Cognitive biases significantly affect human judgment in legal decision-making, impacting fairness and objectivity. We investigated whether large language models (LLMs) exhibit similar cognitive biases. Replicating the study of humans by Teichman et al. (2023), we examined the presence of outcome bias, anchoring bias, and anti-inference bias in seven advanced LLMs (GPT-4o, GPT-o3-mini, DeepSeek-R1, Claude-Sonnet-4, Llama-3.3-70B, Mistral-Large-2411 and Qwen-235B-A22B), also assessing whether the models’ reasoning capability and knowledge enhancement with legal information moderated these biases. LLMs systematically exhibited human-like cognitive biases: they judged negligence more harshly when outcomes were severe (outcome bias), recommended significantly longer sentences when exposed to irrelevant numerical anchors (anchoring bias), and showed a preference for direct over equally probative circumstantial evidence (anti-inference bias). The impact of reasoning capability was mixed: reasoning models often failed to mitigate outcome and anchoring biases, whereas they showed diminished biases specifically regarding anti-inference judgments. Additionally, knowledge enhancement influenced the presence of bias, with knowledge surprisingly amplifying anchoring and outcome bias while weakening anti-inference bias. These findings underscore that advanced reasoning abilities or knowledge enhancement do not fully eliminate cognitive biases in LLMs, highlighting the critical need for targeted bias mitigation before deploying AI to support legal decision-making.
