Does Generative AI Advice Improve Expert Judgment?

Wang et al. systematically searched for clinical collaboration studies between humans and LLMs published by June 2025, using 10 peer-reviewed studies for the main analysis and 3 preprints for sensitivity analysis. The pooled difference for composite scores of diagnosis and management reported in two trials showed that human+AI outperformed human alone by 4.88 points. However, the 95% prediction interval was wide, ranging from -31.65 to 41.42 points, leaving open the possibility that benefits may not appear in other clinical environments. [1]

In the same review, diagnostic and interpretive accuracy improved in direction but was statistically uncertain, and there was no overall difference in time across three pooled studies. Even in studies where document quality improved, factual errors remained at approximately 26% to 36%. [1] There is no need to deny the observed improvements, but they are not evidence that can be compressed into the statement 'AI advice is effective in medicine.' Implementation conditions, including tasks, interface, user training, and verification steps, alter results.

In a 2024 trial by Goh et al., 50 physicians in internal medicine, family medicine, and emergency medicine were randomized to a group using GPT-4 and standard materials, or a group using standard materials only, to answer up to 6 diagnostic cases. The median diagnostic reasoning score per case was 76% vs. 74%, with an adjusted difference of 2 points (95% CI: -4 to 8 points, P = 0.60). There was no significant difference in time. [3] Under these conditions, LLM access did not improve diagnostic reasoning.

In contrast, in a 2025 trial, 92 physicians answered 5 management cases based on real patients, and the GPT-4 group scored 6.5 points higher on the composite management reasoning score than the standard materials group. This improvement occurred in open-ended tasks including diagnosis, patient background, tests, and treatment prioritization. However, the LLM group took an average of 119.3 seconds longer per case, so quality improvement cannot be equated with speed improvement. [2]

In 2026, a trial was also reported recruiting 60 licensed physicians in Pakistan, randomized after all participants received 20 hours of AI literacy education. Among the 58 who completed it, the group using GPT-4o and standard materials outperformed the standard materials-only group by 27.5 points in adjusted diagnostic reasoning, with no difference in time. [4] However, both groups received the same training, and there was no group using LLMs without training. Therefore, one cannot attribute all this difference to the 20-hour training, nor explain the difference from Goh's US trial solely by training.

These trials differ in participants, cases, evaluation metrics, and pre-use education, so differences cannot be compared as a single average effect of LLMs. Nevertheless, it is important that results varied with the same type of expert and similar models. The unit of adoption must be the 'judgment task' rather than 'profession'.

• Classification: Diagnosis narrowing answer candidates vs. management judgment adjusting multiple constraints
• Input: Whether necessary facts are complete or additional information retrieval is needed
• Verification: Whether users are given the basis and time to detect incorrect proposals
• Loss: Whether incorrect answers are detectable and decisions can be reversed afterward

In the 2024 diagnostic trial, exploratory LLM-only evaluation was 16 points higher than the standard materials group, but giving LLMs to physicians yielded no significant improvement. [3] In the 2025 management reasoning trial, the difference between LLM-only 43.7% and human+LLM 43.0% was not significant. [2] Users cannot simply absorb the capabilities held by the model.

These trials alone cannot identify the mechanisms hindering improvement. It must be verified separately whether humans rejected correct AI proposals, adopted incorrect ones, or whether question phrasing or interface presentation was insufficient. Therefore, in adoption evaluations, record not only final accuracy but also proposal acceptance/rejection, correction reasons, incorrect answer detection, and verification time.

• Measure the three conditions: human only, AI only, and human+AI to the extent possible
• Record AI proposal adoption, correction, and rejection with reasons
• Separate the adoption rate of correct proposals from the rejection rate of incorrect proposals
• Have separate evaluators confirm the quality of the completed text and the accuracy of the basis

Meta-analyses of clinical collaboration leave wide prediction intervals, and other trials did not confirm significant improvement in diagnostic reasoning from LLM access. [1][3] These do not immediately mean 'LLM advice is worthless' or 'equivalent to humans.' One must distinguish between states where benefits could not be detected, states where meaningful differences were excluded, and states where transfer to the field is uncertain.

In practice, set minimum necessary quality improvement, acceptable time increase, and upper limits for critical errors before the trial. If results are in the direction of improvement but uncertainty crosses the threshold, classify as undetermined rather than 'promising'; if the difference is small but exceeds the upper limit for critical errors, reject adoption. Record in the review sheet not only average scores but which conclusions the evidence supports and which cannot yet be excluded.

The recommended evidence states are: superior, conditionally acceptable within tolerance, uncertain, concern of harm, and incomparable. Tolerance is defined by the business side before the trial and is not used in the sense of demonstrating statistical non-inferiority. This is not a classification defined by research, but an operational proposal to prevent misconversion of different trial results into adoption decisions. If cases, users, evaluation metrics, and education conditions differ from your own, retain transferability in a separate column even if results are clear, and return to comparison on representative tasks.

For the same incorrect answer, the loss differs between a missed candidate that can be reversed with additional verification and a final judgment that is difficult to correct later. In evaluating expert AI, do not apply uniform weight to accuracy; instead, record the type of error, detectability, reversibility, and affected parties for each case. This is not a new estimate of clinical effect, but a loss design defined by the adopter.

Hypothetical example. If the number of errors where AI recommends unnecessary additional tests equals the number of errors where AI drops necessary responses from candidates, the verification paths and impacts are not the same. Even if humans can correct, record the information and time required for correction. By recording how far incorrect advice progressed and what stopped it, rather than assuming harmlessness because accuracy was reached, determine where to change the interface, education, and scope of application.

In recommended comparisons, view human-only and human+AI not only by final scores but by errors, correction rates, and verification time by loss category. Exclude from the advice scope categories where loss is large and hard to verify, rather than offsetting them with average improvement. Loss weights are not borrowed from research papers; they are agreed upon in advance by business and quality managers, and version-controlled to prevent convenient changes after seeing results.

• Error type: Distinguish misses, over-proposals, ranking errors, and basis inconsistencies
• Loss conditions: Record affected parties, detectability, reversibility, and correction deadlines
• Defense path: Link who stopped incorrect advice and on what basis to each case
• Decision rules: Pre-define categories not offset by average scores and exclusions from application

A role definition where experts make final approval does not guarantee judgment quality. For high-impact cases, combine provisional judgment before viewing AI, questions prompting falsification, direct links to basis, displays showing uncertainty rather than confidence, and independent review. Also clarify the scope of requests to AI: candidate generation, basis search, comparison table creation, or final recommendation.

Effectiveness determination should be based on judgment quality, loss upon incorrect answers, verification time, rework, and variation by user, not a single average score. Do not translate research averages into ROI; fix your company's representative tasks, compare small under the same quality criteria, and expand only the replicated scope. Generative AI advice is not a feature that works upon adoption, but an intervention requiring continuous experimentation with work combinations.

The workflow owner records the representative task set, advice interface, and model version as identifiers in the adoption record, with a minimum for decision quality and maxima for severe errors, verification time, and rework as acceptance criteria. A task that misses a minimum or exceeds a maximum returns to the human procedure. Restart requires the quality owner to approve two observation windows meeting every criterion after the causal handoff is corrected.