Improving AI-Assisted Medical Diagnosis and Triage by the General Public

As large language models (LLMs) become widely accessible, a growing share of the public already turns to AI-powered chatbots for health-related information: surveys suggest that one in six American adults consults AI chatbots for health queries at least once a month. At the same time, diagnostic errors and misplaced triage decisions represent a persistent source of preventable patient harm globally [4]. This has prompted considerable interest in whether LLMs can serve as a reliable "front door" to the healthcare system for patients who lack immediate access to a clinician.

However, incidents involving misguided medical suggestions from LLMs to the general public such as fatal overdose and erroneous diagnosis leading to life-threatening treatment delay have been reported. A recent controlled study shows LLMs perform significantly worse for medical assistance in real-world settings compared to their performance on controlled benchmarks. Bean et al. [1] conducted a large preregistered study with 1,298 UK participants in which laypeople were assigned to receive assistance from one of three popular LLMs (GPT-4o, Llama 3, or Command R+) or to use a source of their choice when assessing ten standardized medical scenarios. Although the LLMs alone correctly identified relevant conditions in up to 94.9% of cases, participants using those same LLMs did so in fewer than 34.5% of cases, significantly worse than the group using their typical home resources (∼60%). Similarly, triage accuracy showed no signifcant difference between LLM users and controls, with an overall correct response rate of 43.0% across all groups. This poor performance can be linked to two main failure modes: users providing incomplete symptom information to the AI, and users failing to correctly interpret or act upon the LLM's output.

To address these failure modes, the study tests a new GPT-4o interface wrapped with a fixed system prompt. Rather than passively responding to whatever the user volunteers, the AI is instructed to ask clarifying questions to gather a proper clinical history before offering any medical suggestions. Once the model judges that enough information is gathered, it provides a structured, easy-to-read response detailing possible conditions, their likelihoods, and a clear triage recommendation.

The trial is structured as a two-arm, single-blind RCT. Both arms retain access to whatever assistance methods participants would normally use at home (e.g., web search), and both arms additionally get a GPT-4o-based LLM interface. The difference lies in which interface: the treatment arm receives the new interface described above, while the control arm receives a standard, unmodified GPT-4o interface with no special prompting. Participants are restricted to using the provided LLM interface (GPT-4o) only, and not other LLMs. AI-overview in web searches will be disabled through an extension. The trial utilizes ten previously validated clinical vignettes covering various medical urgencies, ranging from self-care routines up to ambulance-level emergencies, with each participant randomly assigned two of the ten. To achieve sufficient statistical power (accounting for within-participant clustering across those two responses), the target sample size is calculated at 220 total participants, split evenly with 110 individuals per arm.

The participant pool is drawn from the enrolled students and administrative staff at the Lahore University of Management Sciences (LUMS) in Pakistan. To ensure the sample consists strictly of laypeople, anyone currently enrolled in or who has completed a medical, nursing, or allied health professional degree is explicitly excluded from participating.

The study is driven by two co-primary hypotheses: participants with access to the new LLM interface will identify relevant medical conditions at a higher rate than participants using the standard LLM interface, and they will correctly assess the urgency of the medical scenarios at a higher rate. Accuracy is measured by comparing participant responses to a physician-generated gold-standard list using fuzzy matching, then analyzed via mixed-effects logistic regression with participant-level covariates and a Bonferroni-corrected significance threshold. The study also tracks secondary outcomes - self-reported confidence and time spent per scenario - and several exploratory analyses, including an LLM-alone benchmark, moderation by LLM experience, internet usage patterns, digital literacy, and scenario-level variation in accuracy.