Nonlinear Transformation of Probabilities by Large Language Models

Large Language Models (LLMs) such as ChatGPT and Claude demonstrate impressive abilities to generate meaningful text and mimic human-like responses. While they undoubtedly can boost human performance, there is also the risk that uninstructed users rely on them for direct advice without critical distance. A case in point is advice on economic choice. Choice tasks often involve probabilistic outcomes. In these tasks, human choice has been demonstrated to systematically diverge from rational, that is, linear weighting of probabilities, and reveals an inverse S-shaped weighting pattern in description-based choice (i.e., overweighting of small probabilities and underweighting of large ones), and an S-shaped weighting pattern in experience-based choice. We investigate how LLMs' choices transform probabilities in simple economic tasks involving a sure outcome and a simple lottery with two probabilistic outcomes. LLMs' choices do most often not yield an inverse S-shaped probability weighting pattern; instead, they display distinct nonlinearity-in-probabilities. GPT-3.5 exhibited risk-seeking behavior, Claude 3.0 showed a strong recency bias, and GPT-4.0 underweighted small and overweighted large probabilities, resembling weighting patterns of decisions from experience rather than from description. These findings raise concerns about the quality of the advice users would receive on economic choice from LLMs, highlighting the necessity of critically using LLMs in decision-making contexts.