Critical Confabulation: Can LLMs Hallucinate for Social Good?

The study titled 'Critical Confabulation: Can LLMs Hallucinate for Social Good?' introduces the idea of using LLMs to create narratives that fill historical gaps left by social and political inequalities. This approach, termed critical confabulation, aims to reconstruct the stories of 'hidden figures' in history. Through an open-ended narrative cloze task, researchers evaluated the OLMo-2 family of models alongside other baselines, finding that LLMs possess foundational narrative understanding capabilities. The findings suggest that carefully bounded hallucinations can support knowledge production without compromising historical fidelity, highlighting the potential for LLMs to contribute positively to social discourse and historical representation.