LLMs Struggle With Negation, but so do Humans - a Two-step Simulation Approach

While substantial research has addressed the challenges of negation processing in humans, the parallels between human cognitive models and large language models (LLMs) in this regard remain less explored. This paper investigates how GPT-4o processes negation, drawing comparisons to human negation processing, particularly the two-step model of negation. In two experiments using image-sentence pairs, we assess the model’s ability to handle affirmative and negated sentences. Our findings reveal that, like humans, GPT-4o struggles more with negation, exhibiting higher rates of incorrect inferences after negation, and a systematic bias toward completions associated with the negated state of affairs. This study highlights the qualitative similarities between AI and human processing of negation, offering insights into the limitations of current LLMs and suggesting future directions for improving their cognitive alignment with more complex linguistic constructs.