Cognitive bias in large language models: Cautious optimism meets anti-Panglossian meliorism

The recent success of large language models gives new urgency to the question of how model performance should be evaluated. In many tasks, models can be evaluated for the accuracy of their outputs. However, models can also be evaluated along other important dimensions. For example, we can assess models for the transparency or interpretability of their judgments (Creel 2020; Vredenburgh 2022). We can also assess models for the presence of problematic biases (Kelly 2023; Johnson 2020). Most work on biases in large language models focuses on a conception of bias closely tied to unfairness, especially as affecting marginalized social groups. However, recent work has alleged that large language models also show a number of classic cognitive biases familiar from work in the psychology of reasoning, behavioral economics, and judgment and decisionmaking (Dasgupta et al. 2022; Lin and Ng 2023; Jones and Steinhardt 2022). This development is exciting because it raises the possibility of using cognitive bias as a novel metric by which to evaluate the performance of large language models.