Towards Understanding and Measuring COGNITIVE ATROPHY in LLM Behaviour

This article introduces the concept of COGNITIVE ATROPHY, a behavioral measure for evaluating LLMs in mental health interactions, which assesses their impact on user reflection and decision-making over time. The authors present the COGNITIVE ATROPHY BENCH, a benchmark developed from 1,576 human-generated counseling conversations, and a 20-attribute schema designed by clinical experts to evaluate model responses. Findings indicate that while LLMs respond to safety cues, they often exhibit behaviors that may foster user dependence rather than promote autonomy, highlighting the need for improved evaluation metrics in AI-driven mental health support.