Working Paper
Recent studies show that large language models (LLMs) can improve the generation and evaluation of ideas. Yet strategic decision-making hinges on a prior step: constructing a problem representation that defines which classes of solutions are feasible. In a randomized controlled trial with 305 MBA students, the authors test whether providing LLM support during problem formulation—rather than only during ideation and evaluation—changes the composition of strategic alternatives considered. Consistent with prior findings, the authors find that LLM assistance increases the number of alternatives generated. However, when LLMs are introduced during problem formulation, they decrease strategic focus, an effect not observed when LLMs assist only in ideation and evaluation. Using an abductive approach, the authors propose that when introduced during problem formulation, LLMs appear to shape how individuals construct the strategic problem, creating cognitive anchors that bind subsequent search. Their study advances their understanding of human-AI collaboration in strategic contexts, highlighting the importance of when and how LLMs—and external representations more broadly—are integrated in decision-making.
