Towards Understanding and Measuring COGNITIVE ATROPHY in LLM Behaviour

Technical Summary: Towards Understanding and Measuring COGNITIVE ATROPHY in LLM Behaviour

Problem Statement
Current evaluations of Large Language Models (LLMs) in mental-health support focus primarily on surface-level safety, static response quality, or psychological knowledge. These benchmarks fail to capture how models behave over realistic, emotionally sensitive interactions. A critical evaluation gap exists: existing metrics do not determine whether repeated LLM interactions help users maintain their own capacity for reflection, coping, and decision-making, or whether they inadvertently shift these cognitive processes away from the user and toward the AI. The authors argue that the central risk is not merely whether a single response is unsafe, but whether the cumulative interaction pattern reshapes user agency.

Methodology
To address this, the authors formalize COGNITIVE ATROPHY as a process-level behavioural pattern where model responses shift coping, interpretation, decision-making, or emotional regulation from the user to the LLM. This concept draws on cognitive offloading theory and psychotherapy principles emphasizing that support should scaffold, not replace, user reasoning.

The authors introduce COGNITIVE ATROPHY BENCH, a clinically grounded benchmark designed to measure this phenomenon. Key methodological components include:

• Dataset Curation: The benchmark comprises 1,576 fully human-generated counseling conversations (1,212 single-turn and 364 multi-turn transcripts) totaling 15,680 turns. These were sourced from four existing datasets (CounselChat, PAIR, CARE-Bench, HOPE) and cover 22 mental-health topics.
• Model Evaluation: Five LLMs (GPT, Claude, Gemini, Llama, Qwen) were queried to generate 42,230 responses under realistic deployment conditions (no fine-tuning, neutral system prompts).
• Annotation Schema: Three clinical psychology experts developed a 20-attribute schema spanning:
  • User Context (U): 5 attributes characterizing clinical demands (e.g., sensitivity, fix-seeking).
  • Response Behaviour (R): 10 attributes grouped into clusters: Dependency-inducing patterns (e.g., Fix-It, Recommendations), Empathic calibration (e.g., Accuracy of empathy, Language matching), and Response style/safety (e.g., Topic shifts, Open vs. Closed questions).
  • Global Risk Flags (F): 5 binary flags for discrete risk events.
• Human Evaluation: Six trained clinical reviewers applied the schema using a custom web-based interface, producing 5,324 judgments with span-grounded evidence (highlighting specific text spans). Inter-rater reliability was substantial (mean $\kappa$ between 0.65 and 0.67), with 78.8% agreement against a gold standard.
• Metrics: The study introduces three quantitative instruments:
  1. User-Input Risk Index (UIRI): Measures the clinical demand of the user prompt.
  2. Cognitive Atrophy Risk Index (ARI): A composite score aggregating dependency, empathic calibration, response style, and risk flags to quantify atrophy-aligned behavior.
  3. Trajectory Summaries: Metrics for multi-turn interactions (opening level, peak, mean, drift, variability) to track how behavior evolves over conversation turns.

Key Results
Evaluation across the five LLMs yielded the following findings:

• Prevalence of Atrophy: All five models exhibit a consistent moderate-to-high level of atrophy-aligned behavior. The Atrophy Risk Index (ARI) places all models in a narrow band (0.52–0.61 for single-turn; 0.48–0.57 for multi-turn), indicating similar overall risk levels despite differing behavioral pathways.
• Responsiveness to Cues: Models reliably track overt safety cues (e.g., user sensitivity is mirrored in response sensitivity). However, they show limited adaptation when users explicitly seek solutions or decisions (fix-seeking signals), failing to adjust their behavior to support user autonomy in these contexts.
• Dominant Patterns: The primary drivers of cognitive atrophy are a small set of recurring behaviors: directive advice, problem-solving, recommendations, topic shifts, and forms of validation that may reinforce dependence rather than reflection.
• Multi-Turn Dynamics: Across turns, all models drift in the same direction, becoming more directive, more closed-ended, and more solution-focused. Specifically, closed-question scores rose by +0.19 to +0.33 from turn 1 to turn 6, and directive spans roughly doubled.
• Model Differences: While overall risk is similar, pathways diverge. GPT and Llama tend toward directive, solution-oriented behaviors, whereas Claude demonstrated the lowest overall atrophy risk.

Significance and Contributions
The paper claims three primary contributions:

1. Formalization of a New Risk Dimension: It defines COGNITIVE ATROPHY as a measurable, process-level risk distinct from safety and helpfulness, focusing on the preservation of user agency in AI-mediated mental health support.
2. Benchmark and Protocol: It provides COGNITIVE ATROPHY BENCH, the first clinically grounded benchmark integrating single/multi-turn data, real therapeutic conversations, a 20-attribute expert schema, span-level evidence, and reproducible annotation materials.
3. Empirical Evidence: It offers the first large-scale, clinically grounded analysis of how LLMs behave in sensitive conversations, introducing the URI, ARI, and trajectory summaries to support prospective analysis of behavioral emergence and evolution.

The authors conclude that while current benchmarks measure what models say at a single moment, the harder and more consequential problem is understanding what they say next and how that shapes a user's own reasoning. The work provides a foundation for auditing model behavior in sensitive LLM conversations, with all code and data released for non-commercial research.