Abstract
This article investigates the ability of large language models (LLMs) to successfully answer questions about text when there is “semantic underspecification.” Although humans make substantial use of semantic underspecification, and seem to be able to tease out meaning, LLMs do not successfully respond to underspecified text with the same level of correctness. We found whereas humans were able to disambiguate underspecified text, LLMs required additional information specification to correctly answer queries. These findings have several implications.