Chatbots at the Bedside: What the Physician Trials Actually Show

At 2 a.m., a 58-year-old woman with epigastric pain has a lipase three times the upper limit of normal and an ultrasound read as gallbladder sludge without wall thickening. Your differential has three items on it. The hospitalist wants to know whether this is gallstone pancreatitis before accepting her. Your hospital's approved chatbot is open in another tab, and the question is what to type into it and how much of the answer to trust.

Bottom line: use a chatbot after you have built your own differential, and point it at management questions. Feed it a clean summary you wrote instead of raw chart data, and check every lab interpretation and guideline claim it returns before acting on it.

The diagnostic reasoning trial

Goh and colleagues ran a single-blind randomized trial in 50 US physicians, stratified by career stage, from November 29 to December 29, 2023 [1]. Participants came from Stanford, Beth Israel Deaconess, and the University of Virginia. Of them, 26 were attendings and 24 were residents, and 88% practiced internal medicine. Only 5 (10%) were emergency physicians [1].

Each physician had 60 minutes for up to 6 clinical vignettes based on real patients, using history, examination, and initial labs [1]. The intervention group had ChatGPT Plus (GPT-4) plus their usual resources. The control group used conventional resources such as UpToDate or Google and was told not to use an LLM [1]. No one received training in prompting [1]. The primary outcome was a structured reflection score covering the differential, supporting and opposing findings, the final diagnosis, and next steps [1].

The chatbot did not improve physician performance. The median score was 76% with the LLM and 74% without it, a difference of 2 percentage points (95% CI, −4 to 8) [1]. The subgroups split this way [1]:

• Attendings: 79% vs 75% (difference 0.5 points; p = .92)
• Residents: 76% vs 74% (difference 3 points; 95% CI, −6 to 11)
• Physicians using LLMs more than monthly: 79% vs 74% (difference 5 points; 95% CI, −7 to 16)

Time per case fell by 82 seconds, and that interval crossed zero (95% CI, −195 to 31) [1].

The management reasoning trial

The same group then tested management reasoning in 92 physicians from November 2023 to April 2024 [2]. This cohort included 67 attendings, 25 residents, and 18 emergency physicians (20%) [2]. The five cases were built by experts from real encounters and revealed information in sequence. They asked about treatment choices, testing strategies, and risk, the decisions that follow a diagnosis [2]. Participants got only basic LLM training [2].

Physicians with GPT-4 scored 43.0% against 35.7% for conventional resources, a difference of 6.5 points (95% CI, 2.7 to 10.2; p < 0.001) [2]. The gain came at a cost in time. Physicians with the chatbot spent 119.3 seconds more per case (95% CI, 17.4 to 221.2) [2]. After adjustment for time, the advantage was 5.4 points (95% CI, 1.7 to 9.0) [2].

The question domains split the result [2]:

• Management questions: 40.5% vs 33.4% (difference 6.1; 95% CI, 2.5 to 9.7)
• Diagnostic questions: 56.8% vs 45.8% (difference 12.1; 95% CI, 3.1 to 21.0)
• Factual recall: difference 9.6, with an interval crossing zero (95% CI, −3.1 to 22.3)
• General management knowledge: difference 3.3, with an interval crossing zero (95% CI, −1.3 to 7.9)

In the exploratory harm review, high-likelihood harm was 3.8% with the chatbot and 2.9% without it. Severe-harm ratings were 7.6% and 7.5% [2]. The rubric gave credit only for correct answers and did not penalize wrong ones, so it does not capture the cost of a bad suggestion [2].

Why the chatbot alone did as well

In both trials the model working alone matched or beat the physicians who had it. In the diagnostic trial, GPT-4 alone scored a median of 92%. That was 16 points above the conventional-resources group (95% CI, 2 to 30), while physicians with GPT-4 scored 76% [1]. In the management trial, GPT-4 alone scored 43.7%, essentially the same as physicians using it (43.0%) [2].

The authors of the diagnostic trial said realizing that potential depends on better human-computer interaction [1]. Physicians with the tool did not use it in a consistent way, and none were trained to prompt it [1]. In the diagnostic trial, the researchers wrote the model-alone prompt with established prompt-design principles, which is a better input than a clinician typing quickly between patients [1].

Both trials used curated vignettes, summarized by clinicians, with no patient interviewing, no data collection, and no clinical context [1]. That is the most favorable input a model will ever see.

What happens with raw ED data

Hager and colleagues tested what happens when the input is closer to a real chart. They built a dataset from 2,400 MIMIC-IV patients who presented to the ED with abdominal pain and had appendicitis, cholecystitis, diverticulitis, or pancreatitis [3]. The models were five 70-billion-parameter open-access models based on Llama 2. GPT-4 was not tested, because the MIMIC data-use agreement bars sending records to an external service [3].

On 80 cases, the models performed significantly worse than four hospitalists. The gap was 16 to 25 points, with residents averaging 87.5% and a senior physician 92.5% [3]. The failure patterns matter more than the scores, because they describe how a chatbot goes wrong on clinical data [3]:

• Lab interpretation: given each result and its reference range, models identified low values correctly only 26.5% to 70.2% of the time.
• Information order: reordering the same findings changed accuracy by as much as 18.0 points.
• Information quantity: more information often lowered accuracy, and removing normal labs improved it.
• Wording: asking for a "primary" instead of a "final" diagnosis shifted accuracy by up to 10.6 points.
• Treatment: models rarely recommended source control for severe disease, such as drainage of infected pancreatic necrosis.

The authors concluded that these models are not ready for autonomous clinical decision-making and need extensive clinician supervision [3]. The models tested are older than the chatbots in use today, so the scores do not transfer directly. The failure patterns are the lesson, because each one comes from the structure of the input [3].

Using a chatbot on shift

The evidence supports a narrow set of habits:

• Order: write your own differential first, because the diagnostic trial found no gain when physicians used the tool freely [1].
• Target: ask management questions such as next tests, treatment options, and disposition risks, where the measured benefit was largest [2].
• Input: give a short summary you wrote, with abnormal values flagged, instead of pasting raw lab panels [3].
• Verification: confirm every lab interpretation, dose, and guideline citation against the source.
• Time: expect about two extra minutes per case, and do not use the tool when that time is not available [2].
• Tool: use only a model your institution has approved for patient information.

Key Takeaways

• In 50 physicians, GPT-4 access moved diagnostic reasoning scores from 74% to 76%, a difference whose interval crossed zero.
• In 92 physicians, GPT-4 raised management reasoning scores by 6.5 points and added about 2 minutes per case.
• The model working alone matched or beat physicians who had access to it, which points to how clinicians use the tool as the limiting step.
• With real ED data, open models classified low lab values correctly only 26.5% to 70.2% of the time, and reordering the same facts changed accuracy by up to 18 points.

FAQ

Does ChatGPT improve physician diagnostic accuracy?

Not in the only randomized trial. In 50 physicians, GPT-4 access produced a median diagnostic reasoning score of 76% versus 74% with conventional resources (difference 2 points; 95% CI, −4 to 8) [1]. GPT-4 alone scored 92%, higher than either physician group [1].

Can large language models help emergency physicians with management decisions?

In a randomized trial of 92 physicians, 20% of them emergency physicians, GPT-4 assistance raised management reasoning scores from 35.7% to 43.0% [2]. Physicians spent about 119 seconds more per case, and harm ratings were similar between groups [2].

Are AI chatbots safe for autonomous clinical decisions in the ED?

No. In 2,400 ED abdominal pain cases, open-access LLMs performed significantly worse than physicians, misread lab values, and failed to follow treatment guidelines [3]. The authors concluded the models need extensive clinician supervision [3].

Dr. Chet's Take

I have built clinical AI education for emergency physicians, and the question I hear most often is whether the chatbot makes them better. This article gives the honest answer: it depends on what you ask and how you ask it. The diagnostic trial is the one people quote to dismiss these tools, and the management trial is the one enthusiasts quote to promote them. None of these studies watched a physician who was interrupted three times while typing the question, and that is the condition every emergency physician actually works under. Read together, the trials describe a tool that helps most when the physician has already done the thinking and needs a second set of eyes on the plan. That framing matters, because a physician who expects a diagnosis machine will be disappointed, and one who expects a reviewer will get value.

That being said, these trials were run on vignettes in a quiet room with an hour on the clock. The ED is the opposite environment. The extra time that bought a better management score is time a single-coverage night shift does not have, and nobody has measured what happens to the patients in the waiting room while the physician types. I would also point out that emergency physicians were a small minority of the participants. The specialty that makes the most decisions per hour under the most uncertainty is barely represented in the evidence, and that should bother every department chair reading these results. Chatbots also never get tired, and the end of a long shift is exactly when a physician is most likely to accept an answer without checking it.

If you are an emergency physician, practice on easy cases first. Write a three-line summary, ask one management question, and see whether the answer changes your plan. If you lead a department, measure use before you encourage it, and track whether consults, imaging, or length of stay move in either direction. If you sit on an AI governance committee, ask for an ED-specific evaluation with real charts before approving a chatbot for clinical use. If you teach residents, have them commit to a plan out loud before they open the chatbot, then discuss what the tool added and what it got wrong. Keep a shared log of the wrong answers so the whole department learns from them. None of the three studies tested a chatbot during live patient care.

— Chester Shermer, MD, FACEP | Emergency Medicine, 25+ Years Clinical Experience | State Surgeon

Continue Your Training

If you're an emergency physician (or any clinician treating patients daily) trying to understand how AI will actually impact your clinical practice — not just the hype — I put together a free practical guide. You can download the AI in EM Survival Guide here.

Browse All Courses at Global MedOps Command

Relevant Reading on Global MedOps Command:

How to Avoid Becoming an AI Casualty — Dr. Shermer's guide to navigating AI tools in clinical and operational settings without compromising judgment or patient outcomes.

Emergency Department Efficiency Playbook — Practical systems for throughput, triage optimization, and operational efficiency.

Read more from Dr. Shermer on Medium

Connect with Dr. Shermer: LinkedIn — Chester "Chet" Shermer, MD, FACEP.

References

1. Goh E, Gallo R, Hom J, et al. Large language model influence on diagnostic reasoning: a randomized clinical trial. JAMA Netw Open. 2024;7(10):e2440969. PubMed 39466245
2. Goh E, Gallo RJ, Strong E, et al. GPT-4 assistance for improvement of physician performance on patient care tasks: a randomized controlled trial. Nat Med. 2025;31(4):1233-1238. PubMed 39910272
3. Hager P, Jungmann F, Holland R, et al. Evaluation and mitigation of the limitations of large language models in clinical decision-making. Nat Med. 2024;30(9):2613-2622. PubMed 38965432

Related simulation training: EM-Sim · Books by Dr. Shermer