Patchwork, Misattention, and Pandering: Theory and Evidence on (Generative) AI Biases in Strategy

Artificial intelligence (AI) is often presented as a remedy to human cognitive limitations such as bounded rationality and bias. However, we argue that (generative) AI systemsespecially large language models (LLMs)-exhibit their own systematic biases, grounded not in psychology but in their underlying architecture and training procedures. We identify three key sets of biases in LLMs that constrain their capacity to solve strategic problems: (1) the patchwork bias: the misuse or misapplication of heuristics (or "patches") and the corresponding lack of global coherence; (2) the misattention bias: the tendency to excessively focus on surface cues and biased consensus views at the expense of true root causes and contrarian views; and (3) the pandering bias: the prioritization of agreeable or engaging responses over accurate or original ones. We first discuss the source of these biases, which result directly from the architectural features of LLMs that generate their remarkable capacities. We then conduct and report on empirical demonstrations of each bias in the context of strategic problem solving. We discuss the limitations of existing solutions designed to overcome these biases. Finally, we conclude with a discussion of the implications for strategic problem solving.