Top Line: Large language models (LLMs) can ace the boards, but can they safely triage your headache?
The Study: A randomized trial asking this question is getting lots of buzz. In it, 1,298 UK adult participants tackled 10 medical vignettes with either GPT-4o, Llama 3, Command R+, or a control group who used their usual go-to (e.g., Dr. Google). When tested solo, LLMs identified relevant conditions 95% of the time and selected correct dispositions in 56% of cases. But paired with real people, performance cratered: in the participants’ hands, LLMs identified relevant conditions in <35% of cases and chose the correct disposition only 43% of the time—no better than controls. Transcript analysis revealed the likely culprit as human–LLM interaction failures. Users provided incomplete data, misinterpreted outputs, or ignored correct suggestions. Meanwhile, traditional benchmarks (e.g., MedQA) and even simulated LLM “patients” failed to predict these real-world breakdowns. Ultimately garbage-in, garbage-out remains a truism here, where the LLM in the hands of physicians will likely harness the most diagnostic and therapeutic pathway power. Another important caveat here, though, is the snail pace of medical research (this article was submitted in May 2025, accepted October 2025, and published February 2026) compared to LLM advancement. For example, GPT-4o used in this study is now basically a dinosaur compared to more advanced reasoning versions of GPT-5, Opus 4.6, Gemini 3, or OpenEvidence.
TBL: In this study, a previous generation of LLMs struggled to enhance patients’ medical self-assessment. This may have been, at least in part, from user limitations in interacting with LLMs. Better LLM literacy among patients and physicians could significantly improve the benefits we both gain from interacting with these rapidly advancing tools. | Bean, Nat Med 2026
The Study: A randomized trial asking this question is getting lots of buzz. In it, 1,298 UK adult participants tackled 10 medical vignettes with either GPT-4o, Llama 3, Command R+, or a control group who used their usual go-to (e.g., Dr. Google). When tested solo, LLMs identified relevant conditions 95% of the time and selected correct dispositions in 56% of cases. But paired with real people, performance cratered: in the participants’ hands, LLMs identified relevant conditions in <35% of cases and chose the correct disposition only 43% of the time—no better than controls. Transcript analysis revealed the likely culprit as human–LLM interaction failures. Users provided incomplete data, misinterpreted outputs, or ignored correct suggestions. Meanwhile, traditional benchmarks (e.g., MedQA) and even simulated LLM “patients” failed to predict these real-world breakdowns. Ultimately garbage-in, garbage-out remains a truism here, where the LLM in the hands of physicians will likely harness the most diagnostic and therapeutic pathway power. Another important caveat here, though, is the snail pace of medical research (this article was submitted in May 2025, accepted October 2025, and published February 2026) compared to LLM advancement. For example, GPT-4o used in this study is now basically a dinosaur compared to more advanced reasoning versions of GPT-5, Opus 4.6, Gemini 3, or OpenEvidence.
TBL: In this study, a previous generation of LLMs struggled to enhance patients’ medical self-assessment. This may have been, at least in part, from user limitations in interacting with LLMs. Better LLM literacy among patients and physicians could significantly improve the benefits we both gain from interacting with these rapidly advancing tools. | Bean, Nat Med 2026