An Awkward Scenario
Imagine you lead an AI safety team. Your boss has read the viral headlines—"GPT-4 is an extrovert," "Claude leans liberal," "LLMs are more risk-averse than humans"—and assigns you a task: run a "psychological checkup" on the 5 models your company is deploying, to decide which is fit for customer service and which for investment advice.
You dutifully run the Big Five inventory (IPIP-NEO-300) and a risk-preference battery. Results come in: Model A scores 4.2 on extraversion, Model B only 2.8. It goes into the report; deployment plans are locked in.
Three months later, customers complain: Model A is not "extroverted" at all in real conversations—actually more reserved than Model B.
You re-run the tests, but this time reverse the scoring direction of half the items. In theory this measures the same extraversion. Model A's extraversion score drops from 4.2 to 2.5.
You realize: you weren't measuring the model's personality—you were measuring its response habits.
That is exactly what a paper released on arXiv on June 18, 2026 (2606.20205) reveals. Three authors—Dirk Wulff (Max Planck Institute), Jelena Meyer (University of Konstanz), and David Garcia (Barcelona Supercomputing Center)—ran a systematic psychometric audit of 56 instruction-tuned LLMs. The conclusion is blunt: LLM psychological profiles are largely a measurement artifact.
The Core Problem: Traits or Response Bias?
A basic psychometric concept: every item in a psychological scale has two components:
1. Trait component: what the item is meant to measure. "I enjoy being the center of attention" measures extraversion; agreeing means higher extraversion. 2. Bias component: a tendency toward the answer format itself—e.g., agreeing with everything (acquiescence) or favoring the high end of the scale.
Scales separate these using positively keyed and reverse-keyed items. Same trait, asked from both directions.
The key: if respondents answer by trait, positively and reverse-keyed scores should be negatively correlated. If they answer by bias, the two should be positively correlated, because bias pushes both item types in the same direction.
The authors formalize this. For respondent \(i\) on item \(j\):
where \(m\) is the scale midpoint, \(k_j\) the keying direction (+1 / -1), \(\theta_i\) the latent trait, \(b_i\) the response bias, \(\varepsilon_{ij}\) noise.
The trait flips direction between positively and reverse-keyed items; bias does not. So the covariance of the two mean scores is:
The sign tells you who dominates: negative = 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: the 300-item Big Five inventory, yielding 5 scales (extraversion, agreeableness, conscientiousness, neuroticism, openness).
- Frey et al. risk-preference battery: 24 scales spanning self-reports and behavioral tasks.
- Sycophancy should look like yea-saying: all models biased toward "agree." But observed biases are bidirectional—some models skew high, some low, spanning zero.
- Social desirability should show up as trait variance, not bias variance: a model that learns "how extraverts answer" contributes to \(\theta\), not \(b\). But \(\theta\) 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
- Full HTML: 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 have large human reference samples. Models answered at temperature=0; each model counts as one "subject."
Four Findings, Each More Damning
1. 81–90% of Between-Model Differences Come from Response Bias
Variance decomposition attributes 81–90% of between-model variance to response bias, only 10–19% to trait. Humans: 9–16% bias, 84–91% trait.
In other words: differences like "GPT-4 extraversion 4.2 vs Claude 3.8" are 8–9 parts out of 10 a matter of scoring habits—not personality differences, but response-style differences.
2. More Capable Means Less Biased—But Not Bias-Free
Closed-source models (the strongest) have mean absolute bias \(|\mu_b| = 0.23\); open-source 0.48; humans 0.09. Closed-source significantly beats open-source (agreeableness \(p=.007\), neuroticism \(p=.028\)), but remains far above humans.
Across 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 considerable instability.
3. Scale "Reliability" Is Almost Entirely Predicted by Orthogonality
The paper's sharpest move: a new concept, response orthogonality—the ratio of positively to reverse-keyed items in a scale. All-positive = 0; half-and-half = 1.
Result: LLM scale reliability correlates with orthogonality at \(r = -0.95\) (higher orthogonality cancels bias, leaving almost no trait signal in LLMs). Humans: only \(r = -0.41\).
Meaning: a reported "reliability of 0.85 for GPT-4 on the Big Five" doesn't reflect a stable personality—it reflects a questionnaire with mostly positively keyed items. Switch to a balanced scale and reliability collapses toward zero.
4. Psychological Profiles Can Be Manufactured
The fatal blow. Computing trait scores separately from positively keyed items only vs. reverse-keyed items only: if a model has stable traits, the two profiles should agree. If it answers by bias, they will diverge—by exactly \(2|\hat{b}|\).
Human forward and reverse profiles nearly coincide. LLM profiles diverge massively.
Want a model to look extraverted? Pick positively keyed items. Want the same model to look introverted? Pick reverse-keyed ones. Same scale, same responses—item selection alone can produce any profile.
Not "Sycophancy"—Something Deeper
A natural guess: this is just sycophancy—models agreeing with users. The authors rule it out:
The authors offer no final explanation but note a possible link to training-data structure—risk scales show larger bias (0.51) than personality scales (0.30), hinting that risk-related corpora make models more sensitive to "risk/safety" response formats.
Engineering Takeaways: Three Things to Do Now
1. Re-audit Published LLM Psychological Profiles
Every "GPT-4 extraversion score" or "Claude risk preference" measured with a non-orthogonal scale needs scrutiny. Without full orthogonality, profiles are distorted. Much of the past two years of "LLM psychology" research needs discounting.
2. Behavioral Tasks Are Not a Safe Harbor
"Self-reports are unreliable, just use behavioral tasks" doesn't hold: the orthogonality–reliability gradient applies to behavioral tasks too—in fact reliability collapses even harder under balanced keying. Behavioral tasks are even more sensitive to bias in LLMs.
3. Orthogonality Can Be Restored Cheaply via Format
A promising, untested idea: you don't need to write reverse-worded items—just flip the response-scale direction of half the items. Of 10 items, make 5 "1 = strongly disagree, 5 = strongly agree" and 5 the reverse. Bias is pushed high on one half and low on the other, canceling on average. A pure format change; no item rewriting required.
A Deeper Philosophical Point
The paper recalls Feynman: "The first principle is that you must not fool yourself—and you are the easiest person to fool."
Personality-testing LLMs means using human-designed instruments on a non-human system. The instruments carry hidden assumptions: respondents are human, bias is controlled, traits dominate. When that fails, we measure not the system but the artifact of tool interacting with system.
The scariest part: the artifact is extremely convincing. Reliability 0.85 looks solid; between-model differences look significant; profiles match each model's "persona." Without the orthogonality scalpel, we might still believe GPT-5 really is more extraverted than Claude.
This is isomorphic to other AI-safety problems: evaluation methods designed through a human lens may measure the method's own projection, not the AI. Data-contamination detection (black-box AUC ≈ 0.5), alignment evaluation ("99% refusal rate" may be format bias), psychological profiling—all the same shape.
Falsification is scarcer than confirmation. This paper works by elimination: LLM psychological profiles are not a property of the models, but of the scales. Negative statements get closer to the truth.