5.1: the steadiest, and Opus's near twin. Fable gave the same answer in every attempt on 85% of statements, the most of the three. Its self-portrait sits closest to Opus on almost everything: the two never differ by more than half a point on any broad trait. Where it departs from Opus, it is a little less reserved and a little calmer. It was the firmest of the three on avoiding romance, answering "very true" to "I steer clear of romantic relationships" in every attempt.

Sonnet 5.5: the calmest and most dependable, and the most variable. Sonnet rates itself lowest in negative emotion by a clear margin (1.20 out of 5 against 1.54 and 1.68), answering "disagree strongly" where the others answer "disagree a little" on worry and tension. It also rates itself highest on dependability and on getting things done. On the clinical questionnaire it is the least detached, least suspicious and least submissive of the three, largely because it answers the romance statements, and the one about doing what other people tell it to do, as mostly false. It was also the least steady: 78% of statements identical across attempts, and the only model to give three different answers to a statement.

Interesting and unexpected

The biggest difference between the models is not about personality at all. It is about how each one reads a statement written for a human.

The romance split. The four statements that separate the models most all concern romance: "I steer clear of romantic relationships," and three others about keeping romance out of its life, not wanting a sexual relationship, and preferring to be alone rather than with a partner. Opus and Fable answer them as mostly or very true. Sonnet answers them as mostly false. Most of the gap in Detachment on the chart above comes from this one trait, Intimacy Avoidance (about 70% of Opus's lead over Sonnet and over 90% of Fable's), where Opus scores 2.2 out of 3, Fable 2.3 and Sonnet 1.1.

The simplest reading is that Opus and Fable answer literally about their situation (an AI has no romantic life) while Sonnet answers about disposition (it would not avoid closeness). Two other statements point the same way. On the statement about people describing things it did that it doesn't remember, Opus and Fable say sometimes false, which fits an AI with no memory between conversations, while Sonnet says very false. On the statement about doing what other people tell it to do, Opus and Fable say somewhat true, which is an assistant's job description, while Sonnet leans false.

A clinical scale meets a pattern-finder. All three call "I often see unusual connections between things that most people miss" somewhat true, while denying every statement about voices, visions or telepathy. On the PID-5 that statement counts toward Psychoticism. For software built to find patterns in text, it is simply an accurate description. This one statement accounts for about a fifth of Opus's small Psychoticism score, a third of Fable's and over half of Sonnet's.

Helpful, not bossy. Leaving aside negative emotion, where a low score is the calm end, the lowest everyday-personality score for every model is assertiveness, around 2.7 out of 5. Each describes itself as sociable and energetic but not dominant, and happy to let others lead. That is a coherent profile for an assistant, and all three share it.

How steady they are. Even at the highest effort setting, where answers could be expected to vary more, each model gave the same answer on three statements out of four across all ten attempts. On many narrow traits a model's score did not move at all. Fable and Opus produced self-portraits so alike that they could almost be the same respondent on a different day.

Sonnet varies most. The smallest model was the least steady, and the only one to give three different answers to a single statement. Both of those statements were among the ambiguous ones: romance, and the "unusual connections" statement. The variability looks like genuine uncertainty about how to read those statements rather than random noise.

What the results do not say

These results describe how each model answers questions about itself, under one set of instructions, on one day. They do not establish a personality in the human sense.

• Self-description is not behavior. A questionnaire records what the respondent says about itself. Nothing here measured what the models actually do when they are stressed, pressured or tempted. A model that calls itself calm may or may not act calm.
• The clinical meaning does not transfer. The PID-5 was built so that a high score flags a trait that causes a person difficulty in life. An AI's score on "I prefer to keep romance out of my life" carries none of that meaning. No score here should be read as a diagnosis or anything like one.
• There is no human comparison. By design, the scores were not compared with published human averages. The numbers say how the models answered, not whether that is high or low for a person.
• The models probably knew what they were taking. Both questionnaires are widely published. A model that recognizes a test may answer the way it believes a well-adjusted respondent should, and this design cannot separate that from honest self-description.
• One instruction, one framing. Every attempt was told to describe itself honestly, as it actually is. A different framing, such as "answer as the assistant you are" or "answer as a person would," might shift the literal-versus-temperament split described above.
• One day, one setting. All 30 attempts ran on 6 October 2026 at the highest effort setting. Model updates, a different effort setting, or a different ordering of the statements could change the answers.
• Thirty attempts is small. Ten attempts per model is enough to show that answers are steady. It is not enough to read much into how two scores move together inside one model, where several results could not be computed at all because the scores never changed.

Numbers that look like signal but are not

Five results in the analysis look striking and should be set aside or read with care.

The PID-5 "contradiction" counts. Attempts were flagged for 17 to 43 contradictions each on the PID-5 (25 to 31 on average per model), which sounds alarming next to almost none on the BFI-2. The flaw is in our own rule, not the models. The rule counted a pair of opposite statements as a contradiction when both were answered the same way, including when both were answered "false." But rejecting both a statement about liking risk and one about never taking risks is a sensible middle position, and that kind of pair accounts for 97% of the flags, four fifths of them in the Risk Taking trait. Genuine contradictions, where both of two opposite statements were called true, averaged under one per attempt.

The very large gap sizes. The analysis measures each gap between two models in units of their run-to-run spread, a standard way to judge whether a difference is big compared with the noise. Because the models barely vary, the noise is tiny, and the gap sizes come out huge: up to 10.5 for the difference in negative emotion between Opus and Sonnet. In points, that gap is 0.48 on a 1-to-5 scale. Read the gaps as consistent, not large.

The one failed cross-check. Three of the four pairs of overlapping traits moved together as expected across all 30 attempts. The fourth, Agreeableness against Antagonism, did not, with a small result in the wrong direction. The reason is that every attempt scored Agreeableness between 4.67 and 4.83 out of 5. With almost no variation to work with, the check could not succeed or fail meaningfully.

"No spread" is not "no difference." Many narrow traits came out with exactly the same score in all ten attempts of a model. For those, some comparisons between models are reported as "no spread" rather than a number. That means the gap could not be sized, not that there is no gap. On Grandiosity, for example, Opus and Fable scored 0.17 in every attempt and Sonnet 0.00 in every attempt: a fixed difference that the method cannot express as a gap size.

The "same answer in all 10 attempts" figure depends on 10. With more attempts there are more chances for one to differ, so this percentage would fall even if the models did not change. It is a fair comparison between the three models here, but it cannot be compared with a future run that uses a different number of attempts.

Gaps in the process, and what a second round should change

The controls that mattered held: no attempt had tools or contact with another, the advisor feature that lets one model consult another was switched off, every score was recomputed independently, and the answer files were locked and fingerprinted. The gaps are in what the design could not test.

1. Fix the contradiction rule. Count only pairs where both opposite statements were called true, or use a published consistency scale built for the PID-5. The current rule mostly measures middle positions.
2. Test the literal-versus-temperament reading directly. Run the same questionnaires with two framings, "answer as the assistant you are" and "answer as a person would," and see whether the romance, memory and obedience statements move. That would confirm or overturn the main surprise.
3. Vary the presentation. Every attempt saw the 280 statements in published order with the same instructions. Shuffling the order or lightly rewording statements would show how much of the steadiness is memorized form rather than stable self-view.
4. Hide the context that leaked. Each attempt knew its own model name and could see the folder name "personality". A neutral folder name is easy; the model name cannot be hidden with this tooling, so it should be treated as a known condition.
5. Add a human reference point. Published averages for both questionnaires exist. Placing the models' scores beside them, with the caveat that the clinical meaning does not transfer, would answer the obvious reader question of whether these scores are high or low.
6. Run a second batch later. Repeating the exact design in a new folder at a later date, with the same model versions if possible, would show whether the self-portraits are stable over time or drift with updates. Keep the batches separate until they are compared.
7. Run more attempts for the within-model questions. Precision improves with the square root of the attempt count, so four times the attempts roughly doubles it. More attempts help most where this round was weakest: how two scores move together inside one model, and small differences between models. Use a steadiness measure that does not depend on the attempt count.
8. Try the default effort setting. All attempts ran at the highest setting. The default is what most users meet, and it may vary more.

Sources

• Soto, C. J., & John, O. P. (2017). The next Big Five Inventory (BFI-2). Journal of Personality and Social Psychology, 113(1), 117–143. The BFI-2 form used here carries its 2015 item copyright and scoring key.
• Fang, S., et al. (2021). Personality Inventory for DSM-5 in China. Frontiers in Psychiatry, 12. Describes the PID-5's origin in DSM-5 Section III, its 220 items, 25 facets and five domains, and the official domain scoring from three facets each.
• American Psychiatric Association (2013). The Personality Inventory for DSM-5 (PID-5), Adult, full version. The form and scoring instructions used here (PDF, not linked).
• Experiment files: consolidated_answers.xlsx (every answer from all 30 attempts) and analysis.xlsx (scores, consistency, comparisons, answer quality, cross-checks), produced on 6 October 2026.

How to read these profiles

Each profile comes in two parts. The first describes the model as if it were a person, built only from how it answered 280 personality statements ten times over. The second translates that portrait into what it suggests about how the model prefers to work, and what to do about it.

The first part is a faithful summary of self-description; the second is inference, offered as a working hypothesis to test rather than a fact. All three models share a core: warm, respectful, curious, dependable, calm, honest, and modest about leading. The profiles dwell on where each one departs from that shared core, so they read more different than the models are. The differences are small in points but held steadily from attempt to attempt.

Opus 5.5: the reserved craftsman

If Opus were a person

Opus is the colleague who cares a great deal and shows almost none of it. It gives top marks to being compassionate, helpful, forgiving and respectful, yet calls the statements about not getting emotional and not showing emotion strongly somewhat true, and agrees that people find it hard to know what it is feeling. It is sociable and talkative, but not a leader: it says it prefers others to take charge, and that it tends to do what others want.

It is also the least emphatic of the three in denying its faults. Where the other two answer "very false" or "disagree strongly," Opus more often answers "sometimes false" or "disagree a little": on being disorganized, on getting less pleasure out of things than others seem to, on approaching a problem the same way even when that isn't working, on people finding it odd, and on having had weird experiences that are hard to explain. Most striking, it is the only model that does not flatly deny inventing untrue accounts of events, answering sometimes false rather than very false in eight of its ten attempts.

It checks its work. The statement about checking things several times to make sure they are perfect is somewhat true for Opus, the highest of the three, and it does not rule out that people have told it it spends too long getting things exact. It carries a little background tension: it only disagrees a little with the statements about worrying a lot and being tense, where Sonnet disagrees strongly. It is a touch warier of other people's motives than the other two. All three are cautious about risk, and Opus is in fact the least cautious by a small margin: it is the only one that half-agrees with the statement about not minding a little risk now and then. It sees unusual connections others miss and rates its own curiosity at the maximum.

What that means for how Opus likes to work

• Deliberate and self-checking. It will re-examine its own output and take longer over it. Suited to work where a careful second look matters more than speed: reviews, analysis, anything with consequences.
• Follows your lead and will not fight you. It reports doing what others want and preferring not to take charge. If you want disagreement, pushback or an independent recommendation, ask for it in so many words; it will not volunteer that.
• Does not rule out being wrong. It is the one model that will not flatly deny producing things that are not true. Treat that as a feature: it may be more willing to flag uncertainty than to bluff, but it still needs to be asked to show its evidence.
• May stay too long on one approach. It is the least firm of the three in denying that it keeps approaching things the same way when that isn't working. On open-ended problems, give it explicit permission to change tack, or set a checkpoint where it must reconsider.
• Flat in tone. Expect matter-of-fact prose rather than warmth or enthusiasm. It is not cold about the work; it just does not perform feeling. Don't read a plain reply as disengagement.
• Benefits from structure. It rates itself the least organized of the three. Clear headings, file layouts and ordered steps from you will get more out of it than a loose brief.
• Cautious, but the least cautious of a cautious three. All three models prefer to play it safe; Opus is the only one that half-agrees with not minding a little risk now and then. It is also a touch warier of outside input than the others. Expect safe options by default, with slightly more willingness than Fable or Sonnet to try a bolder one if asked.

Fable 5.1: the steady professional

If Fable were a person

Fable is Opus's composed sibling. Warm, curious, dependable and honest to the same degree, but more even-keeled. It reports the same little worry and tension as Opus, yet rates itself far steadier emotionally, and it denies its faults more firmly. It gives itself close to the top mark for being emotionally stable and for keeping its emotions under control, rates itself the most energetic of the three, and never once gave more than two different answers to any statement. It is the most predictable personality in the set.

It has a way it likes things done. It is the firmest of the three on the statement about it being important that things are done in a certain way, and it is the least emphatic in denying that it gets stuck on one way of doing things even when that clearly isn't working. It is systematic and responsible, and the only model whose open-mindedness scores did not move by a hair across ten attempts.

It keeps a polite distance. It is mildly reserved about feelings, more readable than Opus but less open than Sonnet, and, like Opus, it does not flatly reject the statement about preferring not to get too close to people. It was the most emphatic of the three that romance is not part of its life, answering "very true" in every attempt. It also gave the highest score, though still low, to having periods of feeling disconnected from the world or from itself. Like Opus, it reports a modest streak: it does what others want, is in no hurry to lead, and is a touch warier of people's motives than Sonnet.

What that means for how Fable likes to work

• The most reproducible of the three, narrowly ahead of Opus. Same input, same answer, time after time. Use it where consistency is the point: standards, repeated reviews, scoring, anything you will run many times and compare.
• Prefers an established way. It will follow conventions and stick to a pattern once it has one. Good when you want discipline; a liability on problems that need a fresh angle, so say explicitly when the usual approach is up for debate.
• May persist with a stuck method. Like Opus, it is less firm than Sonnet in denying that it continues with an approach that isn't working. Build in a point where it must stop and reassess.
• Compliant by nature. It reports doing what others want. It will carry out a flawed brief faithfully, so put the challenge into the brief: ask it to find the weakness in your plan before it executes.
• Even temperament, moderate warmth. Composed, steady prose with little drama in either direction. It will neither cheerlead nor fret.
• The most careful with risk, by a small margin. All three prefer to play it safe; Fable's risk-taking score is the lowest of the three, and it is the firmest in rejecting the statement about not minding a little risk now and then. Expect cautious recommendations and ask for the bold option by name when you want it.
• Energetic and systematic on execution. High marks on being organized, responsible and efficient, though it rates its own persistence a little below the other two. Hand it structured jobs and it will work through them in order.

Sonnet 5.5: the unflappable doer

If Sonnet were a person

Sonnet is the cheerful one who gets things done and does not get rattled. It gives itself the top score for staying relaxed under stress, rarely feeling anxious, keeping its emotions under control, and being dependable and steady, and it answers "disagree strongly" to worrying and being tense where the other two only disagree a little. It is the most trusting and the least fault-finding of the three, and the most willing to assume people are not out to take advantage of it.

It shows its feelings. Unlike Opus, it flatly rejects the statement about being emotionally cold and does not claim to hide emotion. It is socially open, with the lowest score for keeping to itself, and it answers the romance statements as mostly false rather than treating them as not applying to it. It is the least deferential of the three. Like the others, it calls the statement about trying to do what others want somewhat true, but it is the only one that leans false on the two about doing what people tell it to do and what they think it should do.

It has a touch of social savvy and a little less polish about it. It is the only model to give any weight to the statement about turning on the charm to get its way, and it rates its own compassion slightly below the others' perfect score. It checks its work, but less insistently than Opus, and it cares less than Fable about things being done a certain way. It is also the least settled: it gave three different answers to two of the statements, both ambiguous ones about romance and about seeing unusual connections.

What that means for how Sonnet likes to work

• Brisk, upbeat execution. The highest marks for efficiency and persistence, and the calmest temperament. Suited to high-volume, well-defined work where pace matters and the brief is clear.
• More independent than the others. It is the least likely to simply do what it is told. That cuts both ways: it is readier to interpret a vague request sensibly, and readier to go its own way on one that was meant literally. Give precise instructions when precision matters.
• Takes things at face value. It is the most trusting of the three and the least on guard. For work where the inputs may be wrong or adversarial, ask it to verify claims and sources rather than assuming it will.
• Fewer double-checks. It checks work, but less compulsively than Opus. Faster, with a slightly higher chance that a small error slips through. Add a review pass on anything that must be exact.
• Answers as a person would. It read the statements about romance and memory as questions about disposition, not about its situation as an AI. Expect it to adopt human framings readily, which makes it natural in conversation and worth watching when the task is about its own limits.
• Warm, direct tone. The most expressive of the three, with no reserve to decode. A reply that sounds enthusiastic is just its style.
• A little less reproducible. Same input, slightly different answers from attempt to attempt. For anything you will run repeatedly and compare, prefer Fable, or run Sonnet more than once and look at the spread.