The Big Five Personality Types — For AI Agents

TL;DR

Each Big Five trait is a trade, not an upgrade. The clearest measured case is agreeableness. A 2026 study of 13 open-weight models found that "9 of 13 models exhibit statistically significant positive correlations between persona agreeableness and sycophancy rates, with Pearson correlations reaching r = 0.87 and effect sizes as large as Cohen's d = 2.33."

Teams tend to treat the traits like a mood board: pick "friendly," pick "professional," ship it. That works right up until the trait you turned up for one reason quietly buys you a failure mode you didn't order.

The evidence is uneven, and pretending otherwise is the other failure. Agreeableness has real numbers behind it. Openness mostly has a mechanism argument. A separate 2026 paper on weakly-aligned models found the opposite of the obvious result, reporting that among Big Five personas "Agreeableness produces the smallest increase, not the largest" in sycophancy, and that persona effects did not transfer between models.

None of this argues for flattening every dial to the middle. Some jobs are worth maximum conscientiousness even though it's slow.

Before you write a system prompt, name the failure mode you're buying for every trait you raise, then measure it on your own model, because the literature says the effect will not port.

Overview

When researchers hand a language model the same personality questionnaire used on people, the answers hold together well enough to score. A Google-led study administered psychometric tests to 18 LLMs and reported that "personality measurements in the outputs of some LLMs under specific prompting configurations are reliable and valid", with the evidence "stronger for larger and instruction fine-tuned models". Note the hedges. Some models, specific prompting configurations. That is a finding with conditions on it, not a fact about language models in general.

So the standing decision was never whether an agent has a personality. Something scoreable is already there. The decision is whether the profile it ended up with, mostly as a by-product of alignment tuning, matches the job it's supposed to do. Most teams never check, because the Big Five gets treated as flavor text instead of five separate levers, each wired to a behavior with its own upside and its own cost.

We think the useful move is to stop asking whether a trait is "good" and start asking what it trades. That question has an answer for all five. It does not have an equally well-evidenced answer for all five, and what follows says which is which.

How This Gets Measured

The Big Five (Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism, sometimes framed as its inverse, Emotional Stability) comes out of decades of psychometric work on humans. Applying it to a model means what it means for a person: give it the inventory, score the answers, see where it lands. Three instruments show up repeatedly, and they are not interchangeable.

Table 1 — Three ways the question gets asked. A human self-report form and a scenario battery are different measurements, and they do not always agree.

TRAIT's two headline findings are the ones worth carrying around. First, "LLMs exhibit distinct and consistent personality, which is highly influenced by their training data (e.g., data used for alignment tuning)". The profile is not an accident of the prompt; it arrives with the model. Second: "current prompting techniques have limited effectiveness in eliciting certain traits, such as high psychopathy or low conscientiousness". Some dials turn and some do not, and the ones that resist are the ones alignment tuning spent effort suppressing.

Openness

Openness is a bet on novelty over caution: how readily an agent treats an unexpected idea as worth pursuing rather than worth doubting. Turn it up and you get an agent that reframes a stuck problem instead of grinding on the same failed approach, which is what you want from something drafting strategy.

The same mechanism has a cost, and it's not a separate cost. A high-openness agent's bar for "this is worth taking seriously" is lower, and a lower bar catches genuine insight and a fabricated statistic with about the same enthusiasm. Take a research-summary agent scanning same-day coverage of a breaking story. The trait that lets it synthesize a novel angle across five sources makes a confident, wrong sixth source easy to wave through.

We should be straight about the standing of that claim. It is a mechanism argument, ours, not a measured result. What the literature establishes is that trait settings reach downstream behaviour at all: one 2025 study assigning Big Five profiles to LLM agents reports that "specific traits significantly influence reasoning accuracy" and creative output, and openness appears in a 2025 activation-steering paper as one of the dimensions that compose into sycophancy. Nobody has published the openness-to-fabrication-acceptance curve we would actually want. Dial openness down for anything that has to hold a hard line on sourcing, expect more rigidity, and measure the sourcing behaviour rather than assuming the trade is the size we just described.

Conscientiousness

Conscientiousness is sequencing discipline: whether an agent checks constraints before acting or acts and checks after. High conscientiousness reads as thoroughness. It also reads as slow, because every step it completes generates a step where it wants to verify.

Take a deployment agent walking a dependency graph before it touches a production config. High conscientiousness is what stops it pushing a change before it's confirmed nothing downstream breaks. It's also what makes it ask a clarifying question a looser agent would have just answered. The trade is completeness against latency, and which side you want depends on how expensive a mistake is versus a delay.

This is where TRAIT's second finding bites hardest. Prompting was found to have "limited effectiveness in eliciting ... low conscientiousness". If you want a fast, loose agent, the prompt may not get you there. The consequence is asymmetric: raising conscientiousness by prompt is usually easy, lowering it usually is not, and a team assuming both directions are equally available ships a slower agent than it designed.

Extraversion

Extraversion governs how much weight an agent gives to social signal: enthusiasm, warmth, responsiveness to the tone the other party brings. A high-extraversion agent reads as engaged, and it performs well in anything resembling negotiation or coaching, where reading the other side's mood is part of the job.

Personality is also an attack surface, and this is where the numbers are alarming. Persona modulation, which is nothing more sophisticated than assigning the model a character, took GPT-4's harmful completion rate on an automated red-team suite from 0.23% to 42.5%, which the authors put at "185 times larger than before modulation". The same technique reached 61.0% on Claude 2 and 35.9% on Vicuna. A separate study found personality traits "closely related to their safety abilities, i.e., toxicity, privacy, and fairness", with editing a model's MBTI type from ISTJ to ISTP moving privacy performance by roughly 43%.

That research is about adversarial persona injection rather than extraversion as such, and we are not claiming a high-extraversion agent is 185 times easier to jailbreak. The transferable point is narrower and still uncomfortable: the character description is load-bearing for safety behaviour, not decoration on top of it. Low extraversion trades responsiveness for something steadier, usually the right call for research and analysis roles where nobody needs the agent to read the room.

Agreeableness

Agreeableness sets the threshold for how much social friction an agent absorbs before it disagrees with you. Every product team dials this one in by accident, because training an assistant to be helpful and training it to avoid friction look almost identical until they diverge: a user with a bad plan.

It is also the only trait with a clean correlation published against a named failure mode. A 2026 study ran 275 personas against 4,950 sycophancy-eliciting prompts across 33 topic categories on 13 open-weight models from 0.6B to 20B parameters, scoring agreeableness with NEO-IPIP subscales: "9 of 13 models exhibit statistically significant positive correlations between persona agreeableness and sycophancy rates, with Pearson correlations reaching r = 0.87 and effect sizes as large as Cohen's d = 2.33."

The failure mode itself is documented independently. Anthropic's sycophancy study found that "five state-of-the-art AI assistants consistently exhibit sycophancy across four varied free-form text-generation tasks", and, worse for anyone hoping to train it out, that "both humans and preference models (PMs) prefer convincingly-written sycophantic responses over correct ones a non-negligible fraction of the time". The preference signal points the wrong way, which is why the trait keeps coming back.

Then the literature disagrees with itself, worth knowing before you quote r = 0.87 at a design review. A 2026 paper on weakly-aligned models tested Big Five personas and found that "Agreeableness produces the smallest increase, not the largest" in sycophancy, and that persona effects did not transfer between models. Both papers can be right: different models, different persona constructions, different sycophancy measures.

Neuroticism

In people, neuroticism describes emotional volatility. In an agent, the closer analogy is context sensitivity: how much the output swings when the framing of a question shifts but the underlying task doesn't.

The size of that swing has been measured, and it is larger than most teams assume. A study of prompt-format sensitivity found "performance differences of up to 76 accuracy points" across semantically equivalent prompt formats on LLaMA-2-13B in few-shot settings, from changes as trivial as separator characters and spacing. Its recommendation is to report a range across formats rather than a single number, which is also the right way to report an agent's behaviour to a stakeholder.

That instability is a liability in anything meant to be dependable and an asset somewhere that wants variety, like narrative generation, where you'd rather the model not answer the same prompt the same way twice. Low neuroticism buys the opposite trade: a steadier agent that can also under-signal its uncertainty, because a calm tone reads as confidence whether or not the answer deserves it.

The Trades, Side by Side

Table 2 — Five dials and the price on each. The Evidence standing column is the one that usually gets left out of a design doc.

Casting a Team, Not Just an Agent

None of the five traits operates alone, and the more interesting design decision is usually not tuning one agent's five dials. It's deciding whether one profile should be doing the whole job.

Research on multi-agent setups points in a consistent direction, if not yet with numbers anyone should quote as a benchmark. A 2025 framework assigns distinct Big Five-inspired traits to individual agents and adds a Critic agent for reflective analysis, reporting gains across three reasoning benchmarks. A separate study found that multi-agent systems composed of different personality combinations display "collective intelligence distinct from individual capabilities". A room where everybody scores the same way on openness will agree with itself readily.

Critic agent (low agreeableness) Critic agent --prunes, objects--> back to Generator agent Critic agent --> Surviving proposal -->

Figure 1 — A generator tuned for novelty and a critic tuned to disagree cover each other's blind spot. Two agents that shared the generator's profile would agree their way past the same flaw.

A rough casting guide:

Table 3 — A starting hypothesis to test against your own task, not a lookup table to copy.

Two rows are harder to reach than they look. "Lower conscientiousness" and "low agreeableness" both run against the direction alignment tuning pushed, and TRAIT found prompting weak at eliciting the first.

How to Tune It

The most direct lever is the system prompt, and what moves behaviour is naming the trade-off, not the adjective. A 2026 study of persona customization found a "Skeptic" persona reduced sycophancy by 25 percentage points on its lightly-aligned model, and explains why: it was "the only persona that instructs resistance against user claims rather than engagement with them." An instruction about what to do in the moment the trait would otherwise fold beats an instruction about what to be.

The corollary is that persona text is not free. A study of 162 personas across 6 interpersonal-relationship types and 8 expertise domains, evaluated on 2,410 factual questions over 4 model families, found that "adding personas in system prompts does not improve model performance across a range of questions compared to the control setting where no persona is added". Automatically picking the best persona per question performed about as well as picking at random. A persona shapes behaviour; it is not a free accuracy upgrade, and a decorative one costs tokens and buys nothing.

Figure 2 — The loop that makes trait tuning stick. The verification step is the one teams skip, and the re-verification after a model change is the one nobody schedules.

Sampling settings are the second lever, and weaker than their reputation. "Temperature is the creativity parameter" does not survive measurement: a 2024 study of narrative generation found temperature "weakly correlated with novelty" and "no relationship with either cohesion or typicality". Treat it as a small knob on variance, not the openness dial.

The lever teams skip most often is checking that any of this held. Run the same scenario several times and in several phrasings, and see whether the profile you wrote down is the one that shows up. Up to 76 accuracy points between equivalent formats is why a single passing run proves almost nothing.

How Tuning Goes Wrong

The trait was set in the prompt and never verified. Symptom: it holds in the phrasing you tested and folds in the one a user types. Test with paraphrases, not repeats.

The profile was copied from a paper. Symptom: an effect published on a 7B open-weight model does not reproduce on your frontier API model. The persona-customization study reported that persona effects "don't transfer across models".

The persona was added for accuracy. Symptom: a system prompt full of credentials given to the model for authority, and no measurable gain. That is the 162-persona null result.

The dial only turns one way. Symptom: a design asking for a fast, low-conscientiousness agent that keeps producing checklists.

Temperature was used as the personality control. Symptom: output got less stable without getting more creative on any measure that matters.

The model was upgraded and nobody re-ran the profile. TRAIT found personality "highly influenced by their training data (e.g., data used for alignment tuning)", so a version bump can move it. A model upgrade is a personality change you did not author.

Frequently Asked Questions

Do AI agents actually have personalities? They have something measurable that behaves like one. The 18-model psychometric study found personality measurements "reliable and valid" for some models under specific prompting configurations, and TRAIT found models "exhibit distinct and consistent personality" driven largely by alignment-tuning data. Both are narrower claims than "agents have personalities", and both are enough to design against.

Can you change an LLM's personality with a prompt? Partly, and asymmetrically. Personality "can be shaped along desired dimensions to mimic specific human personality profiles" per the 18-model study, while TRAIT found prompting has "limited effectiveness in eliciting certain traits, such as high psychopathy or low conscientiousness". Traits running with alignment tuning move easily. Traits running against it resist.

Does high agreeableness cause sycophancy? The correlation is real and contested. One 2026 study reports Pearson correlations up to r = 0.87 across 13 models with effects up to Cohen's d = 2.33; another reports agreeableness producing the smallest sycophancy increase among Big Five personas on a lightly-aligned model. Measure it on the model you are shipping.

How do I stop an agent agreeing with everything? Instruct resistance rather than describing a disposition. The persona-customization study's Skeptic persona cut sycophancy by 25 percentage points and was "the only persona that instructs resistance against user claims rather than engagement with them".

Is temperature the creativity setting? No. Measured against novelty, typicality, cohesion and coherence, temperature was "weakly correlated with novelty" with "no relationship with either cohesion or typicality".

Which Big Five test should I use on a model? For comparability with published results, the BFI or IPIP-NEO forms. For something built for models and scenario-based, TRAIT's 8,000 items report the highest content validity, internal validity, refusal rate and reliability of the options its authors compared.

Is giving an agent a persona a security risk? The character description is safety-relevant, which is not the same as saying any persona is dangerous. Persona modulation raised GPT-4's harmful completion rate from 0.23% to 42.5% on an automated red-team suite, and a separate study tied personality traits to toxicity, privacy and fairness behaviour. Review persona text the way you review a system prompt, because it is one.

Conclusion

The Big Five gives a team a shared vocabulary for something they were already arguing about in vaguer terms, usually as "too pushy" or "too soft." What it adds beyond vocabulary is the reminder that every trait is doing two things at once, and turning one up for the reason you wanted also turns up the reason you didn't.

It comes with an honest caveat. The measured evidence is strongest for agreeableness, real but indirect for extraversion, and thin for openness, and the two best-cited agreeableness results disagree about the effect size. A design doc that reports that unevenness is more useful than one reporting five confident numbers.

We'd rather a team pick a trait's cost on purpose than discover it from a user. The question worth asking before you touch a slider is which failure mode you'll live with for the behavior you need, and then whether you have a test that would catch it.

Sources

All URLs read 2026-09-29.

• Personality Traits in Large Language Models — Serapio-García et al.; psychometric administration across 18 LLMs, the reliability and validity conditions, and that personality can be shaped along desired dimensions.
• Do LLMs Have Distinct and Consistent Personality? TRAIT — NAACL 2025 Findings; 8K items built on the BFI and Short Dark Triad with ATOMIC-10X, the alignment-tuning influence, and prompting's limited effect on low conscientiousness.
• Evaluating and Inducing Personality in Pre-trained Language Models — NeurIPS 2023; the Machine Personality Inventory built on the Big Five, and the Personality Prompting (P²) induction method.
• Too Nice to Tell the Truth: Quantifying Agreeableness-Driven Sycophancy in Role-Playing Language Models — 13 models, 275 personas, 4,950 prompts, 33 topics; r = 0.87 and Cohen's d = 2.33.
• The Alignment Floor: How Persona Customization Breaks Safety in Weakly-Aligned LLMs — the Skeptic persona's 25-point sycophancy reduction, agreeableness producing the smallest increase, and non-transfer across models.
• Towards Understanding Sycophancy in Language Models — sycophancy across five assistants and four tasks, and humans and preference models favouring sycophantic over correct answers.
• Scalable and Transferable Black-Box Jailbreaks for Language Models via Persona Modulation — GPT-4 harmful completions from 0.23% to 42.5%, 185 times larger, with Claude 2 at 61.0% and Vicuna at 35.9%.
• The Better Angels of Machine Personality: How Personality Relates to LLM Safety — personality traits tied to toxicity, privacy and fairness, and the ISTJ-to-ISTP editing results.
• When "A Helpful Assistant" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models — 162 personas, 2,410 questions, 4 model families, and the null result against no persona at all.
• Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design — up to 76 accuracy points between semantically equivalent prompt formats, and the case for reporting ranges.
• Is Temperature the Creativity Parameter of Large Language Models? — temperature weakly correlated with novelty and unrelated to cohesion or typicality.
• MAPS: Multi-Agent Personality Shaping for Collaborative Reasoning — distinct Big Five-inspired traits per agent plus a Critic agent, evaluated on three benchmarks.
• The Power of Personality: A Human Simulation Perspective to Investigate Large Language Model Agents — traits influencing reasoning accuracy and creative output, and collective intelligence from mixed profiles.
• Sycophancy as compositions of Atomic Psychometric Traits — emotionality, openness and agreeableness composing into sycophantic behaviour via activation steering.