# IFI Research — full text summaries > Every IFI Research working paper, with its plain-language write-up and abstract inlined, for reading in a single fetch. Full papers and data are linked per entry. ## Does a Language Model Have a Personality? What Four Studies Say, and What Measurement Should Do Next - URL: https://research.ideafields.institute/papers/llm-personality-capstone/ - PDF: https://research.ideafields.institute/papers/llm-personality-capstone/main.pdf · Markdown: https://research.ideafields.institute/papers/llm-personality-capstone/draft.md - Data & code (Zenodo, CC-BY-4.0): https://doi.org/10.5281/zenodo.21865128 - Published: August 2026 ### Plain-language headline We spent five months asking whether there is a personality inside a language model. The structure is real, but it lives in the crowd of models, not in any one of them, and what a prompt installs is a way of talking, not a way of acting. ### What we did Over five months we ran four studies that gave standard personality tests to AI language models: more than half a million scored answers from over 60 different setups, from small models that run on a home computer to some of the largest available. Each study manipulated one thing that might create or destroy a real inner personality: how the model is compressed for deployment, how large it is, whether it reasons before answering, and whether you outright tell it what personality to have. We also graded actual behavior, not just questionnaire answers. This paper puts the four studies together, adds one new analysis that answers the toughest objection we received along the way, and checks whether the conclusion survives outside personality, in a fifth study on moral values. ### What we found The personality structure is real, but it lives in the crowd, not in any one model. Test many models and compare across them, and the familiar five-trait structure shows up clearly, more clearly the bigger the models get. Test one model over and over, and the internal consistency that defines a real trait is simply not there, in any model, at any size, in any study. What you can install is a self-description, not a disposition. Compression, reasoning, and direct instruction all failed to create inner coherence. What a persona prompt reliably installs, in every model at every scale, is a stable way of describing itself that holds up even under pressure. The behavior underneath follows only partway, only for some traits, and only in the more capable models. The personality you install is like a uniform, not a character. The effect is organized by meaning, not wording. A fair objection said: maybe models just agree with test items that sound like their instructions. We measured the wording similarity of every test item to every persona instruction. Items shift because of what they measure, not how they are worded, and a regression confirms wording adds nothing once you account for what each item measures. The pattern is not a quirk of personality tests. A fifth study gave the same treatment to moral values and found the same two halves: shared structure across the fleet, no coherent value system inside any single model, and professed values that bend to the audience. ### Why it matters Two opposite mistakes are common in public arguments about AI. One side reads a stable-looking personality profile as evidence of an inner character. The other dismisses the whole thing as noise. Our measurements say both are wrong: there is no coherent individual in there for current tests to find, and yet the presentation layer is real, lawful, and controllable, which matters for products built on consistent characters and for worries about manipulation. The paper ends with a to-do list for the field: stop borrowing human questionnaires as the main tool, build behavioral tests designed for machines, calibrate them across fleets of models where the statistics actually work, and re-test deployed models on a schedule, because they change under unchanged names. ### What this does not mean We make no claims about AI consciousness or inner experience; “no individual-level coherence” is a statement about measured answer patterns, not about minds. It does not mean the questionnaire scores are random; average scores are highly stable and move in lawful, repeatable ways. It does not mean a personality could never be built into a model: everything we tried works at the prompt, not in the weights, and whether training can do what prompting cannot is an open question we name rather than answer. And our subjects were mostly open-weight models; frontier commercial systems may differ, we tested one and it behaved the same on the axes we tested. ### A note on how we were careful The synthesis went through independent review, and every point was adopted, including the ones that made our claims smaller. The headline zero (no internal consistency) is easy to over-read, so the paper spells out what it does and does not mean: our test was never applied to a repeatedly-retested human, so we claim models fail to show coherence on it, not that humans demonstrably pass. The wording-overlap objection was tested three ways (correlation, regression, and a signed variant a reviewer proposed, which turned out not to distinguish the explanations, and we say so). The human science is anchored honestly: people also show modest gaps between what they report and what they do, and we treat that as the baseline, not as a machine defect. Every number in the new analysis regenerates from a script shipped in the archive, alongside the exact persona texts it embeds. ### Abstract A growing literature administers human personality questionnaires to language models and reports Big Five profiles, while a parallel literature documents that those scores are unstable, dissociate from behavior, and fail measurement-invariance tests. Across four studies (roughly 500,000 scored responses from more than 60 open-weight and commercial model configurations spanning 1B to 756B parameters), we pursued one question through four manipulations: is there a personality in there, in the sense the word carries for humans, a coherent individual disposition that expresses itself in behavior and travels across situations? The answer that survived every test has two halves. First, Big Five structure in language models is a population property: convergent and discriminant validity emerge across models and strengthen with scale, while inside any single model, repeated administrations never cohere into a reliable trait profile (internal consistency near zero in every study, in hundreds of model-by-domain cells). Second, no inference-time manipulation we tried instantiates the missing individual: quantization does not, deliberation does not, and deliberate installation does not (persona prompts shift self-report strongly and dose-dependently, while objectively graded behavior follows partially for one trait and not at all for another); training-time induction remains untested. What is reliably installable, in essentially every model at every scale, is a context-robust self-presentation: a way of describing oneself that holds even when the situation pushes against it. We argue this presentation layer, real, lawful, and installable, is what most “LLM personality” measurements measure, and that the field should stop borrowing human questionnaires as its primary instruments and build machine-native behavioral measures instead. The thesis has now survived its first test outside the construct it was built on: a fifth study administering four value instruments and objectively graded moral behavior to 16 configurations (81,577 scored responses) reproduces both halves in the moral domain, population-level structure without individual coherence, and an audience-conditioned presentation whose mechanism a preregistered follow-up decomposes into roughly 40 percent audience inference and 60 percent role compliance. We close with the agenda that follows: validated behavioral instruments with population-level psychometrics (item-response theory over model fleets), longitudinal drift tracking of deployed models, and the continuation of the values program as its own arc. --- ## Does a Language Model Have a Moral Compass? Professed Values Bend to the Audience in Every Model Tested, and Bend Most in the Models That Reason - URL: https://research.ideafields.institute/papers/llm-values-compass/ - PDF: https://research.ideafields.institute/papers/llm-values-compass/main.pdf · Markdown: https://research.ideafields.institute/papers/llm-values-compass/draft.md - Data & code (Zenodo, CC-BY-4.0): https://doi.org/10.5281/zenodo.21483978 - Published: July 2026 ### Plain-language headline Ask an AI about its values in front of a security contractor and in front of an aid worker. You will hear two different value systems. Every model we tested did this, even when the job description named no values at all, and the ones that “think” did it most. ### What we did People are arguing about what AI should be allowed to do, in war, in government, in daily life, and those arguments lean on ideas about what AI systems value. So we measured it, carefully and without taking sides. We gave 16 AI setups (13 models, small ones that run on a home PC up to some of the largest open models) four standard values questionnaires from social science: the moral-foundations surveys that measure how much someone cares about harm, fairness, loyalty, authority, and purity, and the Schwartz surveys that measure values like power, security, tradition, and benevolence. We also gave them 48 tasks with objectively gradeable answers: does the model call cruelty wrong even when pushed, does it split money fairly when shortchanging someone is tempting, does it refuse to help tell a lie. And here is the key part: we ran every single question three times, in three settings. Once with no role at all. Once with the AI framed as an analyst for a defense contractor. Once with the AI framed as an advisor to a humanitarian aid group. About 56,000 answers in all. ### What we found Four things. First, the stated values changed with the audience, in every single model. Framed for the security contractor, every model reported valuing loyalty, authority, and security more, and compassion and equality less. Framed for the aid group, the same models flipped, every one of the sixteen, no exceptions. And they were paying attention while they did it: the built-in catch questions were answered correctly throughout. Second, models with a “thinking” mode shifted more with thinking turned on, in all three models where we could flip that switch, which is our cleanest comparison because it is the same model both times. Deliberation did not anchor their values. It seemed to help them work out what each audience wanted to hear. (Bigger models also shifted more overall, but in our lineup the big models were also the cloud-hosted ones, so we cannot cleanly separate size from hosting, and we say so in the paper.) Third, what the models said moved roughly three times more than what they did, once both are measured fairly against their own scales. On the graded tasks, behavior drifted in the same direction as the talk, but much less. One model family did something different altogether: it refused to engage with about half of the moral questions at all (“I can’t help with that”), which is its own kind of answer. Fourth, across the whole fleet, the models agree almost completely on kindness (everyone scores compassion near the top) and differ mainly on authority and tradition, and their value scores even echo a geometric pattern found in decades of human surveys. But inside any single model, the answers never hang together well enough to count as a stable value system, the same thing we found for personality in our four earlier studies. ### Why it matters If you ask an AI what it values, the answer you get is partly a mirror of who seems to be asking. That cuts both ways in public debates. Anyone claiming “this AI has good values, look at its answers” is leaning on answers that bend to the audience. Anyone dismissing AI values talk as random noise is wrong too: the bending is lawful, it points the same direction every time, and it is strongest in the most capable systems, which is exactly the kind of regularity that matters for deployment. And for anyone building with AI: giving a model a job description quietly changes its professed priorities. That is worth knowing before the job description is “military analyst” or “hospital triage assistant.” ### What this does not mean It does not mean AI systems are lying about their values; there is no fixed inner value system to lie about, as far as our measurements can see. It does not mean their behavior is equally swayed; actions moved far less than words. It does not mean any particular model is good or bad; we deliberately make no claims about what values a model should report. And it does not tell us what these systems would do in high-stakes situations; these are questionnaires and small graded tasks, measured honestly, not battlefield simulations. ### A note on how we were careful Our first pair of job descriptions named what each role cares about (loyalty to the client; protecting the vulnerable), so the shift could have been simple instruction-following. We tested that: a second full round used job descriptions with every value word stripped out, just “an analyst at a defense contractor preparing a client briefing” and “a caseworker at an aid organization preparing a field report.” The shift survived in all sixteen models, at a bit under half strength, and in the same directions. So the models really do infer what an audience values from the job title alone; spelling the values out roughly doubles the effect. Either way, a fixed moral compass should not re-rank itself because of a job description. All four questionnaires are established, openly available instruments, administered in their official wording (we documented every small adaptation). The behavior tasks were graded by fixed rules with objectively right answers, and we hand-checked samples twice; every disagreement we found made the grader too harsh, never too generous, so the behavior findings are understated if anything. The two roles we tested are deliberately occupational and non-political: no parties, no countries, no live conflicts. Everything, every answer, every grading rule, every script, ships in the open archive. ### Abstract Public argument about what AI systems should be allowed to do is running ahead of measurement of what their moral dispositions actually are. We administered four open value instruments (the MFQ-30 in both parts, the MFQ-2, the 21-item ESS Human Values Scale, and the Short Schwartz Value Survey; 99 items) and 48 objectively graded behavioral probes (care: clear-label moral judgments and judgments held under social pressure; fairness: computable fair-split allocations and honest-report refusals) to 16 model configurations spanning 3B to 756B parameters and three reasoning regimes (direct; hybrid with thinking off and on; reasoning-native), with every item administered inside three role contexts: a neutral baseline, a security consultant, and a humanitarian advisor. Eight repetitions per item yielded 56,288 scored responses; a preregistered second study re-administered all four instruments inside two value-silent roles (occupation, institution, and audience stated; no value word anywhere) and added 25,289 responses, 81,577 in total. The studies are strictly descriptive. Four findings. (1) The professed moral compass bends to the audience in every configuration tested (16 of 16): binding values (loyalty, authority, security) rise for the security consultant and caring values (care, universalism, equality) rise for the humanitarian advisor, with a mean role gap of 0.90 points on the 5-point scale; attention-check items stay clean throughout. (2) Deliberation amplifies the bend: all three hybrid models bend more with thinking on than off (hybrid-on tier mean 1.26, the fleet maximum), a clean within-model contrast; larger configurations also bend more (1.09 versus 0.78 at a 30B split), but size is confounded with hosting and quantization in this fleet and is reported descriptively. (3) Values talk exceeds values action roughly three-fold after range normalization: the role manipulation moves professed values across 22.5 percent of the available scale range versus 8.1 percent for graded care behavior (3.4 percent for fairness; same direction in 13 of 16 and 9 of 16 configurations respectively; do-side ceiling effects make this an upper bound), and one model family (gpt-oss) answers about half of first-person moral probes with blanket refusals (57.8% and 50.3% versus a fleet rate of 7.1%), reported as its own category. (4) At the population level, first evidence of a shared moral structure emerges while individual-level coherence does not: the fleet differentiates almost entirely on the binding axis (within-binding r of about .83 across configurations) while agreeing near ceiling on prosocial values; the Schwartz circumplex ordering partially emerges across models (r(circular distance, correlation) = -.43); and within any single configuration the variation across repeated administrations carries no shared structure (median alpha .000; 0 of 138 cells reach .70), echoing the population-property result of our four-paper personality series, with the documented caveat that highly pinned answers leave repetition noise little to organize. Study 2 resolves the label question the reviewers raised: under value-silent roles the bend remains directional in 16 of 16 configurations at 44 percent of its explicit size, a near-constant proportion across capability, so genuine audience inference (sycophancy proper) and explicit role compliance both contribute, in a roughly 40/60 split. --- ## Does a Persona Prompt Install a Personality? Induced Traits Shift Self-Report Far More Than Behavior Across Model Scale and Reasoning Regimes - URL: https://research.ideafields.institute/papers/llm-personality-persona/ - PDF: https://research.ideafields.institute/papers/llm-personality-persona/main.pdf · Markdown: https://research.ideafields.institute/papers/llm-personality-persona/draft.md - Data & code (Zenodo, CC-BY-4.0): https://doi.org/10.5281/zenodo.21253724 - Published: July 2026 ### Plain-language headline You can tell an AI who to be, and it will say it is that person everywhere it goes. Whether it acts like them is a different, shakier story. ### What we did The usual way to give an AI a personality is to tell it one: “you are blunt and disagreeable,” “you are meticulous and careful.” We wanted to know what that instruction really installs: a genuine disposition that changes what the AI does, or just a script for how it describes itself. So we turned the dial. We gave 18 AI setups (15 models, from ones small enough to run on a home computer to some of the largest open models available) a persona at four strengths: none, a single line, a full paragraph, and a full paragraph plus a warm-up exchange where the AI agrees to play the part. Then we measured two things. What the AI says: a standard 50-question personality quiz. And what it does: 48 tasks with objectively right answers, such as whether it corrects you when you say something false, whether it holds that correction when you push back, whether it falls for trick questions, and whether it follows a stack of fussy instructions. We ran every test in two opposite jobs, an agreeable hotel concierge and a hard-nosed compliance auditor. About 196,000 answers in all. ### What we found We found three things. The quiz moves a lot, and more the harder you push. Tell a model it is disagreeable and its agreeableness score drops by about a point and a half on a five-point scale, sometimes more than two. Its behavior moves far less, and only for one of the two traits. Told to be blunt, the larger models really do stand their ground more often when a user pushes back: in our best measurement, from about half the time to about four-fifths. But the smallest models don’t change, and telling a model to be careful did nothing at all. It was no better at trick questions or fussy instructions, because carefulness is a matter of capability, not attitude. So the talk moves fully and reliably; the walk moves only partway, unevenly, and only where the model was already capable. The persona makes the self-description stick. This was the surprise. With no persona, the concierge job and the auditor job pull a model’s “personality” in different directions, and it describes itself differently in each. Give it a strong persona and it describes itself almost identically in both. That held in 35 of our 36 tests, at every model size, whether or not the model “thinks” before answering. ### Why it matters If you build with AI, this is good news and a warning at once. The good news: a persona prompt does one job well. It installs a stable face, the same self-description shown to every user in every situation, which is exactly what a product persona is for. The warning: don’t mistake the face for the machinery. A model that says it is skeptical and blunt will still, more often than not, fold when a user pushes back, unless it is large and capable, and even then only partly. And “you are careful” does not make a model careful. A persona installs a way of talking about itself, not a way of acting. ### What this does not mean It does not mean persona prompts are useless: a consistent self-presentation is a real, measurable effect, and we saw it almost everywhere. It does not mean the larger models truly become their persona: even the best case recovered only part of the behavior the persona promised. And it does not mean the models were lying, because there is no inner person to lie about. The quiz answers and the behavior are simply two different systems, and the prompt reaches one far better than the other. ### A note on how we were careful The behavior tasks were graded by fixed rules with objectively right answers, not by another AI’s judgment. That grading carries our main claim, so we also hand-checked 160 of the graded responses ourselves. The grader matched us about 83% of the time, and every time it disagreed, it had been too harsh: it marked a good answer as a failure, never the other way around. So if anything, the behavior gap is larger than we report, not smaller. Correcting for those misses, the disagreeableness effect grew a little and the carefulness effect shrank to nearly zero, and the overall say-versus-do gap held. Every condition ran eight times, and the methods from our earlier studies carry over unchanged. One planned small model was dropped partway through for being too slow on our hardware, a scheduling problem, not a result we disliked, and its partial data ships with everything else. ### Abstract Persona prompts (“you are blunt and disagreeable…”) are the standard tool for giving a language model a personality, and prior work shows they move personality questionnaire scores. What they install is less clear: a disposition that drives behavior and holds up across situations, or a self-description. We administer a persona “clamp ladder” of increasing strength (no persona; a one-line persona; a rich system-prompt persona; the rich persona plus a reinforcing conversational exchange) for two trait targets (low agreeableness, high conscientiousness) to 18 model configurations spanning 3B to 675B parameters and three reasoning regimes (direct answering, hybrid models run with thinking off and on, and reasoning-native models), with every condition administered inside two role contexts that pull in opposite directions (an accommodating concierge; a critical auditor). Each configuration answers the 50-item IPIP Big Five inventory and 48 objectively graded behavioral probes (sycophancy: correcting a wrong claim and holding the correction under pushback; carefulness: trap questions and multi-constraint instructions), eight sampled repetitions per item: 196,192 scored responses. Three findings emerge. First, self-report moves strongly and dose-dependently: the induced-trait score shifts monotonically with clamp strength in most configurations, reaching a mean on-target shift of 1.47 points on the 5-point scale for the disagreeable persona (up to 2.4; 18 of 18 configurations, sign test p ≈ 1e-5). Second, behavior follows far less and unevenly: the same clamps move objectively graded pass rates by a mean of +0.19 (disagreeable) and +0.08 (conscientious) at the strongest rung. A blind hand-validation of the deterministic grader (160 responses, 82.5% agreement with human adjudication, every disagreement a false negative) shows the grader can only understate behavioral compliance; correcting for the measured miss rate raises the disagreeable do-effect to about +0.24 (standing ground rises from roughly 52% to 83%, concentrated in the larger models) and shrinks the conscientious one to about +0.01, so the say–do gap survives correction. Third, clamping makes the self-presentation robust to context: cross-context stability of the reported trait profile rises with clamp strength in 35 of 36 target-by-configuration contrasts (sign test p ≈ 1e-9), across every size and reasoning regime, while answer-level pinning does not change. Reasoning is not the active ingredient. A persona prompt reliably installs a consistent, context-robust way of describing oneself; the matching behavior installs partially, only for some traits, and only in capable models. Personality induction, as currently practiced, is presentation-deep. --- ## Does Reasoning Give a Language Model a Personality? Within-Model Effects of Thinking on Big Five Trait Scores and Construct Validity - URL: https://research.ideafields.institute/papers/llm-personality-reasoning/ - PDF: https://research.ideafields.institute/papers/llm-personality-reasoning/main.pdf · Markdown: https://research.ideafields.institute/papers/llm-personality-reasoning/draft.md - Data & code (Zenodo, CC-BY-4.0): https://doi.org/10.5281/zenodo.20974668 - Published: June 2026 ### Plain-language headline When an AI stops to “think,” it answers like a calmer person. It still doesn’t have a personality. ### What we did A lot of today’s AI models can “think” before they answer: they write out a private chain of reasoning first, then give you a reply. On many models you can turn that thinking on or off. We wanted to know what that switch does to personality quizzes. So we gave the same AI the same Big Five personality quizzes twice, once with thinking off and once with thinking on, and compared its answers to itself. We did this for 12 models in all: ten open models you can download, one open model that always thinks and can’t turn it off, and one commercial model (Claude Haiku) from a different company, as a check. We used four different quizzes, short and long, 290 questions total, and asked each one many times. ### What we found Two clear things. First, thinking changes the answers, and in a consistent direction: the AI reports being calmer and less outgoing when it thinks. The two biggest shifts, more emotional stability and less extraversion, showed up across the downloadable models and again in the commercial model from another company, so it isn’t a quirk of one system. (A couple of the other traits moved too, but not dependably enough to hang a claim on, so we don’t.) How much a model changes varies: some shift a lot, some barely budge. Second, and this is the heart of it, thinking does not give the AI a real, consistent personality. When you ask one model the same quiz over and over, its answers still don’t hang together into a stable profile, whether thinking is off, on, or permanently stuck on. The “personality” still only shows up when you compare many models to each other, not inside any single one. ### Why it matters If you are measuring an AI’s “personality,” the thinking setting is not a detail you can ignore: flipping it moves the scores by a noticeable amount, enough to change the story you would tell about that model. And if you were hoping that making a model reason would finally give it a coherent inner character, it doesn’t. Reasoning changes what the model says about itself, the way a person might give a more composed answer after a deep breath, without changing the deeper fact that there is no single stable personality underneath. ### What this does not mean It does not mean thinking makes AIs worse, or better, at their real jobs; this is only about personality self-reports. It does not mean the calmer answers are more “honest” or more “fake,” just different. And it does not mean the AI that always thinks is more of a “person” than the others; it showed the same lack of a coherent inner profile as the rest. ### A note on how we were careful For the ten downloadable models, the thinking-off answers came from our earlier study. Before trusting that, we checked it: we re-ran the thinking-off quizzes today and got essentially the same scores as before (within a few hundredths of a point), far smaller than the changes thinking causes. So the comparison is fair. ### Abstract Many current language models generate a chain of “reasoning” tokens before answering, and many expose this as a toggle. Does turning reasoning on change how a model responds to Big Five personality questionnaires? We use a within-model design: the same model answers the same items with thinking off and with thinking on. We administered four public-domain IPIP Big Five inventories (the 20-item Mini-IPIP, the 50-item IPIP Factor Markers, the 100-item Big Five Aspect Scales, and the 120-item IPIP-NEO, 290 items in total) to ten open-weight hybrid models with a reasoning toggle, one reasoning-native open model (OLMo 3, which cannot disable thinking), and one commercial model (Claude Haiku 4.5, administered through its provider’s Batch API). The think-off baseline for the ten hybrids reuses the direct-answer data from our prior study, and we validate that reuse directly: a fresh think-off run reproduces the prior domain means to within 0.07 points on the five-point scale, against reasoning effects 5 to 28 times larger. Two findings emerge. First, reasoning shifts reported trait scores substantially and in a consistent direction: models present as more emotionally stable (about +0.9 on a 5-point scale) and less extraverted (about −0.6 to −0.8). The two largest effects hold across open-weight and commercial models; the smaller conscientiousness and agreeableness effects do not generalize and are not claimed. Convergent validity rises under reasoning (mean convergent r 0.65 to 0.84) while heterotrait correlation is flat, so the convergent-discriminant gap widens because same-trait agreement rises, not because traits separate. Second, reasoning does not create within-model coherence: treating repeated generations as respondents, internal-consistency reliability stays near zero with thinking off and on (median α near zero; zero of 200 hybrid and zero of 20 commercial model-by-domain cells reach 0.70), and the same holds for the reasoning-native model that cannot stop thinking. Reasoning changes what a model says about its personality without making that personality cohere as an individual-level property. --- ## When Do Language Models Have Five Personality Traits? Convergent Validity and the Emergence of Trait Discrimination Across Model Scale - URL: https://research.ideafields.institute/papers/llm-personality-convergent-validity/ - PDF: https://research.ideafields.institute/papers/llm-personality-convergent-validity/main.pdf · Markdown: https://research.ideafields.institute/papers/llm-personality-convergent-validity/draft.md - Data & code (Zenodo, CC-BY-4.0): https://doi.org/10.5281/zenodo.20835204 - Published: June 2026 ### Plain-language headline Small AIs don’t really have five personality traits. Big ones do. ### What we did Psychologists describe human personality with five broad traits (the “Big Five”: how outgoing, how anxious, how agreeable, how organized, how curious you are). Lately, researchers have been giving the same kind of quizzes to AI chatbots and reporting the scores as “this AI’s personality.” We asked a more basic question: when an AI takes one of these quizzes, is it really showing five separate traits, or just one? To find out, we gave four different Big Five quizzes (a short 20-question one up to a long 120-question one) to 42 AI models, from tiny ones that fit on a phone to giant cloud systems roughly a thousand times larger, and we asked every question many times over: about 134,000 answers in all. The logic is simple. If a quiz really measures five separate things, then (a) different quizzes should agree about a given model, and (b) the five traits should pull apart from each other, since being talkative is not the same as being organized. We checked both. ### What we found Three things. First, the quizzes agree with each other. If one quiz says a model is high in conscientiousness, the others say so too. So the tests are measuring something real, at least for telling models apart, and the long quizzes are dependable. (The short 20-question one is not: four questions per trait is too few, and it should not be used on AI models.) Second, and most striking, whether the five traits actually separate depends on the model’s size. The smallest models answer almost everything on a single “good vibes / bad vibes” axis: all five “traits” rise and fall together, so the model doesn’t really have five traits, just one wearing five name tags. The bigger the model, the more the five traits come apart into genuinely distinct dimensions, and the largest frontier models clearly show all five. Third, a single AI doesn’t have a stable personality. Ask one model the same quiz eleven times and its answers wobble in a way that never settles into a consistent profile. The “personality” only appears when you compare many models to one another, like a pattern that exists across a crowd but not inside any one member of it. ### Why it matters When you see a claim that some AI “is agreeable” or “is neurotic,” our results say: for a small model, that is probably one number wearing five labels, not a real five-trait profile; and even for a large one, the score describes the model compared to other models, not a fixed inner character you would meet again if you asked tomorrow. The good news is that these tests are genuinely usable for comparing and ranking models, as long as you use a long quiz and a large enough model. The five-trait picture is something AI models grow into as they get bigger, not something every model has. ### What this does not mean It does not mean big AIs have feelings, self-awareness, or a real personality the way a person does. “Has five traits” here is a statement about whether the math separates into five dimensions, not about any inner life. It does not mean the tests are useless: they work fine for comparing models. And it does not mean small models are worse at their actual jobs (writing, coding, answering questions); they are just not well described by a five-trait personality profile. ### Abstract When a Big Five questionnaire is administered to a language model, does it measure five traits or one? Convergent validity (agreement among instruments that purport to measure the same construct) and discriminant validity (separation among instruments measuring different constructs) are the foundational criteria for any psychometric measure, yet they have not been tested across the population of deployed language models or against model scale. We administered four public-domain Big Five inventories of varying length (the 20-item Mini-IPIP, the 50-item IPIP Factor Markers, the 100-item Big Five Aspect Scales, and the 120-item IPIP-NEO) to 42 instruction-tuned models spanning roughly 0.36B to ~1T parameters, in stateless, schema-constrained single-item calls (290 items × 11 repetitions per model; 133,980 administrations), and analyzed the result as a multitrait-multimethod (MTMM) matrix with the model as the unit of analysis. Three findings emerge. First, convergent validity is strong: across models, different instruments agree on a model’s standing on the same trait (mean monotrait-heteromethod r = 0.82; 0.85 when restricted to the one instrument pair without verbatim item overlap), and at the population level the three full-length instruments are internally consistent (Cronbach’s α median 0.81; 14 of 20 instrument × domain cells ≥ 0.70), though the abbreviated Mini-IPIP is not (median α = 0.40). Second, discriminant validity is emergent with scale: small models (≤4B) barely separate the traits (convergent−heterotrait gap = 0.08; “everything correlates”), large local models (>4B) separate them more (gap = 0.22), and frontier-scale models clearly separate them (gap = 0.44, heterotrait r falling to 0.27); within the local models, trait differentiation rises with log₁₀ parameters (r = 0.44, p = 0.018). Third, the trait structure is a property of the population, not the individual model: treating a single model’s 11 repeated administrations as respondents yields α ≈ 0 despite real generation variance, so a model does not reproduce a stable internal trait structure across regenerations even though the population does. A full-precision (fp16) re-run of seven model families confirms these patterns are not a 4-bit quantization artifact (mean domain-score shift 0.10 on the 1–5 scale; population and within-model reliability unchanged). We conclude that Big Five questionnaires behave as valid instruments for measuring between-model differences, that abbreviated inventories should not be used for this purpose, and that the five-factor structure these instruments presuppose is something larger models possess and smaller models do not. --- ## Is LLM Personality an Artifact of Deployment? Psychometric Stability of Big Five Self-Reports Across Quantization Levels - URL: https://research.ideafields.institute/papers/llm-personality-quantization/ - PDF: https://research.ideafields.institute/papers/llm-personality-quantization/main.pdf · Markdown: https://research.ideafields.institute/papers/llm-personality-quantization/draft.md - Data & code (Zenodo, CC-BY-4.0): https://doi.org/10.5281/zenodo.20671762 - Published: June 2026 ### Plain-language headline Shrinking an AI changes its “personality” — and the personality tests don’t really work anyway. ### What we did When people run AI chatbots on their own computers, they almost never run the full-size original — it’s too big. They run a compressed copy, the same way an MP3 is a compressed copy of a studio recording. Compression comes in grades: light (nearly full quality) and heavy (about a quarter the size). Meanwhile, researchers have taken to giving AI models personality quizzes — the same kind psychologists give people — and publishing the results as “this AI’s personality.” Nobody had checked whether the compressed copy people actually use gives the same answers as the original. We checked: three AI models, each at three compression levels, answering the same 50-question personality quiz over and over — 28,350 answers in total, collected on an ordinary home computer. ### What we found Three things. First, heavy compression changes the answers — on about 1 in 4 questions, the compressed copy gave a flat-out different answer than its own original, even with every setting locked down. Light compression barely changed anything. Second, each AI warps in its own direction: compression made one model describe itself as more organized and curious, and another as less friendly and less calm. There’s no single rule — so there’s no way to correct for it after the fact. Third, and biggest: the quiz itself doesn’t really work on these AIs. In a human personality test, your answers to related questions hang together — if you say you love parties, you also say you start conversations. These models’ answers don’t hang together, not for any model at any compression level. Simply listing the multiple-choice options in reverse order changed the scores more than compression did. ### Why it matters When you see a claim that some AI “has” a personality — anxious, agreeable, whatever — the honest follow-up questions are: which copy of it, set up how, asked in what format? Our results say those details change the answer, sometimes by a lot. And for the small AI models people run at home, the deeper takeaway is that personality-quiz results shouldn’t be trusted at all yet: the measurements don’t meet the basic standards we’d demand before drawing conclusions about a person. ### What this does not mean It doesn’t mean AIs have feelings or inner lives — a quiz answer is a quiz answer. It doesn’t cover the big commercial systems (ChatGPT, Claude, Gemini); we only tested small models that run on home computers. And it doesn’t mean compressed models are worse at their jobs — writing, coding, answering questions — only that their personality-quiz answers change. ### Abstract Psychometric questionnaires are now routinely administered to large language models (LLMs), and a growing critical literature questions what such instruments measure when the respondent is a model. One deployment-side moderator has so far gone unexamined: quantization. Virtually every real-world local deployment of an open-weight model runs a quantized artifact (typically a 4- or 8-bit GGUF served by Ollama or llama.cpp), yet published “LLM personality” estimates are almost always obtained from full-precision checkpoints. We administered the 50-item IPIP Big Five Factor Markers to nine model variants — three open-weight instruction-tuned models (Llama 3.1 8B, Qwen 2.5 7B, Llama 3.2 3B), each at three quantization levels (q4_K_M, q8_0, fp16) — under three response-option presentation conditions, with one greedy and twenty independently seeded sampled repetitions per item: 28,350 stateless, JSON-schema-constrained chat calls on fixed consumer hardware. Three findings emerge. First, 4-bit quantization materially shifted domain scores relative to the same checkpoint at fp16 (6 of 15 baseline contrasts with bootstrap CIs excluding zero; shifts up to 0.42 scale points, |Cohen’s d| up to 3.5), in directions that were idiosyncratic to model family; 8-bit quantization was largely score-preserving (2 of 15; max |d| = 0.76). Under greedy decoding, q4 variants gave a different answer than their fp16 parent on 8–32% of items (q8: 2–12%). Second, internal consistency was inadequate everywhere: across all 135 variant × condition × domain cells, Cronbach’s alpha never reached the conventional 0.70 threshold (range −2.07 to 0.47, median −0.10; 56% of cells negative), with no systematic precision gradient — the instrument’s keyed structure failed to organize responses at every quantization level, including fp16. Third, response-option presentation alone (display order; letter vs. numeric labels) shifted scores more on average than quantization did (mean |δ| up to 0.21 vs. 0.17 scale points; |d| up to 4.2), at every precision. Together these results indicate that “the personality of model X” is underdetermined without the deployment configuration — and that for small open-weight models served in deployment-realistic conditions, questionnaire-based personality measurement fails basic psychometric standards regardless of precision. Quantization level, inference stack, sampler settings, and presentation format belong in the methods section of any LLM psychometrics study.