Research Shows AI Language Models Exhibit Human-Like Cognitive Biases

Researchers have identified patterns in large language models that closely resemble well-known human cognitive biases, including anchoring, availability heuristic, and framing effects. Studies found that irrelevant numbers and the way information is presented can systematically shift AI-generated responses, with models like GPT-4, Claude 2, and Gemini Pro all showing anchoring tendencies. For example, when an unrelated number is introduced before an estimation task, the model's output tends to shift toward that figure. Unlike human cognitive biases, which stem from mental shortcuts in cognition, AI bias manifests as a pattern in model output rather than an internal psychological process. Experts stress that understanding these patterns is important, as they can affect the reliability and consistency of AI-generated judgments.