Do Large Language Models Show Negation Bias? A Replication of Beukeboom et al. (2010)

LLM bias research has largely focused on content-level associations, overlooking biases arising from linguistic form and pragmatic choice. One such phenomenon is the negation bias, where negated expressions are preferentially used for stereotype-inconsistent behavior and trigger systematic inferences. While LLMs’ handling of negation in semantic tasks is well studied, it is unclear whether they reproduce these pragmatic inferences in social contexts. We replicate Beukeboom et al. (2010) with Mistral-7B-Instruct-v0.2, varying the valence and polarity of trait descriptions. The model inferred more negative expectations from negated negative traits (e.g. not stupid) and more positive expectations from negated positive traits (e.g. not smart), and showed stronger situational but weaker dispositional attributions. These results mirror human negation-bias patterns, suggesting that instruction-tuned models can produce pragmatic-like bias in judgment tasks. The findings underscorethe need to extend LLM bias evaluation beyond content to linguistic form.