Study Tests Patient Cognitive Bias in LLM Consultations

A simulation-based comparative study published in the Journal of Medical Internet Research (JMIR) establishes that patient cognitive bias meaningfully degrades LLM diagnostic performance in health consultations. Researchers Yi Zuo, Qifeng Wan, and Shalong Wang developed a simulated patient agent to generate unbiased and confirmation-biased consultations using 1,273 MedQA-USMLE cases, then evaluated six LLMs of varying capacities through multi-turn dialogues. The primary finding: user cognitive bias reduced diagnostic accuracy by 10-40 percentage points (P < .001), with smaller models occasionally performing near chance level. A secondary metric, the bias-influenced error proportion (BIEP), exceeded 33% -- meaning a substantial fraction of model errors directly reflected the user's misconceptions rather than independent model reasoning.

The study used two bias-simulation modes: unbiased consultations and confirmation-biased consultations in which the simulated patient agent steered dialogue toward a preconceived diagnosis. Authors measured three outcomes: diagnostic accuracy, bias-induced accuracy decline (BIAD, loss under bias), and bias-influenced error proportion (BIEP, fraction of errors aligned with user misconceptions). They then tested four prompt-based mitigation strategies, four temperature settings, and a dual-system framework inspired by dual-process cognitive theory -- System 1 being a standard foundation model and System 2 being o1-Mini as a deliberative reasoning layer.

Prompt engineering and temperature adjustments produced limited or inconsistent improvements -- neither reliably counteracted patient confirmation bias. In contrast, the dual-system framework increased accuracy by 10-39 percentage points and recovered most or all of the bias-driven performance gap (P < .001). This suggests architectural interventions, rather than prompting alone, are needed for bias-resilient clinical AI.

For practitioners building or evaluating patient-facing AI tools, the study introduces a concrete and previously underspecified failure mode: users themselves are a source of reasoning error. Standard benchmarks such as MedQA do not capture this dimension; the study's BIAD and BIEP metrics provide a practical evaluation vocabulary. The dual-system result offers a deployment path -- pairing a fast response model with a slower deliberative reasoning model may be a scalable safeguard for higher-stakes medical applications.