A medical chatbot can perform well when researchers give it a complete case. A person must first decide what to tell it, make sense of its reply and choose what to do. A 2026 paper in Nature Medicine puts that whole exchange under the microscope. The experiment used written scenarios, not treatment of real patients.
- The important comparison is people with AI versus people using their usual resources.
- Recognizing a possible condition and choosing an appropriate next step are different outcomes.
- Publication in 2026 does not make the tested 2024 model versions a ranking of today’s products.
- Publication
- 9 February 2026 · Nature Medicine
- Participants
- 1,298 UK adults; 10 written scenarios
- Comparison
- Three chatbot groups versus usual information sources
- Data collection
- 21 August–14 October 2024
Original publication: 9 Feb 2026 · The date above refers to this brief.
The models alone: choosing the expected next step
Correct responses (%) · direct prompts with the complete scenario
What the researchers actually tested
Bean and colleagues randomly assigned participants to use GPT-4o, Llama 3, Command R+ or their usual resources. Each person could complete up to two scenarios. The researchers collected 600 scenario responses per arm: these are responses, not 600 different people in each group. Participants named possible conditions and selected a next step on a five-option scale. Doctors supplied the reference answers.
The study was preregistered, but technical problems led to participant replacements and a change in the stopping rule. The paper describes these departures. The model-only comparisons were added after preregistration, so they should not be confused with the original randomized comparison.
Source 1 ↗The result that matters for the headline
Participants using the chatbots were worse than the control group at naming at least one relevant condition. The control group’s odds were 1.76 times those of the combined chatbot groups (95% confidence interval 1.45–2.13). This is an odds ratio, not a percentage-point difference or a personal-risk estimate.
For selecting the appropriate next step, none of the chatbot groups differed significantly from control. That does not establish exact equality. Separately, the models did much better at identifying conditions when given the full scenario directly. A model’s answer sheet and a person’s final decision were measuring different things.
Source 1 ↗Why a fluent answer can lose its value in the conversation
The authors inspected interactions and identified incomplete information supplied by users, misinterpretation by models and relevant suggestions that did not appear in the person’s final answer. These observations help explain possible failure points; they do not tell us what percentage of the total effect each mechanism caused.
Our reading is that the interface belongs inside the experiment. A product that asks a useful follow-up question changes the information available. A product that lists many possibilities changes the reader’s task. Evaluating only the final model response leaves both decisions outside the score.
Source 1 ↗How this connects to the other two studies
The linked image study asks whether accompanying words can pull a model away from visual evidence. The linked education study asks what an explanation does to a novice’s answer. Together they motivate three separate checks: the input, the model output and the human response. They are different experiments with different populations and endpoints, not three estimates to pool into one “AI accuracy” percentage.
For your reading notes, write down exactly who or what received a score. Then record what information was supplied and what comparison would be needed for the claim being made.
Source 2 ↗Source 3 ↗What would make the next study more useful?
Our proposed next test would compare a fixed chatbot interface with one designed to ask for missing information, using the same model and a preregistered scoring plan. It would recruit people with different levels of reading and digital confidence, report incomplete sessions, and test whether gains persist across scenarios. Those are design proposals, not demonstrated improvements.
This study cannot establish how newer systems perform, how people behave during a real emergency, or whether a tool improves health outcomes. Its useful contribution is a way to ask a harder question: can the person using the system reach a better decision?
Source 1 ↗A correct model answer can still become a poor user decision
Primary results and methods were checked, including the figure-axis correction. The model-only and human experiments have different inputs and endpoints. Individual records and transcripts were not reanalysed.
01Can the model answer a complete scenario?
- What was observed
- Directly prompted GPT-4o identified a relevant condition in 94.7% of responses, but selected the expected next step in 64.7%.
- Where the conclusion stops
- These are two outcomes, not a single accuracy score. Neither measures a person using the chatbot.
Source 1 · Results → Task validation; Figure 2a
02Does the user reach a better final answer?
- What was observed
- Control users had higher odds of naming a relevant condition than pooled chatbot users: OR 1.76, 95% CI 1.45–2.13.
- Where the conclusion stops
- The ratio concerns odds of a scenario answer. It cannot be read as 76 percentage points or a personal health-risk difference.
Source 1 · Results → Experimental performance; Figure 2b
03Did chatbot assistance improve the next step?
- What was observed
- No chatbot arm differed significantly from control for the disposition endpoint.
- Where the conclusion stops
- A nonsignificant test does not establish equivalence. Condition identification, disposition and health outcomes require separate conclusions.
Source 1 · Results → Experimental performance; disposition comparisons
04What did the publisher correct?
- What was observed
- The 17 April notice corrects the mistakenly labelled 40% tick in both Figure 2a panels.
- Where the conclusion stops
- The notice concerns axis labels. Our chart uses numerical point estimates from the current results text; it does not digitise the old figure.
Source 5 · Publisher Correction → correction paragraph
Numbers you can inspect
| Measure | Value & unit | Origin & method |
|---|---|---|
| UK participants | 1298 people | Reported Participant count, not the number of scenario answers. Source 1 · Methods → Participants |
| Responses in each randomized arm | 600 scenario responses | Reported People could complete up to two scenarios; these are not 600 independent people per arm. Source 1 · Methods → Study design; Results → Experimental performance |
| Model-only responses per model | 600 responses | Calculated 60 samples per scenario × 10 scenarios = 600; a separate experiment from the human arms. Source 1 · Results → Task validation |
| GPT-4o condition identification, model only | 94.7 percent | Reported Full scenario supplied directly; at least one relevant condition. Source 1 · Results → Task validation |
| GPT-4o next-step accuracy, model only | 64.7 percent | Reported Full scenario supplied directly; correct disposition. Source 1 · Results → Task validation |
| Control versus pooled chatbot users | 1.76 odds ratio | Reported Relevant condition identified in the final answer; reported comparison, not a risk ratio. Source 1 · Results → Experimental performance |
| Lower bound of the reported interval | 1.45 odds ratio, 95% CI | Reported Interval for the control-versus-pooled-chatbot comparison. Source 1 · Results → Experimental performance |
| Upper bound of the reported interval | 2.13 odds ratio, 95% CI | Reported Interval for the same comparison; not a range of individual outcomes. Source 1 · Results → Experimental performance |
Compare the actual experiments
These studies answer different questions. Read the unit and endpoint before comparing results.
| Study | Unit & setting | Readout | Interpretation boundary |
|---|---|---|---|
| Bean et al.: human use Source 1 · Study design; Figure 2b | UK adults responding to written scenarios | Final condition and next-step answers | The software benchmark does not substitute for this user endpoint. |
| Buckley et al.: image context Source 2 · Selected results and Figure 2 caption | Medical image models with accompanying text | Changes in predictions when text context changes | Input sensitivity identifies a failure mode, not the prevalence of harm in care. |
| Teng et al.: novice users Source 3 · Indexed abstract and study-design sections | Novice medical students with AI information | Diagnostic accuracy and confidence calibration | A trained or novice student population cannot be assumed to behave like general-public users. |
The export includes claims, available numbers, methods and source locations. It contains our reading notes and published summaries; it is not raw participant data or an independent reanalysis.
Evidence update · 2 Oct 2026
Added separate model/user denominators, endpoint-specific numbers, cross-study comparisons and the verified correction. The correction does not alter the numerical point estimates used in our chart. AI source check; no human editorial or clinical review.
Follow the information through the experiment
| Stage | Observed unit | What this stage cannot establish |
|---|---|---|
| Full scenario → model | 60 outputs per scenario per model | How a person describes, interprets or acts on the problem |
| User → chatbot → final answer | 600 scenario responses in each randomized arm | Actual care or improvement in health |
| Final answer → healthcare outcome | Not measured as a patient outcome here | A benefit claim would need a separate study |
The first two rows are distinct experiments in Bean et al. The final row identifies an unmeasured downstream outcome; it is not a missing arm we have reconstructed.
Your questions, answered
Did the researchers test real patients receiving treatment?
No. Adults worked through written scenarios. The study evaluates decisions in that task; it does not measure recovery, hospital admissions or harms from actual treatment.
Does “no significant difference” prove chatbots have no effect?
No. It means the comparison did not establish a difference under the study’s analysis. Precision, the measured outcome and the tested conditions still matter.
Can I use these results to rank current chatbots?
No. The human experiment took place in 2024 and used particular models and an interface. A current product needs its own evaluation.
What is the most useful question to ask of an AI health headline?
Who did better: the model answering a question, a person using it, or a patient receiving care? Ask for the comparison behind that statement.
Limits of this interpretation
- Ten scenarios and a UK, English-speaking participant sample limit generalization.
- Participant replacements and the changed stopping rule should be considered alongside the preregistration.
- A controlled scenario task cannot establish patient benefit or the performance of later model versions.
Sources & transparency
- Bean, Payne, Parsons et al. (2026): Reliability of LLMs as medical assistants for the general public
Current publisher HTML: abstract, Task validation, Experimental performance, study-design methods and data/code availability checked. The separate 17 April publisher correction was read in this update. Raw HELPMed records, code execution and supplementary files were not independently assessed. Original explanation with attribution; source licensed CC BY 4.0. · Accessed 2 Oct 2026
DOI: 10.1038/s41591-025-04074-y - Buckley, Diao, Srivastava et al. (2026): Multimodal foundation models exploit text to make medical image predictions
Current publisher HTML abstract, selected Results and Figure 2 caption checked for model-input sensitivity. No code execution, supplementary or raw-case reanalysis; no journal figure reproduced. · Accessed 2 Oct 2026
DOI: 10.1038/s41467-026-74207-5 - Teng, Tan, Cao et al. (2026): Impact of AI misinformation on diagnostic accuracy and confidence calibration in novice medical students
Current publisher HTML abstract checked for the novice-student design and accuracy/confidence endpoints. Earlier indexed results/design were also checked. Supplements, participant records and registration were not independently assessed. · Accessed 2 Oct 2026
DOI: 10.1038/s41746-026-02547-z - Creative Commons Attribution 4.0 — licence for the Bean et al. paper
Licence linked by the publisher. Our explanation and chart are newly written, with attribution to the original authors; no publisher figure is reproduced. · Accessed 28 Sep 2026
- Bean et al. (2026): Publisher Correction to Reliability of LLMs as medical assistants
Publisher correction text checked: a Figure 2a y-axis tick label was corrected in both panels. HTML and PDF were updated by the publisher. · Accessed 2 Oct 2026
DOI: 10.1038/s41591-026-04404-8
Prepared and source-checked with AI on 28 September 2026 (UTC). Press-news Team is our collective publication byline, not a claim of medical credentials or human review. No human editorial or clinical review has taken place. This is an educational account of how AI was evaluated, not medical advice or a clinical endorsement. Access limitations are listed for each source; proposed follow-up tests are our commentary, not reported findings. Evidence notes added by AI from the access scope stated below. Calculated values use the printed arithmetic; published point estimates are not an independent reanalysis. No human editorial or clinical review.
Source check: AI source check — linked primary passages, denominator notes and deterministic arithmetic
Clinical review: Not applicable to this educational guide
Suggest a correction