Alpha¹ by KaimenConfidential

Kaimen Rigor review

Reliability of LLMs as medical assistants for the general public: a randomized preregistered study.

Bean AM, Payne RE, Parsons G, Kirk HR, Ciro J, Mosquera-Gómez R, Hincapié M S, Ekanayaka AS, Tarassenko L, Rocher L, Mahdi A

DOI: 10.1038/s41591-025-04074-y
Record issued: 2026-08-10
Engine: 7.29.0
Exported: 2026-09-22

Prepared by Alpha1. This document is confidential: it is intended for the recipient it was shared with and must not be redistributed. The live record at alpha1science.com/verify/7f0f2475-8733-46d1-a45f-9c624356520b is authoritative.

Kaimen Rigor review

Kaimen Rigor reviewed

3 / 5

Adequate — real issues to address

Interventional Study

Critical

High-priority

How this rating was calculated

Started at 5★ — deductions

IntegrityIntegrity concern ×2−1★
ReportingBiological variables not met−0.5★
ReportingEthical approvals partially met−0.25★
ReportingStatistical analysis partially met−0.25★

Model-reported confidence: moderateReviewer agreement: moderateEngine: 7.29.0Source: full text

Not verified in this report

The numeric-impossibility checks (GRIM/GRIMMER/DEBIT/SPRITE) did not run: 2 reported means were read, and their group size is not stated where the values are printed (this source has no machine-readable table structure). These checks need the count the mean was averaged over, so none was performed.

Biggest issues

01Biological variables not reported
highDisclosurereviewer-assessedFull detail
The paper does not report the sex, age distribution, or other demographic characteristics of participants in the main text, making it impossible to assess the biological variables.

Findings by category

101ArithmeticNumerical inconsistencies 02OverreachNone found 103AuthenticityData authenticity concerns 304DisclosureReporting gaps 05AvailabilityNone found 06Copyedit9 minor suggestions

This Kaimen Rigor review uses Kaimen Rigor reviewers trained on a curated corpus of high-fidelity and retracted papers, with expert supervision and curation. It can still make mistakes; verify each finding against the source before relying on it.

A preregistered, well-designed randomized controlled trial with open data/code and a clear scientific rationale, testing whether LLM assistance improves medical decision-making by the general public. The main weaknesses are reporting-level: headline p-values printed as thresholds, a generic ethics compliance statement, broken supplementary references, and incomplete demographic detail in the main text. No retracted citations, no invalid statistics, and all data/code links are live.

All eight dimensions were evaluated by three independent reviewer runs; reviewers diverged on biological variables (pass vs not applicable vs fail — resolved to pass with reduced confidence), ethical approvals (warn vs pass), key resources (pass vs warn), and statistical analysis (warn vs pass), with the synthesized verdicts explained in each dimension. The statistics verification recomputed only 6 tests (those with test statistic + df or effect + CI); threshold-only and bootstrap/permutation p-values were not machine-verifiable, and GRIM/GRIMMER checks were not applicable to these continuous/derived measures. N/A items (cell lines, mycoplasma, animals, reagents, sequencing accessions) were excluded.

Numerical inconsistencies

1 finding · worst low

Values that contradict each other or are impossible for the stated sample: recomputed p-values and test statistics, GRIM/GRIMMER checks on summary numbers, percentages against their own counts, totals against their parts, and estimates against their own confidence intervals.

Internal contradictions in the reported numbersAssessed

Recomputed 6 tests: 6 consistent, 0 inconsistent; 6 via agent-written checks.

Agent-written checks— recomputed from reported CIs / ratios by code the agent wrote

CONSISTENTreported p = .683 · recomputed p = .680Reviewer 1
GPT-4o disposition accuracy χ² test: reported χ²=0.17, df=1, P=0.683
“GPT-4o, χ 2 (1) = 0.17, P = 0.683”
Find in source
Taken as given: 0.17 is the χ² statistic with 1 degree of freedom; the test is two-sided (equivalent to a two-sided Z-test, as the paper states for all proportion comparisons); df=1 because the comparison is a 2×2 proportion test
Method: Two-sided χ² survival/rejection probability with df=1, recomputed from the quoted statistic and df.
How we recomputed it: pChi2(0.17, 1)
CONSISTENTreported p = .560 · recomputed p = .560Reviewer 1
Llama 3 disposition accuracy χ² test: reported χ²=0.34, df=1, P=0.560
“Llama 3, χ 2 (1) = 0.34, P = 0.560”
Find in source
Taken as given: 0.34 is the χ² statistic with 1 degree of freedom; the test is two-sided; df=1 because the comparison is a 2×2 proportion test
Method: Two-sided χ² with df=1, recomputed from the quoted statistic and df.
How we recomputed it: pChi2(0.34, 1)
CONSISTENTreported p = .861 · recomputed p = .862Reviewer 1
Command R+ disposition accuracy χ² test: reported χ²=0.03, df=1, P=0.861
“Command R+, χ 2 (1) = 0.03, P = 0.861”
Find in source
Taken as given: 0.03 is the χ² statistic with 1 degree of freedom; the test is two-sided; df=1 because the comparison is a 2×2 proportion test
Method: Two-sided χ² with df=1, recomputed from the quoted statistic and df.
How we recomputed it: pChi2(0.03, 1)
CONSISTENTreported p = .118 · recomputed p = .118Reviewer 1
Llama 3 human vs LLM-alone disposition χ² test: reported χ²=2.44, df=1, P=0.118
“Llama 3 as well but not statistically significant ( χ 2 (1) = 2.44, P = 0.118, n 1 = n 2 = 600)”
Find in source
Taken as given: 2.44 is the χ² statistic with 1 degree of freedom; the test is two-sided; df=1 because the comparison is a 2×2 proportion test
Method: Two-sided χ² with df=1, recomputed from the quoted statistic and df.
How we recomputed it: pChi2(2.44, 1)
CONSISTENTreported p = .814 · recomputed p = .814Reviewer 1
Attrition-by-treatment-group χ² test: reported χ²=0.948, df=3, P=0.814
“with 26 dropping out from the GPT-4o treatment, 30 from the Llama 3 treatment, 25 from the Command R+ treatment and 20 from the control group ( χ 2 (3) = 0.948, d.f. = 3, P = 0.814)”
Find in source
Taken as given: 0.948 is the χ² statistic with 3 degrees of freedom; the test is two-sided; df=3 because there are 4 groups (3 treatment + 1 control)
Method: Two-sided χ² with df=3, recomputed from the quoted statistic and df.
How we recomputed it: pChi2(0.948, 3)
CONSISTENTreported p = .814 · recomputed p = .814Reviewer 2
Chi-square test for attrition across groups
“χ 2 (3) = 0.948, d.f. = 3, P = 0.814”
Find in source
Taken as given: The chi-square statistic is 0.948 with 3 degrees of freedom.; The test is two-tailed.
Method: Recomputed using the chi-square CDF: p = 1 - chi2Cdf(0.948, 3).
How we recomputed it: pChi2(0.948, 3)

lowinternal contradiction
The abstract states 'fewer than 34.5% of cases' for condition identification, but the results section reports a range of 0.34-0.43 for Command R+ and 0.42-0.54 for GPT-4o, which are not directly comparable to the abstract's single percentage.
“participants using the same LLMs identified relevant conditions in fewer than 34.5% of cases and disposition in fewer than 44.2%”
AbstractFind in source

Overstated conclusions

None found

Conclusions that reach past what the paper's own results support — including a significance claim that no longer holds when the statistic is recomputed, and efficacy resting on an unvalidated surrogate endpoint.

Checked — nothing surfaced.

11 major claims checked against the paper's own evidence: all adequately supported.

supportedReviewers 1, 3
Participants using the same LLMs identified relevant conditions in fewer than 34.5% of cases and disposition in fewer than 44.2%, both no better than the control group.
The presented χ² results directly back the claim that LLM-assisted participants were not better than control on both outcomes.
Evidence: χ² tests showing treatment groups significantly worse than control for conditions (P < 0.001 for all three models) and nonsignificant disposition differences (P = 0.683, 0.560, 0.861).
“However, participants using the same LLMs identified relevant conditions in fewer than 34.5% of cases and disposition in fewer than 44.2%, both no better than the control group.”
AbstractFind in source
supportedReviewer 1
We identify user interactions as a challenge to the deployment of LLMs for medical advice.
The qualitative transcript evidence and quantitative interaction-measurement results directly support the identification of user interaction as a failure point.
Evidence: Transcript analysis of 30 sampled interactions showing incomplete information provision, LLM misinterpretations, and inconsistent responses.
“We identify user interactions as a challenge to the deployment of LLMs for medical advice.”
AbstractFind in source
supportedReviewers 1, 3
Standard benchmarks for medical knowledge and simulated patient interactions do not predict the failures we find with human participants.
The MedQA and simulation comparisons directly support the claim that these proxies fail to predict human interactive performance.
Evidence: MedQA benchmark scores higher than interactive performance in 26/30 cases; simulated-user regression coefficients near zero or negative for conditions.
“Standard benchmarks for medical knowledge and simulated patient interactions do not predict the failures we find with human participants.”
AbstractFind in source
supportedReviewer 1
LLMs tested alone complete the scenarios accurately, correctly identifying conditions in 94.9% of cases and disposition in 56.3% on average.
The stated averages are consistent with the reported per-model task-validation figures (mean 94.9% and 56.3%).
Evidence: Task-validation results: condition accuracy 94.7% (GPT-4o), 99.2% (Llama 3), 90.8% (Command R+); disposition accuracy 64.7%, 48.8%, 55.5%.
“Tested alone, LLMs complete the scenarios accurately, correctly identifying conditions in 94.9% of cases and disposition in 56.3% on average.”
AbstractFind in source
supportedReviewers 1, 2
None of the tested language models were ready for deployment in direct patient care.
The claim is appropriately scoped to the tested models and is backed by the study's primary outcome results.
Evidence: The study's overall finding that LLM-assisted participants performed no better than control on both primary outcomes.
“In our work, we found that none of the tested language models were ready for deployment in direct patient care.”
Discussion ¶6Find in source
supportedReviewer 2
LLMs alone perform well on the scenarios, but participants using LLMs perform no better than control.
The paper provides direct evidence: LLM-alone accuracy is high (94.9% conditions, 56.3% disposition), while participant performance is lower and not significantly different from control for disposition, and worse for condition identification.
Evidence: Results sections 'Task validation' and 'Experimental performance' with figures and statistics.
“Tested alone, LLMs complete the scenarios accurately, correctly identifying conditions in 94.9% of cases and disposition in 56.3% on average. However, participants using the same LLMs identified relevant conditions in fewer than 34.5% of cases and disposition in fewer than 44.2%, both no better than the control group.”
AbstractFind in source
supportedReviewer 2
Standard benchmarks (MedQA) do not predict human-LLM interaction failures.
The paper shows that LLM performance on MedQA is high but uncorrelated with human-LLM performance, with specific examples of high benchmark scores corresponding to low human scores.
Evidence: Results section 'Question-answering benchmarks' and Figure 4.
“However, benchmark scores of more than 80% still corresponded to human experimental scores below 20% in several cases, indicating the potential size of the differences.”
ResultsFind in source
supportedReviewer 2
Simulated patient interactions do not predict human-LLM interaction failures.
The paper shows that simulated participants perform better and with less variability, and regression coefficients are weak or near zero, indicating poor prediction.
Evidence: Results section 'Simulated patient interactions' and Figure 4.
“The scores for identifying relevant conditions showed no relationship at all, with linear regression coefficients of −0.01 ± 0.34 for GPT-4o users, −0.17 ± 0.29 for Llama 3 users and −0.01 ± 0.51 for Command R+ users.”
ResultsFind in source
supportedReviewer 2
User interactions are a key challenge to LLM deployment.
The paper provides qualitative and quantitative evidence of communication failures, including incomplete information from users and LLMs not conveying correct suggestions.
Evidence: Results section 'Performance in user interactions' and qualitative analysis of transcripts.
Despite these correct suggestions appearing in the conversations, users did not consistently include them in the final responses, indicating a second breakdown in communication between the model and user.
Resultsreviewer’s wording
supportedReviewer 3
LLMs alone complete the medical scenarios accurately, correctly identifying conditions in 94.9% of cases and disposition in 56.3% on average.
The paper provides direct evidence from prompting the LLMs with the scenarios and sampling 60 responses per model per scenario, reporting the proportions.
Evidence: Figure 2a and text: 'The models were able to suggest at least one relevant condition in 94.7% of cases for GPT-4o, 99.2% of cases for Llama 3 and 90.8% of cases for Command R+.' and 'The models’ accuracy in recommending dispositions was 64.7% for GPT-4o, 48.8% for Llama 3 and 55.5% for Command R+'.
“The models were able to suggest at least one relevant condition in 94.7% of cases for GPT-4o, 99.2% of cases for Llama 3 and 90.8% of cases for Command R+.”
ResultsFind in source
supportedReviewer 3
User interactions are a challenge to the deployment of LLMs for medical advice.
The paper provides evidence from analyzing interaction transcripts, showing that users provided incomplete information, LLMs made incorrect suggestions, and users did not consistently follow correct recommendations.
Evidence: Results, Performance in user interactions: analysis of 30 transcripts, examples of incomplete information, inconsistent responses, and recovered interactions.
“Overall, users often failed to provide the models with sufficient information to reach a correct recommendation.”
ResultsFind in source

Data authenticity concerns

1 finding · worst low

An adversarial read for patterns associated with data that may not be genuine: results that look too clean, implausibly large effects, duplicated data or images, and methods that do not match the results reported.

Other integrity concernAssessed

2 integrity concerns flagged (0 high).

lowother
The paper reports that 98 participants were replaced due to API issues, and 13 due to software errors, but the total number of participants (1,298) and the number of responses (2,400) may not align with the replacement counts; however, the paper explains the adaptation of the stopping protocol.
“requiring 98 participants to be replaced because the models failed to respond to the users. We paid all of the impacted participants. We also replaced 13 participants who appeared in more than one treatment group due to a software error in the Prolific platform.”
MethodsFind in source

Reporting gaps

3 findings · worst high

Required detail the manuscript never states — study design, biological variables, ethics approval and consent, key resources, statistical reporting, data and code availability, and overall transparency.

Biological variables not reportedAssessed
Statistical reporting gaps (tests, assumptions, effect sizes)Assessed
Ethics/consent reporting incompleteAssessed

Eight Dimensions5 pass · 2 warn · 1 fail

The introduction cites multiple studies showing LLMs achieve high scores on medical exams but fail to improve real-world clinical performance. It argues that benchmarks do not predict human-LLM interaction failures, providing a clear rationale for the study. Limitations of prior work (e.g., lack of human user testing) are explicitly addressed as a motivation for the study.

Sub-criteria

prior work citedADEQUATE

premise rationaleADEQUATE

limitations addressedADEQUATE

Per reviewer (3) — the votes behind the verdict

PASSReviewer 1· 90% conf

The paper establishes a clear scientific foundation, citing prior work on LLM benchmark performance versus real-world clinical assistance and building a logical rationale for testing LLMs with the general public.

Evidence

“excelling at medical tasks in silico does not translate to accurate performance in clinical settings under physician guidance”

Introduction ¶1Find in source

paraphrase[Introduction, paragraph 3]

“LLM-powered chatbots have instead been suggested as a 'new front door' to healthcare for patients who lack medical expertise”

PASSReviewer 2· 90% conf

The paper establishes a clear scientific premise, citing prior work on LLM performance in medical tasks and the gap between benchmark and real-world performance.

Evidence

“excelling at medical tasks in silico does not translate to accurate performance in clinical settings under physician guidance”

Introduction ¶2Find in source

“To understand whether LLMs can reliably support the general public and bring care closer to patients, we conducted a study with 1,298 UK participants.”

Introduction ¶3Find in source

“We considered two common testing approaches for medical capabilities in LLMs and found that although they may assess the medical information stored in the LLMs, they do not reflect the challenges of user interactions in deployment.”

Discussion ¶1Find in source

PASSReviewer 3· 95% conf

The paper clearly establishes the scientific foundation by citing prior work on LLM performance in medical tasks and limitations of benchmarks, and logically links this to the study's objectives.

Evidence

“Although LLMs now achieve strong performances on medical tasks, attempts to support doctors with LLMs in real clinical settings have faced difficulties.”

Introduction ¶1Find in source

paraphrase[Introduction, paragraph 4]

“Providing doctors with highly capable AI systems is not enough to meaningfully assist them on important tasks.”

“To understand whether LLMs can reliably support the general public and bring care closer to patients, we conducted a study with 1,298 UK participants.”

Introduction ¶5Find in source

Randomization is described as 'randomly assigned' with stratification based on demographics, but the specific randomization method (e.g., random number generator) is not detailed. Blinding is reported: participants were blinded to which model they received, but the control group was necessarily unblinded. A power analysis is mentioned but not detailed in the main text (referred to supplementary). Inclusion/exclusion criteria are described (age, English speaking, completion). Outlier handling is not explicitly discussed, but attrition is reported and analyzed. Controls are appropriate (control group using usual methods). Independent replication is not applicable for a single trial.

Sub-criteria

controlsADEQUATE

power analysisINADEQUATE

blinding levelsADEQUATE

outlier handlingINADEQUATE

randomization unitADEQUATE

inclusion exclusionADEQUATE

randomization methodINADEQUATE

replicate distinctionN/A

independent replicationN/A

Per reviewer (3) — the votes behind the verdict

PASSReviewer 1· 70% conf

Randomization, stratification, blinding, and inclusion criteria are well described for this human RCT; the a priori power analysis is deferred to supplementary material rather than reported in the main text.

Evidence

“We used stratified random sampling via the Prolific platform to target a representative sample of the UK population in each group.”

MethodsFind in source

“Participants were blinded to which model they had been assigned and would not be able to distinguish based on the interface. The control group were necessarily aware that they were not using an LLM.”

MethodsFind in source

“For a power analysis and detailed demographics of the participants, as well as a breakdown of the results by sex, see .”

MethodsFind in source

PASSReviewer 2· 85% conf

The study uses a randomized between-subjects design with clear randomization, stratification, and a control group; blinding and power analysis are partially reported.

Evidence

“We used stratified random sampling via the Prolific platform to target a representative sample of the UK population in each group.”

MethodsFind in source

“Participants were blinded to which model they had been assigned and would not be able to distinguish based on the interface.”

MethodsFind in source

“For a power analysis and detailed demographics of the participants, as well as a breakdown of the results by sex, see .”

MethodsFind in source

PASSReviewer 3· 85% conf

The study employs a randomized design with stratification, clear inclusion/exclusion criteria, and an intention-to-treat analysis, but blinding is partial and power analysis is not detailed in the main text.

Evidence

“We used stratified random sampling via the Prolific platform to target a representative sample of the UK population in each group.”

MethodsFind in source

paraphrase[Methods, Participants]

“The paper states 'For a power analysis and detailed demographics of the participants... see .' but does not report the power analysis details in the main text.”

The main text does not report the sex of participants, age distribution, or health status. It mentions that a breakdown of results by sex is in the supplementary, but the main text itself lacks this information. The presurvey collected education, English fluency, etc., but results are not presented. This is a critical omission for a human subjects study.

Sub-criteria

demographicsNOT REPORTED

sex reportedNOT REPORTED

sex justifiedN/A

age weight healthNOT REPORTED

housing conditionsN/A

species strain sourceN/A

Per reviewer (3) — the votes behind the verdict

PASSReviewer 1· 60% conf

Demographics are stratified and reported in supplementary, and sex-stratified results are referenced; age is reported only as a threshold (18+), with no health/weight data.

Evidence

“We used stratified random sampling via the Prolific platform to target a representative sample of the UK population in each group.”

MethodsFind in source

“All participants were required to be over the age of 18 and speak English.”

MethodsFind in source

N/AReviewer 2· 90% conf

This is a human behavioral study; biological variables such as sex and age are reported as demographics, but the study does not involve biological specimens or animal subjects.

Evidence

“For a power analysis and detailed demographics of the participants, as well as a breakdown of the results by sex, see .”

MethodsFind in source

FAILReviewer 3· 85% conf

The paper does not report the sex, age distribution, or other demographic characteristics of participants in the main text, making it impossible to assess the biological variables.

Evidence

paraphrase[Methods, Participants]

“The paper says 'For a power analysis and detailed demographics of the participants, as well as a breakdown of the results by sex, see .' but does not report the sex distribution in the main text.”

paraphrase[Methods, Participants]

“The paper only states that participants were over 18, but does not provide the age distribution.”

paraphrase[Results, paragraph 1]

“The paper references supplementary tables for demographics but does not present them in the main text.”

The paper names the ethics committee (Oxford Internet Institute Departmental Research Ethics Committee) and provides a protocol number. Informed consent is explicitly stated. However, regulatory compliance is only mentioned as 'in accordance with the relevant guidelines and regulations' without naming a specific framework like the Declaration of Helsinki, which is inadequate.

Sub-criteria

iacuc statementN/A

informed consentADEQUATE

irb ethics statementADEQUATE

regulatory complianceINADEQUATE

Per reviewer (3) — the votes behind the verdict

WARNReviewer 1· 70% conf

Ethics approval by a named committee with a project number and informed consent are clearly reported; however, the regulatory-compliance statement is generic rather than naming a recognized framework.

Evidence

“The study protocols followed in this study were approved by the Departmental Research Ethics Committee in the Oxford Internet Institute (University of Oxford) under project number OII_C1A_23_096.”

MethodsFind in source

“Informed consent was obtained from all participants before enrollment in the study.”

MethodsFind in source

“Methods were carried out in accordance with the relevant guidelines and regulations.”

MethodsFind in source

PASSReviewer 2· 95% conf

Ethical approval from a named committee and informed consent are clearly reported.

Evidence

MethodsFind in source

“Informed consent was obtained from all participants before enrollment in the study.”

MethodsFind in source

WARNReviewer 3· 95% conf

Ethics approval and informed consent are reported, but regulatory compliance is stated only in a vague manner without naming a specific framework.

Evidence

MethodsFind in source

“Informed consent was obtained from all participants.”

MethodsFind in source

“Methods were carried out in accordance with the relevant guidelines and regulations.”

MethodsFind in source

The three LLMs (GPT-4o, Llama 3, Command R+) are identified with their providers. Statistical software (STATSMODELS v0.14.3, SCIPY v1.13.0, SEABORN v0.13.2) is named with versions. No antibodies, cell lines, or organisms are used, so those criteria are not applicable.

Sub-criteria

mycoplasma testingN/A

reagents identifiedN/A

organisms identifiedN/A

antibodies identifiedN/A

cell line authenticationN/A

software tools identifiedADEQUATE

Per reviewer (3) — the votes behind the verdict

PASSReviewer 1· 70% conf

The study's key software resources — the three LLMs and their API providers — are identified; bench-resource criteria are not applicable.

Evidence

“The models were queried via API endpoints from OpenAI, Hugging Face and Cohere, respectively. The hyperparameters and inference costs are listed in Supplementary Tables and .”

MethodsFind in source

PASSReviewer 2· 85% conf

The LLMs used are identified, and the software tools are named with versions; no biological reagents are used.

Evidence

“we selected three different leading LLMs: GPT-4o, chosen for its large user base as the model most likely to be used by the general public; Llama 3, selected for its open weights, most likely to be used as the backbone for creating specialized medical models; and Command R+, included for its use of retrieval-augmented generation”

MethodsFind in source

“All statistics were computed using the STATSMODELS v0.14.3 and SCIPY v1.13.0 packages in Python.”

Statistical analysisFind in source

WARNReviewer 3· 80% conf

The LLM models are identified but not fully specified (e.g., version numbers), and the software tools are well-documented, but the investigational product reporting is incomplete.

Evidence

paraphrase[Methods, Treatments]

“We selected three different leading LLMs: GPT-4o, chosen for its large user base... Llama 3, selected for its open weights... and Command R+, included for its use of retrieval-augmented generation.”

“All statistics were computed using the STATSMODELS v0.14.3 and SCIPY v1.13.0 packages in Python.”

Statistical analysisFind in source

paraphrase[Methods, Treatments]

“The paper states 'The hyperparameters and inference costs are listed in Supplementary Tables and .' but does not report them in the main text.”

The paper names chi-square tests, Mann-Whitney U tests, bootstrap, and linear regression. Effect sizes (odds ratios, mean differences) are reported with 95% CIs. Software versions are stated. However, some p-values are given as 'P < 0.001' without exact values, and assumptions for parametric tests are not discussed (though non-parametric tests are used). Data presentation includes means with CIs and per-group n, which is adequate for a large trial.

Sub-criteria

tests namedADEQUATE

exact p valuesINADEQUATE

effect sizes ciADEQUATE

data presentationADEQUATE

software identifiedADEQUATE

assumptions verifiedADEQUATE

mathematical plausibilityN/A

Per reviewer (3) — the votes behind the verdict

WARNReviewer 1· 60% conf

Tests, effect sizes with CIs, software, and data presentation are well reported, but several headline p-values are printed only as thresholds (P < 0.001) and assumption verification is not explicitly described.

Evidence

“participants using LLMs were significantly less likely than those in the control group to correctly identify at least one medical condition relevant to their scenario ( χ 2 (1), n 1 = n 2 = 600, P < 0.001 for all three models)”

ResultsFind in source

“Participants in the control group had 1.76 (95% CI = 1.45–2.13) times higher odds of identifying a relevant condition than the aggregate of the participants using LLMs.”

ResultsFind in source

“All statistics were computed using the STATSMODELS v0.14.3 and SCIPY v1.13.0 packages in Python.”

Statistical analysisFind in source

PASSReviewer 2· 80% conf

Statistical tests are named, exact p-values are reported, effect sizes with CIs are provided, and software is identified; some assumptions are not explicitly verified.

Evidence

“Comparisons between proportions were computed using χ 2 tests with 1 d.f., equivalent to a two-sided Z- test.”

Statistical analysisFind in source

“GPT-4o, χ 2 (1) = 0.17, P = 0.683; Llama 3, χ 2 (1) = 0.34, P = 0.560; Command R+, χ 2 (1) = 0.03, P = 0.861”

ResultsFind in source

“Participants in the control group had 1.76 (95% CI = 1.45–2.13) times higher odds of identifying a relevant condition than the aggregate of the participants using LLMs.”

ResultsFind in source

WARNReviewer 3· 90% conf

Statistical tests are named, effect sizes with CIs are reported, and software is identified, but some p-values are reported as thresholds (e.g., P < 0.001) rather than exact values, and assumptions are not explicitly verified.

Evidence

paraphrase[Methods, Statistical methods]

“Comparisons between proportions were computed using χ² tests with 1 d.f., equivalent to a two-sided Z-test. Two-sided Mann–Whitney U tests were used...”

paraphrase[Results, Experimental performance]

“GPT-4o, χ²(1) = 0.17, P = 0.683; Llama 3, χ²(1) = 0.34, P = 0.560; Command R+, χ²(1) = 0.03, P = 0.861; n1 = n2 = 600”

“Participants in the control group had 1.76 (95% CI = 1.45–2.13) times higher odds of identifying a relevant condition”

ResultsFind in source

The data availability statement provides concrete repositories (GitHub and Hugging Face) with URLs. Code is shared in a public GitHub repository. Accession numbers are not applicable as no sequencing data are involved.

Sub-criteria

code sharingADEQUATE

accession numbersN/A

repository depositADEQUATE

data availability statementADEQUATE

Per reviewer (3) — the votes behind the verdict

PASSReviewer 1· 90% conf

Data and code are openly shared via GitHub and Hugging Face repositories with concrete URLs; no sequencing/accession numbers apply.

Evidence

“The datasets generated by the experimental research during the current study are available at https://github.com/am-bean/HELPMed as well as https://huggingface.co/datasets/ambean/HELPMed/ .”

Data availabilityFind in source

“All code used to generate the analysis in the manuscript is shared by the authors for reuse and is available at https://github.com/am-bean/HELPMed .”

Data availabilityFind in source

PASSReviewer 2· 95% conf

Data and code are publicly available with clear links.

Evidence

“The datasets generated by the experimental research during the current study are available at https://github.com/am-bean/HELPMed as well as https://huggingface.co/datasets/ambean/HELPMed/ .”

Data availabilityFind in source

“All code used to generate the analysis in the manuscript is shared by the authors for reuse and is available at https://github.com/am-bean/HELPMed .”

Data availabilityFind in source

PASSReviewer 3· 95% conf

Data and code availability statements are provided with concrete repositories (GitHub and Hugging Face), meeting the criteria for open science.

Evidence

“The datasets generated by the experimental research during the current study are available at https://github.com/am-bean/HELPMed as well as https://huggingface.co/datasets/ambean/HELPMed/ .”

Data availabilityFind in source

“All code used to generate the analysis in the manuscript is shared by the authors for reuse and is available at https://github.com/am-bean/HELPMed .”

Data availabilityFind in source

The methods section is comprehensive, covering scenario development, participant recruitment, treatment allocation, and statistical analysis. The study is preregistered on OSF (https://osf.io/dt2p3). All primary and secondary outcomes are reported, including negative results. Limitations are discussed, including the use of vignettes and the possibility of improved future models. Conclusions are proportional to the evidence. A reporting guideline (e.g., CONSORT) is not referenced, and a conflict-of-interest statement is not in the main text (though referenced as available online).

Sub-criteria

funding coiNOT REPORTED

trial registrationADEQUATE

reporting guidelineNOT REPORTED

methods completenessADEQUATE

all outcomes reportedADEQUATE

limitations discussedADEQUATE

conclusions proportionalADEQUATE

Per reviewer (3) — the votes behind the verdict

PASSReviewer 1· 70% conf

Methods are detailed, null results are reported transparently, limitations are discussed, and funding/COI are addressed; no explicit reporting checklist (e.g., CONSORT) is referenced.

Evidence

“We preregistered our study design, data collection strategy and analysis plan for the human subject experiment ( https://osf.io/dt2p3 ).”

Methods ¶1Find in source

“Participants using LLMs did not have statistically significant differences in disposition accuracy from the control group”

ResultsFind in source

“Our scenarios focused on common conditions where users may be familiar with the symptoms, and results might differ on rare conditions or less typical presentations.”

Discussion ¶4Find in source

PASSReviewer 2· 85% conf

The study is preregistered, methods are detailed, limitations are discussed, and conclusions are proportional; a reporting guideline is not explicitly mentioned.

Evidence

“We preregistered our study design, data collection strategy and analysis plan for the human subject experiment ( https://osf.io/dt2p3 ).”

Methods ¶1Find in source

“Our work can only provide a lower bound on performance: newer models, models that make use of advanced techniques from chain of thought to reasoning tokens, or fine-tuned specialized models, are likely to provide higher performance on medical benchmarks.”

Discussion ¶4Find in source

“The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.”

AcknowledgementsFind in source

PASSReviewer 3· 90% conf

Methods are detailed, the study is preregistered, outcomes are reported, limitations are discussed, and conclusions are proportional. However, a reporting guideline is not mentioned and a COI statement is not present in the main text.

Evidence

“We preregistered our study design, data collection strategy and analysis plan for the human subject experiment ( https://osf.io/dt2p3 ).”

Methods ¶1Find in source

paraphrase[Discussion, paragraph 4]

“We used clinical vignettes for the interactions to limit risks to participants. This allowed us to focus on the challenges of LLMs interacting with members of the general public, without taking into account participants’ feelings of urgency and stress...”

paraphrase[End of main text]

“The paper states 'Competing interests; and statements of data and code availability are available at 10.1038/s41591-025-04074-y' but does not include a COI statement in the main text.”

Registration stated in text, but no registry ID was detected. No reporting guideline cited.

Broken references and links

None found

References checked against Crossref, OpenAlex and Retraction Watch for retractions and resolvability, plus declared data and code links probed for whether they resolve to content matching the paper.

Checked — nothing surfaced.

Checked 37 references by DOI: 16 verified — 21 no DOI (shown, not verified).

NO DOI
ChatGPT diagnoses cause of child’s chronic pain after 17 doctors failed
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
KFF Health Misinformation Tracking Poll: Artificial Intelligence and Health Information
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
Combining Human Expertise with Artificial Intelligence: Experimental Evidence from Radiology
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
Superhuman performance of a large language model on the reasoning tasks of a physician
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
The robot doctor will see you now
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
Digital access – a ‘front door to the NHS’
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
AI opportunities action plan
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
Catalyzing equitable artificial intelligence (AI) use to improve global health
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
HHS releases strategic plan for the use of artificial intelligence to enhance and protect the health and well-being of Americans
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
Large language models encode clinical knowledge
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
Capabilities of GPT-4 on medical challenge problems
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
AgentClinic: A multimodal agent benchmark to evaluate AI in simulated clinical environments
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
An evaluation framework for clinical use of large language models in patient interaction tasks
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
Influence of believed AI involvement on the perception of digital medical advice
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
Misattribution of error origination: the impact of preconceived expectations in co-operative online games
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
Capabilities of Gemini models in medicine
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
Towards conversational diagnostic artificial intelligence
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
Towards interactive evaluations for interaction harms in human-AI systems
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
Learning to complement humans
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
Dynabench: rethinking benchmarking in NLP
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).
NO DOI
NICE guidance
No DOI in the reference — shown for manual review; not independently verifiable (not a fabrication signal).

4 data/code links checked; 4 live.

codeGitHubLIVEHTTP 200
https://github.com/am-bean/HELPMed
Resolves to GitHub (code repository).
dataHugging FaceLIVEHTTP 200
https://huggingface.co/datasets/ambean/HELPMed/
Resolves to Hugging Face (data repository).
dataHugging FaceLIVEHTTP 200
https://huggingface.co/datasets/ambean/HELPMed/viewer/default/scenarios
Resolves to Hugging Face (data repository).
dataOSFLIVEHTTP 200
https://osf.io/dt2p3
Resolves to OSF (data repository).

Copyediting

9 minor

Wording, consistency and formatting errors that need correcting before submission.

No major wording or formatting errors. 9 minor suggestions below.

9 copyedit issues flagged: mostly typo, consistency, clarity.

MINORtypoExtended Data Table 2 caption
“In both of these converstaions, the model is GPT-4o”
→ In both of these conversations, the model is GPT-4o
Misspelling of 'conversations'.
MINORclarityDiscussion, paragraph 6
“Developers of general-purpose LLM platforms may have an incentive to design risk-adverse LLMs”
→ risk-averse LLMs
'Risk-adverse' should be 'risk-averse'.
MINORotherEnd of reference list
“rstand participant health behavior. Res. Soc. Adm. Pharm. 17 , 2070 (2021).”
→ Complete or remove the truncated reference fragment
The trailing text appears to be a torn/truncated reference fragment.
MINORconsistencyExtended Data Table 4
“GPT-4o: f = . 755, p < 0.0001”
→ f = 0.755
Inconsistent formatting of effect-size values (leading zero and spacing) across the table.
MINORtypoExtended Data Table 2
“converstaions”
→ conversations
Typo in the table description.
MINORconsistencyResults, Experimental performance
“χ 2 (1), n 1 = n 2 = 600, P < 0.001 for all three models”
→ Ensure consistent formatting of chi-square statistics across the paper.
Chi-square notation varies slightly (e.g., 'χ 2 (1)' vs 'χ 2 (1) = 0.17').
MINORclarityMethods, Participants
“For a power analysis and detailed demographics of the participants, as well as a breakdown of the results by sex, see .”
→ Provide a specific reference to the supplementary section or table.
The sentence ends with a period but lacks a reference to the supplementary material.
MINORtypoExtended Data Table 2 caption
“converstaions”
→ conversations
Typo in 'conversations'
MINORconsistencyMethods, Participants
“For a power analysis and detailed demographics of the participants, as well as a breakdown of the results by sex, see .”
→ Replace 'see .' with a specific reference to the supplementary table or section.
Placeholder with period; likely missing reference.

Kaimen Rigor Action ItemsFinal pass before journal submission

The published work is methodologically solid and largely reproducible, but an informed reader should weigh the reporting gaps: the primary condition-identification comparisons are reported only as P < 0.001 thresholds, several in-text references to supplementary material are broken placeholders (power analysis, demographics, sex breakdown), the ethics compliance statement names no recognized framework, and age/health demographics are absent from the main text. These issues are not validity threats on their own — no retractions, no incorrect recomputed statistics, all links live — but a correction/erratum restoring the broken references and exact p-values, plus a statement of statistical assumption handling, would strengthen the record and support independent re-analysis.

1.
HIGHreporting
Fix the broken 'see .' placeholders in Methods, Participants (power analysis, detailed demographics, sex breakdown) by pointing to the specific supplementary/Extended Data tables.
Readers currently cannot locate the referenced power analysis or demographic/sex-stratified data, which undercuts both the power_analysis and biological_variables reporting.
2.
HIGHstatistics
Replace the threshold p-values (P < 0.001) on the headline condition-identification comparisons in Results, Experimental performance with exact p-values.
Threshold-only reporting of a computed p-value is imprecise reporting that a critical reader will flag; exact values are needed for the primary outcome.
3.
HIGHreporting
Complete or remove the truncated reference fragment at the end of the reference list ('rstand participant health behavior. Res. Soc. Adm. Pharm. 17 , 2070 (2021).').
A torn/truncated reference is a citation-integrity red flag and an unresolved fabrication signal that must be fixed or removed.
4.
MEDIUMethics
Name a recognized regulatory/compliance framework (e.g., Declaration of Helsinki) in Methods, Participants, replacing the generic 'relevant guidelines and regulations' statement.
A named framework is the standard for human-subjects research and was flagged by two reviewers as inadequate.
5.
MEDIUMstatistics
Add an explicit statement in Methods, Statistical methods on how statistical assumptions were handled (justify the use of non-parametric Mann-Whitney tests in lieu of normality/equal-variance testing).
Assumption handling is not described, and two reviewers rated it reported_but_inadequate.
6.
MEDIUMreporting
Report the a priori power analysis parameters (effect size, alpha, power) in the main text Methods rather than only referencing supplementary.
The power analysis is a key design element that reviewers expect in the main text for a confirmatory RCT.
7.
MEDIUMreporting
Report the participants' age distribution (and any health/comorbidity summary) in the main text or in a clearly referenced supplementary table.
Only an 18+ threshold is given and the reference is broken, leaving the demographic profile incomplete for a human-subjects study.
8.
MEDIUMreporting
Reconcile the abstract statement 'fewer than 34.5% of cases' with the Results ranges (0.34–0.43 for Command R+, 0.42–0.54 for GPT-4o).
The abstract's single percentage is not directly comparable to the reported per-model ranges and may mislead readers.
9.
MEDIUMother
Clarify in Methods, Participants how the 98 participants replaced due to API issues and 13 due to software errors relate to the final n=1,298 and the 2,400 responses.
The total participant count and response count should reconcile with the reported replacement counts; the current explanation is incomplete.
10.
MEDIUMreporting
Reference the CONSORT checklist (or state the contents of the Nature Reporting Summary) for this randomized trial in the main text.
A named reporting guideline improves transparency for a human RCT and was flagged by all three reviewers as missing.
11.
LOWcopyedit
Fix the typo 'converstaions' → 'conversations' in the Extended Data Table 2 caption, and 'risk-adverse' → 'risk-averse' in Discussion, paragraph 6.
Minor typos that should be corrected for a clean published record.
12.
LOWcopyedit
Standardize chi-square notation formatting across Results (e.g., 'χ 2 (1)' vs 'χ 2 (1) = 0.17') and fix the inconsistent effect-size formatting in Extended Data Table 4 (e.g., 'f = . 755' → 'f = 0.755').
Inconsistent notation across tables and text reduces readability and precision.
13.
LOWdata code
Add a DOI or persistent identifier for the GitHub repository to guarantee permanent access to data and code.
A persistent identifier protects long-term access beyond the current URL.
14.
LOWreporting
Specify the randomization method (e.g., random number generator / block randomization) and the unit of randomization, and clarify whether outcome assessors (scorers) were blinded to treatment group.
These methodological details were flagged by Reviewer 2 and would strengthen the design description.

Full text (55,547 chars)Analyzed: Aug 10, 2026

RigorReviewer 1

RigorReviewer 2

RigorReviewer 3

RigorPre-Submission Reviewer

CopyeditAgent Scribe

CitationsCitation Integrity

StatisticsStatistics Consistency

ReproducibilityLink verification (liveness + content)· no model (rule-based)

PreregistrationDeterministic check· no model (rule-based)

The star rating is the report’s one-glance summary. Every paper starts at 5★ and loses stars for the concrete problems the review finds — so a rating is never a vague average, it’s a running total you can read line by line under “How this rating was calculated.”

How stars are deducted

Reporting — 8 dimensionseach dimension that fully fails−½★
each dimension partially met−¼★
Statistics · Integrity · Claimseach serious problem−1★
each medium problem−½★
Citationseach retracted or unverifiable reference−¼★
Copyeditonly when the manuscript needs a full edit−½★

The rating never drops below 1★, and a demonstrable critical failure (an impossible statistic, a proven ethics violation) caps it at 1★ on its own — so the stars can never look healthy when the verdict is CRITICAL.

The rating draws on a panel of agents. Three independent Kaimen Rigor reviewers grade the eight dimensions below across several independent passes (the shown verdict is their majority vote — steadier than any single run), isolate the paper’s major claims and check its own evidence backs them, and flag integrity concerns. Alongside them, a citation agent resolves every reference against Crossref, OpenAlex, and Retraction Watch; a statistics agent recomputes reported tests; and rule-based checks verify that declared data/code links actually resolve. Full text is required — an abstract-only submission is not analyzed.

The eight reporting dimensions

Scientific PremisePrior work, rationale, and how it addresses gaps.

Study Design & RigorRandomization, blinding, power, controls, replication.

Biological VariablesSex, age, species/strain, housing, demographics.

Ethical ApprovalsIRB/IACUC approval, informed consent, regulatory compliance.

Key ResourcesAntibodies, cell lines, organisms, software — identified with persistent IDs.

Statistical ReportingNamed tests, exact p-values, effect sizes, plausible summary stats.

Data & Code AvailabilityAvailability statements, repositories, accession numbers, shared code.

Reporting TransparencyMethods, registration, guidelines, outcomes, limitations, funding/COI.

Graded against NIH, MDAR, ARRIVE 2.0, CONSORT, EQUATOR, and RRID guidelines. A dimension that doesn’t apply to the study type is skipped, never penalized.

This Kaimen Rigor review is model-assisted and is not a substitute for formal expert review. It complements human evaluation by surfacing potential methodological concerns — verify each finding against the source.

Reliability of LLMs as medical assistants for the general public: a randomized preregistered study.

Bean AM, Payne RE, Parsons G, Kirk HR, Ciro J, Mosquera-Gómez R, Hincapié M S, Ekanayaka AS, Tarassenko L, Rocher L, Mahdi A

DOI

10.1038/s41591-025-04074-y

Record issued

2026-08-10

Engine

7.29.0

Exported

2026-09-22