Paper ID: 2403.05152

Towards a Psychology of Machines: Large Language Models Predict Human Memory

Markus Huff, Elanur Ulakçı

Large language models (LLMs) are excelling across various tasks despite not being based on human cognition, prompting an investigation into their potential to offer insights into human cognitive mechanisms. This study examines ChatGPT's ability to predict human performance in a language-based memory task. Following theories of text comprehension, we hypothesized that recognizing ambiguous sentences is easier with relevant preceding context. Participants, including humans and ChatGPT, were given pairs of sentences: the second always a garden-path sentence, and the first providing either fitting or unfitting context. We measured their ratings of sentence relatedness and memorability. Results showed a strong alignment between ChatGPT's assessments and human memory performance. Sentences in the fitting context were rated as being more related and memorable by ChatGPT and were better remembered by humans, highlighting LLMs' potential to predict human performance and contribute to psychological theories.

Submitted: Mar 8, 2024