What happened
Researchers evaluated the performance of three consumer-facing large language models (GPT-5.2 Instant, Claude Sonnet 4.5, and Gemini 3 Pro Preview) against two early-career urologists in urological oncology. The study used a source-masked with 100 synthetic vignettes and found that the LLMs achieved lower composite success rates than the human clinicians. The LLMs also generated safety-critical errors in 25.0% to 26.0% of their responses, compared to error rates of 2.0% and 1.0% observed among the urologists.
The study used a source-masked with 100 synthetic vignettes to evaluate the performance of three consumer-facing large language models (GPT-5.2 Instant, Claude Sonnet 4.5, and Gemini 3 Pro Preview) against two early-career urologists in urological oncology.
The LLMs achieved lower composite success rates than the human clinicians, with GPT-5.2 Instant achieving success in 70 out of 100 vignettes, Claude Sonnet 4.5 in 64 out of 100 vignettes, and Gemini 3 Pro Preview in 60 out of 100 vignettes.
The LLMs also generated safety-critical errors in 25.0% to 26.0% of their responses, compared to error rates of 2.0% and 1.0% observed among the urologists.
The study found that the LLMs had limited repeatability, with exact three-run response stability reached in only 45.0% of combinations for primary composite , 48.3% of combinations for safety-critical status, and 28.3% of combinations for the complete four-domain score profile.
Source details: bioengineer.org
Why it matters
The study highlights the limitations of large language models in clinical decision support and emphasizes the need for separate evaluation of clinical performance, consequential error rates, and response repeatability before clinical deployment. The findings also suggest that clinician supervision may be necessary to mitigate the observed failure modes.
The study highlights the limitations of large language models in clinical decision support and emphasizes the need for separate evaluation of clinical performance, consequential error rates, and response repeatability before clinical deployment.
The findings also suggest that clinician supervision may be necessary to mitigate the observed failure modes.
The study's implications for the development and deployment of large language models in clinical settings are significant, and further research is needed to fully understand the potential risks and limitations associated with their use.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
A route planner searches possible journeys using explicit rules. What does this illustrate about AI?
What to watch next
The study's findings have implications for the development and deployment of large language models in clinical settings. It is essential to evaluate the performance of these models in real-world scenarios and to consider the potential risks and limitations associated with their use.
The study's findings have implications for the development and deployment of large language models in clinical settings.
It is essential to evaluate the performance of these models in real-world scenarios and to consider the potential risks and limitations associated with their use.
Further research is needed to fully understand the potential risks and limitations associated with the use of large language models in clinical settings.