Artificially Biased Intelligence: Does AI Think Like a Human Investor?

Abstract

We test whether large language models exhibit cognitive biases in financial decisionmaking using a prompt-pair experimental design across 48 models and eleven biases. LLMs display economically significant biases in how they process information (framing, anchoring) and respond to social and narrative cues (herding, authority, representativeness, availability). Model intelligence relates non-monotonically to bias: higher capability reduces susceptibility to framing and representativeness but increases sunk cost, loss aversion, and disposition-style asymmetries. Less capable models behave like momentum investors, overweighting recent returns and social signals; more capable models resemble disciplined value investors, emphasizing fundamentals. Context engineering and prompt preprocessing mitigate some biases but prove ineffective or counterproductive for others, implying that debiasing requires bias-specific validation rather than generic guardrails. These findings have direct implications for AI governance in investment workflows.

Keywords: Large Language Models, Cognitive Bias, Behavioral Finance, Framing Effect, Anchoring Bias, Loss Aversion, Disposition Effect, Artificial Intelligence, AI, LLM, Large Language Models, ChatGPT