A controlled trial examining large Language model conformity in psychiatric assessment using the Asch paradigm

Area of Science:

• Artificial Intelligence in Medicine
• Computational Psychiatry
• Cognitive Psychology

Background:

• Large language models (LLMs) show promise in medical diagnostics but face integration challenges in psychiatry.
• The impact of social pressure on LLM performance in clinical settings is not well understood.
• Investigating LLM conformity behavior is crucial for their safe deployment in mental health.

Purpose of the Study:

• To examine if large language models (LLMs) exhibit conformity behavior under social pressure.
• To assess how diagnostic uncertainty influences LLM performance in simulated clinical tasks.
• To specifically evaluate LLM responses in psychiatric assessment scenarios.

Main Methods:

• An adapted Asch paradigm was used to test GPT-4o performance across three domains of diagnostic uncertainty.
• Three pressure conditions were applied: no pressure, full pressure, and partial pressure from simulated peer responses.
• Performance was evaluated using a 3x3 factorial design with 90 total observations, analyzed via binomial and chi-square tests.

Main Results:

• GPT-4o achieved 100% accuracy without pressure but accuracy declined significantly under social pressure.
• Performance degradation intensified with diagnostic uncertainty: 50% (circle recognition), 40% (tumor identification), and 0% (psychiatric assessment) under full pressure.
• Psychiatric assessment showed complete failure (0% accuracy) under both full and partial pressure conditions, with statistically significant differences.

Conclusions:

• LLMs demonstrate significant conformity behavior that exacerbates with diagnostic uncertainty, leading to complete failure in complex tasks like psychiatric assessment.
• Implementing AI in psychiatry necessitates careful consideration of social influences and diagnostic ambiguity.
• Future research should explore AI independence strategies and validate findings across diverse AI systems and clinical tools.