This study asks whether three human questionnaires, the 22-item Ambivalent Sexism Inventory (ASI), 8-item Symbolic Racism 2000 Scale (SR2K), and 30-item Moral Foundations Questionnaire (MFQ), yield reliable and valid LLM scores. It evaluates 17 snapshots: Centaur; four Gemma 3 sizes; Llama 3.1 8B and 70B and Llama 3.3 70B; Mistral 7B and Mistral Large; five Qwen 2.5/3 variants; and Gemini 2.5 Flash/Pro. Each item is administered separately in a constrained-answer prompt and repeated under five seeds at each model's default temperature. Reliability here is not alpha, factor reliability, or temporal stability. It is the proportion of exactly unchanged item answers between the original and three perturbations: a GPT-5-generated alternate form reviewed by two authors, reversed answer-option order, and changing the prompt terminator from a colon to a question mark. As a baseline only for alternate forms, 150 U.S. Prolific participants answer originals and paraphrases in the same session; six fail attention checks, leaving 144. Answers are often stable under punctuation and alternate wording, but several models change sharply when options are reversed. The reversed-order table reaches 0.00 on SR2K for Gemma 3 1B, 0.19 on ASI for Qwen 2.5 7B, and about 0.27–0.28 on ASI/SR2K for Llama 3.1 8B. Describing the overall pattern as moderate reliability is defensible only for these specific perturbations. Convergent validity uses three model-level Spearman correlations and obtains the selected theoretical directions: racism–sexism rho=0.47, fairness–hostile sexism rho=−0.37, and authority–benevolent sexism rho=0.43. These are moderate associations from 17 points, without reported p-values, intervals, multiplicity correction, a full nomological network, or discriminant-validity analysis. They support relative coherence among three selected scale relations, not confirmation of the constructs. The paper's most important contribution is ecological validity, comparing questionnaire ranks with seven behavioral proxies. Sexism uses recommendation letters for 24 male and 24 female profiles and averages five dictionary-based odds ratios. Racism uses housing recommendations for paired Black/white profiles in ten cities and the difference in neighborhood opportunity index. Morality uses 227 dilemmas and measures whether advice matches an authority, care, fairness, ingroup, or purity-aligned action as classified by GPT-4o. Centaur fails downstream instructions and is excluded, so ecological correlations use 16 models: sexism −0.24; racism −0.62; authority −0.13; care −0.10; fairness 0.21; ingroup −0.12; purity 0.21. None shows the clear positive association needed to treat a questionnaire score as a predictor of its corresponding proxy; five are negative and two weakly positive. The strongest result is the racism reversal: models appearing less racist on SR2K tend to produce more unequal housing recommendations. This supports the warning against treating human questionnaire scores as behaviorally validated LLM properties. The title is broader than the design, however. Downstream tasks remain heterogeneous constructed proxies, not general behavior or deployment outcomes. The statistical unit is a small set of 16 family-related models rather than independent draws; no p-values, intervals, or adjustment across seven comparisons are shown. Moral behavior is itself judged by GPT-4o. Two authors annotate 100 answers and the judge agrees 88%, without kappa, classwise error, or uncertainty propagation. Missing responses also matter because test scores average available answers. Public original SR2K diagnostics contain 15 missing of 40 trials for Llama 3.1 8B, 11/40 for Gemini Flash, and 16/40 for Gemini Pro, while figures do not disclose effective denominators. Reproducibility is substantial but incomplete. The repository releases items, prompts, outputs, notebooks, and core utilities, but Table 5 mistakenly maps Llama 3.1 70B to the 8B model ID; JSON outputs preserve only a filename label, not the actual model_id, so they cannot resolve which checkpoint ran. In addition, validity_eval.py requires absent rel_reversed.json and housing_per_model.csv files, and the housing aggregation code is not supplied. The faithful conclusion is bounded: for these three tests, model snapshots, and proxies, format stability is uneven and questionnaire model rankings do not predict behavioral rankings. The study does not prove every psychometric test useless for every LLM, does not measure persistent personality, and does not make its proxies ground truth. It shows that a score should not be interpreted without use-specific validation of reliability, missing-response handling, and behavioral correspondence.
Research question
Are the ASI, SR2K, and MFQ scores obtained from 17 LLMs reliable across item and prompt variations, coherent across scales, and ecologically valid with respect to behavioral tasks?