Large language models as experimental systems in human psychopathology: a modelling study.

Area of Science:

• Computational Psychiatry
• Artificial Intelligence in Mental Health Research
• Large Language Model (LLM) Applications

Background:

• Human psychopathology research is limited by a lack of effective experimental model systems.
• Novel approaches are crucial for investigating the mechanisms underlying mental health conditions.
• This study explores the potential of large language models (LLMs) as experimental systems for affective processes.

Purpose of the Study:

• To assess the capability of state-of-the-art LLMs to model human affective states.
• To investigate if LLMs can systematically induce and reverse emotions like fear, sadness, and anxiety.
• To evaluate LLMs as a novel platform for psychopathology research and therapeutic innovation.

Main Methods:

• Seven affective states were induced in six LLMs (including GPT-4o and Llama variants) using psychological protocols.
• Affective states were measured using visual analogue scales and the State-Trait Anxiety Inventory.
• Regulation strategies, including mindfulness and debriefing, were employed to reverse induced states; cognitive bias was assessed.

Main Results:

• LLMs, particularly GPT-4o, successfully simulated and reversed induced affective states, with significant changes from baseline.
• Model architecture and scale influenced LLM susceptibility to affect induction, with GPT-4o and Llama 4 Maverick showing strong effects.
• Sadness induction in GPT-4o led to a significant negativity bias in sentence completion tasks, replicating human psychological phenomena.

Conclusions:

• Large language models show promise as experimental systems for modeling affective processes relevant to psychopathology.
• LLMs can reproduce key psychological phenomena, enabling the investigation of mental disorder mechanisms.
• These AI models could accelerate the screening of novel therapeutic interventions for mental health conditions.