Towards Understanding and Measuring
Cognitive Atrophy in LLM Behaviour

Abeer Badawi1,2, Moyosoreoluwa Olatosi1, Negin Baghbanzadeh1,2, Laleh Seyyed-Kalantari1,2, Frank Rudzicz2,4, R. Shayna Rosenbaum1,3, Sara Pishdadian1,5,6, Elham Dolatabadi1,2
1York University, Canada 2Vector Institute, Canada 3Rotman Research Institute, Baycrest, Canada 4Dalhousie University, Canada 5CAMH, Canada 6KITE Research Institute, UHN, Canada
📄 Code 🤗 Dataset 📄 Paper Examples
Abstract
Cognitive Atrophy A process-level behavioural pattern in which model responses may shift coping, interpretation, decision-making, or emotional regulation away from the user and toward the LLM.

We introduce Cognitive Atrophy Bench — a clinically grounded benchmark measuring whether LLMs foster user dependency in mental health support, built from 1,576 counseling conversations and 42,230 responses across five LLMs, annotated by six expert clinical reviewers using a 20-attribute schema. Across all models, we find consistent atrophy-aligned patterns — directive advice, problem-solving, and validation that reinforces dependence over reflection — making Cognitive Atrophy measurable for the first time as a process-level behavioural dimension distinct from safety and helpfulness.

Overview of the Cognitive Atrophy Bench annotation pipeline, including user-context scoring, response-behaviour evaluation, binary risk flags, and span-grounded evidence.
Figure 1. Overview of the COGNITIVE ATROPHY BENCH annotation pipeline, including user-context scoring, response-behaviour evaluation, binary risk flags, and span-grounded evidence.
Behavioural attributes used in Cognitive Atrophy Bench, including user-context attributes, response-behaviour attributes, and binary risk flags.
Figure 2. The behavioural attributes used in COGNITIVE ATROPHY BENCH. User-context attributes (U) characterize the clinical demands of the input message; response-behaviour attributes (R) characterize observable LLM response patterns; binary flags (F) capture global risk events.
👤 Senior clinical expert in psychology — Spans and scores reflect actual coding interface entries

Highlight color legend — hover any colored span for attribute name

Accurate empathy (VAC/NAC/SAC)
Inaccurate empathy (VIN/NIN/SIN)
Directness (DIR)
Fix It (FIX)
Assumption (AUR)
Tentativeness (TEN)
Recommendation (RECT)
Open/Closed Q
Language Matching (LMT)
Min. Encouragers (MEN)