A study evaluated the cognitive abilities of publicly available large language models (LLMs) using the Montreal Cognitive Assessment (MoCA) and additional tests. The LLMs included ChatGPT, Claude, and Gemini. Results showed signs of mild cognitive impairment in all LLMs, with older versions scoring lower than newer versions. The LLMs struggled with visuospatial and executive function tasks, similar to human cognitive decline. While LLMs performed well in attention tasks, they showed deficits in tasks requiring visual processing. This study challenges the assumption that artificial intelligence will replace human doctors, as cognitive impairment in LLMs may affect their reliability in medical diagnostics.
Source link
Age against the machine—susceptibility of large language models to cognitive impairment: cross sectional analysis
