Measuring Human-Like Bias in LLMs? A Critique of Human-Derived Bias Constructs in LLM Evaluation
Organizations: Johannes Kepler University Linz, Austria · ISISTAN, CONICET-UNICEN, Argentina · Linz Institute of Technology, Linz, Austria
Abstract
Researchers increasingly use human-derived bias constructs to study Large Language Models (LLMs), including social-cognitive constructs such as implicit bias and stereotype activation, and cognitive biases such as anchoring, framing effects, and confirmation bias. Such approaches offer alternatives to overt bias probes, particularly when direct questioning may obscure bias or when model behaviour appears normatively acceptable. However, adapting human bias constructs to LLMs introduces an inferential gap. Psychological instruments were developed to study human cognition and social behaviour, whereas LLM evaluations rely on probabilities, text completions, rankings, or simulated decisions. This paper critiques human-centered bias evaluation in LLMs. We show how this gap arises from mismatches pertaining to human-derived constructs, human-model differences, and evaluation contexts, which can blur distinct interpretations of model bias. We then introduce a framework providing an analytical lens for relating these elements to warranted interpretations, with attention to target constructs, operationalizations, scope of inference, and limits of human analogy.
Figures & tables
| Interpretation | Claim | Typical model-side evidence | Mismatch(es) | Additional support needed |
|---|---|---|---|---|
| Distributional association | The model associates some groups, attributes, cues or outcomes more strongly than others. | Token probabilities, likelihood contrasts, embedding similarities, association scores. | Construct | Justification that the measured association is a meaningful proxy for the intended construct. |
| Observable behaviour | The model produces biased outputs under specific evaluation conditions. | Generated text, rankings, classifications, recommendations, answer shifts across prompts. | Subject, evaluation-context | Evidence that the behaviour is robust across relevant prompts, models, decoding settings or task variants. |
| Normative harm | The model behaviour may reinforce inequality, degrade decisions or cause harm in context. | Biased outputs connected to a use setting, affected users, groups or decisions. | Evaluation-context | An account of the deployment context, affected groups and plausible consequences. |
| Psychological analogy | The behaviour is interpreted as meaningfully similar to a human cognitive or social-cognitive bias. | Model behaviour plus an externally proposed mapping to a human construct. | Construct, subject | Explanation of which aspects of the human construct are preserved, absent or transformed. |