Evaluating the o1 reasoning large language model for cognitive bias: a vignette study

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Cognitive Psychology

Background:

• Cognitive biases are systematic errors in judgment common in high-pressure clinical settings.
• Previous AI models like GPT-4 have demonstrated susceptibility to these biases, sometimes exceeding human performance.
• Understanding AI susceptibility to cognitive bias is crucial for safe medical applications.

Purpose of the Study:

• To evaluate the susceptibility of the o1 AI model, a new system with enhanced reasoning, to common cognitive biases in clinical decision-making.
• To compare the o1 model's bias performance against established GPT-4 models and human clinicians.
• To assess the potential of advanced AI reasoning models to mitigate cognitive bias in medical contexts.

Main Methods:

• Ten pairs of clinical scenarios were designed to test specific cognitive biases, with subtle modifications between versions.
• The o1 model generated 1,800 clinical recommendations across paired scenarios to measure systematic bias.
• Performance was benchmarked against prior GPT-4 and human clinician data using established methodologies.

Main Results:

• The o1 model exhibited no measurable bias in 70% of tested vignettes.
• In biased vignettes, o1's bias magnitude was lower than GPT-4 and human clinicians.
• The model showed consistent bias in one vignette (Occam's razor) and was more prone to bias with gap-closing cues.
• Intra-scenario agreement exceeded 94%, indicating lower decision variability than previously reported.

Conclusions:

• Advanced AI reasoning models show promise in reducing cognitive bias and decision variability in medicine.
• However, AI models are not entirely immune to cognitive bias, necessitating further research into failure circumstances.
• AI decision-support tools may offer benefits but require careful evaluation for safe and effective clinical integration.