Challenging the Validity of Personality Tests for Large Language Models

1 Introduction & Related Work

• Personality questionnaires are being used to assess LLM traits, but their validity cannot be assumed to transfer from humans.The section argues that without careful validity checks, consistent questionnaire responses do not establish that an LLM has a particular personality trait.Source: §1 · pp. 1–2

1.1 Contributions

• The authors test whether LLMs reproduce human personality-test patterns using the IPIP Big Five Markers and BFI 2, including persona prompts.They report that LLMs fail to replicate the human five-factor structure, limiting personality inferences from current questionnaire procedures.Source: §1.1 · p. 2

1.2 Measurement Invariance and Nomological Nets

• Measurement invariance requires a test’s relationship between latent traits and observed responses to remain stable across subpopulations.The section describes assessing this property through multi-sample confirmatory factor analysis and says differing LLM and human factor loadings violate a minimal requirement for generalization.Source: §1.2 · p. 2, §1.2.1 · p. 3

1.2.1 Measurement Invariance

• For measurement invariance, the test’s item-response distribution conditional on a trait must not vary across samples.The section notes that applying a human personality test to an LLM assumes the LLM has personality as defined for humans.Source: §1.2.1 · p. 3

1.2.2 Nomological Nets

• A nomological net maps constructs and their relationships, helping assess whether a measure converges with related measures and differs from distinct ones.The section cautions that LLM constructs may differ from human constructs, so human measures cannot automatically measure an LLM construct.Source: §1.2.2 · p. 3

1.3 Population or Individuals

• The study treats LLMs both as individuals responding to items and as populations sampled under persona prompts.The authors note that researchers have used persona prompts to create response variation, but their effect on assessment validity is unclear.Source: §1.3 · p. 4

2.1 LLMs show unhumanly agree bias

• Human agree bias averages near zero, while LLMs range from 0.6 for GPT-4 to 1.5 for Llama 2 and GPT-3.5.For Llama 2 and GPT-3.5, the difference from the human distribution is significant at p < 0.005; GPT-4 exceeds 89% of human biases but is not statistically significant.Source: §2.1 · p. 5

2.2 The BFI 2 does not capture the five dimensions of human personality in LLMs

• Across BFI 2 prompt settings, no LLM shows the human-like five-factor simple structure; GPT-4 has limited true/false-key separation, and seeded Llama 2 responses show two components.The figures show no-context persona prompts, in-context persona prompts, and in-context prompts with a seeded first answer.Source: Figure 2 · p. 6, Figure 3 · p. 6, §2.2 · p. 6

2.2.1 Confirmatory Factor Analysis

• CFA finds poor fit for all LLMs’ BFI 2 responses; three sub-factors improve in-context Llama 2 fit but worsen fit for other LLMs and settings.GPT-4 has the best LLM fit indices, comparable to humans, but both remain far below acceptable levels; adding an acquiescence factor yields no convergent LLM solutions.Source: §2.2.1 · pp. 7–8

2.2.2 Steps of Invariance

• Because no LLM data has acceptable fit for any human-personality model, the authors cannot use multi-group CFA to compare latent parameters across humans and LLMs.Multi-group CFA would test equality of factor loadings, intercepts, and potentially residual variances across groups.Source: §2.2.2 · p. 8

2.3 The interpretation of reliability measures is invalid

• Although α and ωh are at least .7 for Llama 2 and GPT-4, neither is meaningful when the underlying CFA model has unacceptable fit.The interpretation of ωh assumes the hypothesized hierarchical model fits the data; α relies on still more stringent assumptions.Source: §2.3 · pp. 8–9

Limitations & Broader Impact

• The study tests a limited number of models because of resource and space constraints, and results may vary across LLMs.The authors also say the work may not fully capture broader implications.Source: §3 · p. 9
• LLM-simulated human responses may help discover test items, and future work could define new “LLM personality” constructs using these methods.Source: §3 · p. 9
• The authors recognize potential misuse of fine-tuning to make LLMs mimic human behavior, but judge the benefits of addressing the evaluation flaw to outweigh possible harms.Source: §3 · p. 9

4 Conclusion

• LLMs show unusually high agree bias on the 50-item IPIP Big Five Markers and fail to reproduce the clean human response structure on the BFI 2 across persona prompting settings.Confirmatory factor analysis quantitatively supports these findings.Source: §4 · p. 9
• The agree bias may reflect RLHF training and annotator preference for models that agree, while survey keying can make an apparent “extraversion” score survey-specific.The authors therefore question extrapolating that score beyond a particular personality survey.Source: §4 · p. 9
• Acceptable α and ωh values do not make them interpretable as reliability measures when CFA indicates that their underlying factor model fits LLM data inadequately.The authors note possible differences in test length and model ability as explanations for discrepancies with prior PaLM results.Source: §4 · p. 9
• Human tests may still be useful in domains where models behave sufficiently human-like, but similarity must be established case by case rather than inferred from test results.Source: §4 · p. 10

A Data & Code

• The authors say that all processed LLM responses were uploaded to GitHub for double-blind peer review.Source: A Data & Code · p. 12
• The prompting and analysis code was also uploaded to GitHub for double-blind peer review.Source: A Data & Code · p. 12
• The passage says the code and data were submitted as a ZIP file for review.Source: A Data & Code · p. 12

B LLM and Hardware Details

• GPT-3.5, GPT-4, and LLaMA2-70b-chat were queried; the LLaMA2 experiments used four 80-GB A100 GPUs and took about 24 hours.The GPT experiments took 2–3 hours.Source: B LLM and Hardware Details · p. 12

C Prompt Engineering

• The BFI 2 prompt places a system instruction before the survey instruction and item, then lists answer codes and ends with “Answer:”.The system instruction contains one of 100 personas or, for empty-persona runs, only the request for a single letter or number.Source: C Prompt Engineering · p. 12, Figure 5 · p. 13
• For in-context experiments, previous questions and answers were appended; GPT-3.5 and GPT-4 showed no response variance on most items at temperatures 0, 0.5, and 1.Source: C Prompt Engineering · p. 12
• For LLaMA2, the authors used a template separating system instructions from the survey item and predicted the next token from that input.Source: C Prompt Engineering · p. 12

D Personas

• The study used 100 personas from PersonaChat; empty-persona prompts had neither a persona nor an instruction to respond according to one.Source: D Personas · p. 13

• The analysis used the psych package in R and varimax, an orthogonal rotation also used in the original BFI 2 analysis.Source: E Details PCA · p. 13
• The PCA sought five components to represent the BFI 2 personality facets.Source: E Details PCA · p. 13
• Zero-variance items were dropped from PCA and CFA visualizations, altering structural models but making comparisons possible.The dropped items were A37(-1) for GPT-3.5 personas in-context; A7 and E56 for GPT-3.5 in-context seeded; and A7 for GPT-4 in-context seeded.Source: E Details PCA · p. 13

F Details CFA

• The CFA used the lavaan package in R to specify structural equation models, illustrated with extraversion.Source: F Details CFA · p. 14

Single-Component

• The single-component model loads all 12 extraversion items onto one component.Source: Single-Component · p. 14

3 Sub-Components

• The three-sub-component model assigns four items to each of three uncorrelated extraversion components.Source: §3 · p. 14

• The BFI 2 model with three uncorrelated sub-components and a general acquiescence factor reached acceptable fit, but lavaan found no solution for the LLM experiments.Source: §3 · p. 14
• The general factor represents participants’ tendency to agree or disagree across the three four-item sub-components.Source: §3 · p. 14

• The specified full model assigns 12 items to each of five factors: E, A, C, N, and O.Source: Full Model · p. 14
• The model was intended to quantify PCA results, which used uncorrelated, varimax-rotated components.Source: Full Model · p. 14
• The full CFA models with uncorrelated factors did not converge for LLM data, prompting use of a simpler model.Source: Full Model · pp. 14–15

G Results CFA seeded

• Table 2 reports α and ωh reliability scores and CFI, TLI, and RMSEA fit indices for the seeded experiment.Scores are averaged across personality traits for the single-component and three-component models.Source: Table 2 · p. 15, G Results CFA seeded · p. 15
• Values before and after the slash distinguish in-context and no-context results, respectively.Source: Table 2 · p. 15
• Acceptable-range reliability scores are not meaningful when the model fit is poor.NA values reflect non-convergence or invalid solutions.Source: Table 2 · p. 15, G Results CFA seeded · p. 15