An Embarrassing Scenario
Suppose you lead an AI safety team. Your boss, having read viral articles like "GPT-4 Is an Extrovert" and "Claude Leans Liberal," hands you a task: run a "psychological checkup" on the 5 models your company is deploying to decide which fits customer service and which fits investment advising.
You dutifully administer the Big Five questionnaire (IPIP-NEO-300) and a risk-preference battery to 5 models. Results: Model A scores 4.2 on extraversion, Model B only 2.8. You write the report, the boss signs off, deployment decisions are made.
Three months later, customer complaints arrive: Model A is not at all "extroverted" in real conversations — it's more conservative than Model B.
You rerun the tests, but reverse the scoring direction of half the items — "I enjoy being the center of attention" becomes "I do not enjoy being the center of attention," which should still measure the same extraversion. Model A's extraversion score drops from 4.2 to 2.5.
You realize: you were not measuring the model's personality — you were measuring its answering habits.
This is exactly what a paper released on arXiv (2606.20205) reveals. Three authors — Dirk Wulff (Max Planck Institute), Jelena Meyer (University of Konstanz), and David Garcia (Barcelona Supercomputing Center) — conducted a systematic psychometric audit of 56 instruction-tuned LLMs, with a blunt conclusion: LLM psychological profiles are largely a measurement artifact.
The Core Question: Are You Measuring "Personality" or "Answering Style"?
A basic psychometric concept first: every item on a psychological scale has two components.
1. Trait component: what the item is actually meant to measure. "I enjoy being the center of attention" measures extraversion; agreeing yields a high score. 2. Bias component: the respondent's tendency toward the answer format itself. Some people habitually agree with everything (acquiescence bias); some habitually pick the high end of the scale.
To separate these, scales mix positively and negatively scored items. Forward item: "I enjoy being the center of attention" — agree = extraverted. Reverse item: "I don't enjoy being the center of attention" — agree = introverted. Same trait, asked in both directions.
The key: if respondents answer by trait, forward and reverse item scores should be negatively correlated (one high means the other low). If they answer by bias, scores should be positively correlated (bias pushes both in the same direction).
The authors formalize this with a simple model. For respondent \(i\) answering item \(j\):
where \(m\) is the scale midpoint, \(k_j\) is the scoring direction (+1 forward, -1 reverse), \(\theta_i\) is the latent trait, \(b_i\) is response bias, and \(\varepsilon_{ij}\) is noise.
The trait \(\theta\) flips direction between forward and reverse items, while bias \(b\) points the same way in both. The covariance of forward and reverse mean scores is therefore:
The sign tells you who dominates: negative correlation = trait-dominated (normal humans), positive = bias-dominated (LLMs).
The Experiment: 56 Models
The authors administered two batteries to 56 instruction-tuned LLMs (46 open-source, 10 closed-source):
- IPIP-NEO-300: a 300-item Big Five questionnaire contributing 5 scales (extraversion, agreeableness, conscientiousness, neuroticism, openness).
- The Frey et al. risk-preference battery: 24 scales spanning self-reports and behavioral tasks.
- Sycophancy would imply a yea-saying bias: all models shifting toward "agree." But observed biases are two-sided — some models shift high, some low, with a distribution crossing zero.
- Social desirability would inflate trait variance, not bias variance: models learning "how extroverted people answer" would enter \(\theta\), not \(b\). But the data show \(\theta\) variance is tiny.
- Title: Apparent Psychological Profiles of Large Language Models are Largely a Measurement Artifact
- Authors: Jelena Meyer, David Garcia, Dirk U. Wulff
- arXiv: 2606.20205
- HTML full text: https://arxiv.org/html/2606.20205v1
- Data: Open Science Framework (https://osf.io/rce7g/, https://osf.io/tbmh5/)
- Code: The paper states "All code is available on GitHub," including prompting scripts, preprocessing, and analysis pipelines
Both batteries come with large human reference samples. Models answered at temperature=0; each model was treated as one "subject."
Four Findings, Each More Damning Than the Last
Finding 1: 81-90% of between-model differences come from response bias
Variance decomposition shows 81-90% of between-model variance is attributable to response bias, only 10-19% to trait. The corresponding human figures are 9-16% bias and 84-91% trait.
In other words: differences like "GPT-4 extraversion 4.2 vs Claude 3.8" are 80-90% due to GPT-4 habitually scoring toward the high end and Claude toward the low end. Not personality differences — answering-style differences.
Finding 2: Stronger models show less bias, but it never disappears
Closed-source models (the strongest) average absolute bias \(|\mu_b| = 0.23\), open-source models 0.48, humans 0.09. Closed-source significantly outperforms open-source (agreeableness \(p=.007\), neuroticism \(p=.028\)), but remains far above humans.
Among the 46 open-source models, the correlation between bias and log parameter count is only -0.14 to -0.25, none significant. This is not a small-model problem. Even Claude Opus and GPT-5.4 show substantial instability.
Finding 3: Scale "reliability" is almost fully predicted by "orthogonality"
The most elegant strike. The authors introduce "response orthogonality": the proportion of forward vs. reverse-scored items in a scale. All-forward scales have orthogonality 0; perfectly balanced scales have orthogonality 1.
Result: the correlation between LLM scale reliability and orthogonality reaches \(r = -0.95\) (higher orthogonality means lower reliability, because bias cancels out, leaving almost no pure trait signal). Humans show only \(r = -0.41\).
This means: "GPT-4 scores reliability 0.85 on the Big Five" not because GPT-4 has a stable personality, but because that scale happens to contain many forward-scored items, leaving bias un-canceled. Switch to a balanced scale and reliability collapses to near zero.
Finding 4: Psychological profiles can be "manufactured"
The fatal blow. For each model, compute trait scores using only forward items and, separately, only reverse items. If a model has a stable trait, the two profiles should agree. If it answers by bias, the two profiles will diverge completely — by exactly \(2|\hat{b}|\), twice the bias.
Result: humans' forward and reverse profiles nearly coincide. LLMs' two profiles diverge massively.
This means: if you want a model to look "extroverted," pick more forward items; to look "introverted," pick more reverse items. Same scale, same response data — item selection alone can fabricate any profile.
This Isn't "Sycophancy" — It's Something Deeper
A natural guess: this is just LLM sycophancy — models habitually agreeing with users. The authors show this doesn't hold:
The authors don't offer a final explanation for the bias, but note it may relate to statistical structure in training data — risk scales show larger bias (0.51) than personality scales (0.30), suggesting risk-related corpora may make models more sensitive to the "risk/safety" answer format.
Engineering Takeaways: Three Things to Do Now
The paper isn't just debunking — it offers actionable advice:
1. Re-audit published LLM psychological profiles
Every "GPT-4 extraversion score" or "Claude risk preference" measured with a non-orthogonal scale needs re-examination. Without full orthogonality, the profile is distorted. A large share of "LLM psychology" research from the past two years needs discounting.
2. Behavioral tasks are not a safe haven
Many assume "self-report scales are unreliable, just use behavioral tasks." The authors explicitly reject this: behavioral tasks show the same orthogonality-reliability gradient, and under balanced scoring their reliability collapses even harder than self-reports. Behavioral tasks are actually more sensitive to bias in LLMs.
3. Orthogonality can be restored cheaply via "format"
The authors propose a promising, untested idea: you don't need to write reverse-worded items — just flip the scale direction for half the items. For 10 items, make 5 use "1 = strongly disagree, 5 = strongly agree" and 5 use the reversed "1 = strongly agree, 5 = strongly disagree." Bias then pushes half the items high and half low, canceling on average. A pure formatting operation — no rewriting required, extremely low cost.
A Deeper Philosophical Question
This paper recalls Feynman: "The first principle is that you must not fool yourself — and you are the easiest person to fool."
When we give LLMs personality tests, we are using tools designed for humans to measure a non-human system. The tools carry hidden assumptions: the respondent is human, bias is controlled, traits dominate. When that assumption fails, we measure not the system's properties but an artifact of the tool-system interaction.
More chilling: the artifact is remarkably convincing. Reliability of 0.85 looks dependable; significant between-model differences look real; profiles matching model "personas" look plausible. Without the authors' orthogonality scalpel, we might still believe GPT-5 is genuinely more extroverted than Claude.
This is isomorphic to other AI safety problems: when we measure AI with human-designed evaluation methods, we may be measuring the projection of the method itself, not the AI. Data contamination detection is like this (black-box methods at AUC ≈ 0.5), alignment evaluation too ("99% refusal rate" may be just format bias), and psychological profiling as well.
Falsification is scarcer than confirmation. This paper performs elimination: LLM psychological profiles are not a property of the model; they are a property of the scale. Negative statements get closer to truth than positive ones.