Last week I read a new consult series from a major professional Ob society. It carries the endorsement of a second society. Thirty-five pages. Two hundred eighty-four references. The kind of document that tells doctors what to do on Monday morning.

I did not set out to audit it.

Something in one of the radiation tables refused to sit still. Two tables sat side by side, under identical headers, and disagreed with each other. When I checked them row by row, they were offset by exactly two weeks. One counted from conception. The other counted from the last menstrual period. Nobody had noticed.

So I built a prompt and pointed it at the whole document. It came back with nineteen items.

A steroid dose conversion that used two different ratios in the same sentence. They cannot both be right, and neither one is. A recommended waiting time between chemotherapy and delivery that matched exactly the waiting time its own cited study had identified as the worst one. A cancer stage that cannot exist under the current staging system. A study cited as measuring time from one event when it actually measured time from a different one, in a direction that reverses the advice a woman would be given.

Not one of these was a mistake about obstetrics.

Every single one sat on a border with another field. Radiation physics. Blood platelet thresholds. Drug potency. Cancer staging. Lymphoma staging.

That is where documents written by one specialty about five others break.

Why I am writing about prompting instead

I could not have done this three years ago. I could not have done it last year, not this well. What changed is not that I got better at reading. What changed is the machines, and almost everything doctors have been taught about how to talk to them is now out of date.

The old advice was to learn tricks.

Give the model a persona.

Show it examples.

Tell it to think step by step.

In 2023 that advice was correct.

A Microsoft team showed that a stack of prompting tricks, which they called Medprompt, pushed a general model past specialist medical models on nine medical benchmarks and above 90 percent on the medical licensing question set for the first time. Note that this work is a preprint and has not been peer reviewed. (1)

Then the reasoning models arrived, and the same team found the opposite. With the newer models, adding examples to the prompt made performance worse, not better. Their words: in-context learning may no longer be an effective steering approach. Also a preprint. (2)

Peer-reviewed work now says the same thing. A 2025 study compared several chain-of-thought prompting methods across medical question sets and found no significant difference between them. The authors concluded that complex prompting techniques do not significantly improve performance compared with simpler approaches. (3) Another group tested prompting strategies across three models on clinical decision tasks and found the effect went in both directions. It helped the weakest task and was counterproductive for others. That one is also a preprint. (4)

So the tricks are fading. Three things replaced them, and they matter more than any trick ever did.

What actually works now

1. The first is telling the model what job it has and what the answer should look like. Not clever wording. Plain specification. The reasoning models do their own thinking; what they need from you is the task, the constraints, and the format.

2. The second is telling it to prioritize safety, out loud, in the prompt. This is not decoration. In a 2025 study of three reasoning-capable models across clinical scenarios graded by six expert clinicians, critical safety problems appeared in 12 of every 100 responses overall. A safety-first instruction cut those from about 16 in 100 down to about 9 in 100. That is a 45 percent reduction from one sentence. The same study found that even the best prompting left the models poor at communication and empathy, at roughly half of the maximum score, and concluded that current prompt engineering gives only marginal improvement, not enough for reliable clinical use. (5)

3. The third is verification. Tell the model to check its own claims against the original source, to quote what it found, and to mark plainly what it could not verify. This is the whole reason my audit prompt works. It is not clever. It forces the model to label every finding as proven from the document, proven against a retrieved source, or unverified. The unverified ones I take to the authors as questions, not accusations.

4. The fourth (optional) one is to enter the finding into a different AI model and ask it to verify the results of the first one

5. One practical note before you paste it. The free tier of most of these services gives you a smaller model, less thinking time, and a shorter memory, and none of that is enough for a thirty-five page document with 284 references. The gap is measurable: on board-style medical questions this year, models that reason step by step scored 81.5 percent against 75.8 percent for those that do not, and in an oncology text-mining study the same model's accuracy rose from 0.84 to 0.94 simply by being allowed to think longer. If you are going to check a document that tells doctors what to do on Monday, pay for the better model.

The failure nobody warns you about

Here is the most important study of the last year, and it is not about prompting at all.

Researchers gave 1,298 people ten medical scenarios and asked them to work out the likely condition and what to do. Some got help from a large language model. Some used whatever source they liked. Tested on its own, the model identified the right condition 94.9 percent of the time. The people using that same model identified it fewer than 34.5 percent of the time, which was no better than the people with no model at all. (6)

Read that again. The machine knew. The human did not get it out.

The authors located the failure precisely. Their words: the transmission of information between the model and the user is a particular point of failure. People gave the model incomplete information, and the model produced the right answer but did not get it across. They also noted that standard medical benchmarks did not predict this at all. (6)

There is a second version of this problem aimed straight at us. A 2026 evaluation across more than 5,500 medical questions found that models readily abandoned correct reasoning when they were given misleading hints attributed to an expert physician. (7) So the model can be talked out of a right answer by a confident doctor. If you use one of these tools and push back on it because it disagrees with you, understand what you may be doing.

Subscribe now

What this means for clinicians and patients

Do not treat the tool as the doctor. Treat it as the reader who never gets tired at reference number 140. That is a real job and we are demonstrably worse at it.

Be honest that it is not uniformly better. When one group asked models to agree or disagree with an orthopedic guideline, the best prompt reached only 62.9 percent agreement overall. (8) When another group ran 300 manuscripts past both models and 324 ophthalmologists, the humans rejected 73 of every 100 papers and the models rejected 2, repeating the same three generic complaints and omitting line numbers and references. Those authors concluded the tools should not be used for manuscript revision in their current state. (9) Both findings are real and I am not going to hide them.

But in the other direction, a blinded cardiology study this year found model-generated reviews scored higher than human reviews in five of seven quality domains, agreed with the final editorial decision about as often, and took 2 to 6 minutes against a median human turnaround of 17 days. (10) A doctor who dismisses all of this is not being careful. He is being comfortable.

Share

For a patient, the practical message is short. If you ask an AI about your symptoms, the answer you receive is probably worse than the answer the machine actually had, because the handoff is where it breaks. Tell it everything. Ask it what it would need to know to be more certain. Then take what it says to a human.

Conclusion

The prompting advice circulating right now is a year or two behind the models it describes. Personas and clever phrasing are not where the value is anymore. Specification, an explicit safety instruction, and forced verification are, and the largest single source of error is not the machine at all. It is the space between the machine and the person using it.

For our profession the conclusion is sharper. A guideline endorsed by two societies, reviewed at multiple levels, went to press with a dose conversion that fails its own arithmetic and a cancer stage that cannot exist. Nobody checked the numbers against the sources, because by hand nobody can. That is no longer an acceptable reason. Every professional society document should be machine-checked for arithmetic, units, and citation fidelity before publication. Not machine-written. Machine-checked. The two are not the same thing and the distinction is the entire argument.

I am not asking anyone to trust the output. I checked every finding myself, and where a paper sat behind a paywall I could not open, I marked the item unverified instead of claiming it. That is the division of labour that works.

The prompt itself is below. Paid subscribers read on.

Read more

For most of the AI era, medicine has protected itself with a reassuring formulation:

AI will assist physicians. It will not replace physician judgment.

The American Medical Association even prefers the term “augmented intelligence,” emphasizing that AI should enhance, rather than supplant, human clinicians. The American College of Physicians similarly argues that AI should support clinical decision-making rather than replace it.

That position is increasingly difficult to defend as a scientific principle.

In a provocative new JAMA Perspective, Ezekiel Emanuel and colleagues argue that autonomous AI may ultimately provide better cognitive medical care than physicians, including physicians working with AI.

For obstetrics and gynecology, we should take that possibility very seriously.

I suspect they may be right.

We are asking the wrong question

The usual question is:

Will AI replace the ObGyn?

That is too simplistic.

A better question is:

For which parts of being an ObGyn will AI eventually make better decisions than the ObGyn?

And the answer may be: a very large part of what we currently call clinical judgment.

Consider what fills much of our cognitive work:

• obtaining and organizing a history;

• calculating gestational age;

• interpreting laboratory values;

• generating differential diagnoses;

• estimating risk;

• choosing diagnostic tests;

• selecting medications;

• applying guidelines;

• recognizing contraindications;

• integrating multiple comorbidities;

• reviewing longitudinal records;

• predicting complications;

• counseling about probabilities.

These are precisely the kinds of tasks at which modern AI systems are improving extraordinarily quickly.

Emanuel and colleagues summarize evidence suggesting that AI already rivals or exceeds physicians in five fundamental cognitive domains: eliciting clinical information, diagnosis, test selection, guideline-concordant treatment, and chronic disease management.

That matters enormously for ObGyn.

Imagine the autonomous AI obstetrician

A sufficiently capable obstetric AI could simultaneously know:

every blood pressure during pregnancy;

every laboratory result;

every ultrasound measurement;

the fetal growth trajectory;

every medication and potential interaction;

the patient’s prior pregnancies;

every applicable guideline;

the patient’s cesarean probability;

her hemorrhage risk;

her preeclampsia risk;

her venous thromboembolism risk;

and every change in fetal status occurring during labor.

It would not become tired during a 24-hour call.

It would not forget whether the creatinine was 0.8 or 1.2.

It would not confuse one patient with another.

It would not rely on a vaguely remembered Practice Bulletin.

It could continuously update risk as new information became available.

Most importantly, it could compare the current patient with an amount of medical knowledge and clinical data that no individual physician could possibly retain.

That does not mean today’s LLM can safely run a labor floor.

It means we should stop assuming that human cognition is the ceiling against which medical intelligence must be measured.

It probably is not.

The physician in the loop may not always improve care

This is the most uncomfortable part of Emanuel and colleagues’ argument.

We have generally assumed that the safest configuration is:

AI recommendation → physician review → physician decision.

But there is no law of nature saying the physician improves the answer.

A meta-analysis cited in the Perspective found that when humans outperform AI, combining humans with AI can improve performance. But when AI already outperforms humans, the human-AI combination can perform worse than AI alone.

That makes intuitive sense.

Imagine an AI correctly recognizes evolving severe preeclampsia and recommends delivery.

The physician says, “She looks fine. Let’s wait.”

Or the AI identifies subtle fetal deterioration.

The physician dismisses it because the tracing “doesn’t look that bad.”

The human has not added safety.

The human has introduced error.

Medicine tends to describe this as “clinical judgment.”

Sometimes it is.

Sometimes it is simply wrong judgment delivered with confidence.

Obstetrics may be particularly suited to AI superiority

Obstetrics contains an extraordinary mixture of structured data and high-stakes decisions.

• Gestational age.

• Blood pressure.

• Proteinuria.

• Platelets.

• AST and ALT.

• Estimated fetal weight.

• Amniotic fluid.

• Cervical dilation.

• Contraction frequency.

• Fetal heart rate.

• Prior cesarean history.

• Maternal age.

• BMI.

• Diabetes.

• Hypertension.

These variables interact across time.

Humans are not especially good at continuously integrating dozens or hundreds of changing variables.

Computers are.

An autonomous obstetric AI could eventually perform continuous surveillance rather than episodic evaluation.

That is a fundamentally different model of care.

Today, the obstetrician reviews the chart, examines the patient, interprets the fetal tracing, and makes a decision.

Tomorrow, the AI may have been analyzing the pregnancy continuously for months and detecting trajectories that no clinician ever consciously noticed.

The physician may increasingly become the person who acts on the intelligence rather than generates all of it.

And gynecology?

The same transformation is easy to imagine.

• Abnormal bleeding.

• Adnexal masses.

• Infertility.

• Menopause.

• Contraceptive counseling.

• Cervical screening abnormalities.

• Hereditary cancer risk.

• Medication selection.

• Longitudinal surveillance.

Much of outpatient gynecology involves integrating history, laboratory results, imaging, guidelines, risk estimates, and patient preferences.

An autonomous system with complete access to the patient’s longitudinal record could potentially perform much of that cognitive synthesis more consistently than an individual clinician seeing the patient for 15 minutes.

That is not science fiction.

The remaining question is how quickly reliability, regulation, liability, data integration, and real-world validation catch up with capability.

But AI cannot deliver a baby

Correct.

Emanuel and colleagues explicitly acknowledge that procedures remain a major boundary. Surgery, delivery, interventional radiology, colonoscopy, and physical examination still require human clinicians because robotics cannot yet replace those functions in most settings.

But this is not an argument against autonomous medical intelligence.

It simply separates thinking from doing.

The AI may decide that cesarean delivery is indicated.

The obstetrician performs it.

The AI may identify postpartum hemorrhage and recommend the optimal sequence of uterotonics, tranexamic acid, transfusion, balloon tamponade, and escalation.

Humans carry out those interventions.

The fact that a pilot does not personally calculate every navigational parameter does not mean the aircraft lacks automation.

Medicine may evolve in the same direction.

The hardest objection is not technical. It is professional.

Physicians have built professional identity around judgment.

We perform procedures, certainly.

But what distinguishes the physician has traditionally been the claim:

I know what should be done.

If a machine can know what should be done more accurately, more consistently, and with access to vastly more information, that threatens something deeper than employment.

It threatens professional hierarchy.

Yet preserving physician authority cannot be the objective of medical ethics.

The objective is better patient care.

If autonomous AI eventually produces fewer diagnostic errors, fewer treatment errors, more guideline-concordant care, better risk prediction, and better outcomes than physicians, then insisting that a physician retain final control merely because the decision is “medical” would become difficult to justify ethically.

Professional authority is not an entitlement.

It must be earned by superior performance.

There are still important limitations

The case is not yet proven.

The authors appropriately acknowledge that much of the evidence comes from simulated tasks rather than actual autonomous clinical practice. Real patient communication remains a potential failure point. AI can be brittle. Cybersecurity, hallucinations, infrastructure failure, liability, reimbursement, and regulation remain unresolved.

These limitations are substantial.

But they should affect how autonomous AI is introduced, not whether we permit ourselves to consider that it may ultimately outperform us.

The appropriate response is rigorous prospective testing.

If autonomous AI performs worse, do not deploy it.

If physician plus AI performs best, use that model.

But if autonomous AI repeatedly produces better outcomes than physician-controlled care, medicine must be willing to accept that result too.

The ObGyn of the future may have a very different job

Physicians will not disappear.

Obstetricians will still operate, deliver babies, manage emergencies, examine patients, communicate devastating news, navigate values and preferences, and take responsibility in circumstances in which algorithms cannot act physically.

But the assumption that the physician must remain the supreme cognitive authority may not survive.

By 2030, Emanuel and colleagues argue, autonomous AI could be ready for real-world deployment in at least some cognitive medical workflows.

In ObGyn, I think we should prepare for something even more consequential:

The safest obstetrician may eventually be an excellent proceduralist working alongside an autonomous intelligence that makes many cognitive decisions better than any individual physician can.

And eventually, for some areas of care, the question may no longer be:

Should the AI be allowed to make this decision without the doctor?

It may become:

If the AI consistently makes this decision better than the doctor, what is our justification for allowing the doctor to overrule it?

That is the question our profession needs to start confronting now.

Reference

1. Emanuel EJ, Baker-Butler A, Khosla N, Khosla V. Will autonomous AI exceed AI-aided physicians as the best medical care? JAMA. Published online August 17, 2026. doi:10.1001/jama.2026.15380.

Artificial intelligence is rapidly becoming part of scientific writing.

Researchers use it to organize ideas, summarize literature, improve grammar, shorten abstracts, suggest titles, restructure discussions, and convert rough scientific language into polished prose.

These applications can save time. They can also make a weak manuscript sound more convincing than it deserves to sound.

That is the central danger.

Subscribe now

A fluent manuscript is not necessarily a valid manuscript.

Clear language can conceal weak methods. Elegant paragraphs can connect claims that the data do not actually support. A well-structured discussion can make an association sound causal. A polished conclusion can quietly move beyond the study population, the measured outcome, or the limits of the design.

AI is exceptionally good at helping an author make an argument coherent.

It may therefore be equally important to use AI for the opposite purpose.

After the manuscript has been written and revised, I give it to a separate model and tell it:

Your task is not to improve this manuscript. Your task is to find what is wrong with it.

I call this a structured adversarial manuscript review.

The term is not yet standardized in biomedical publishing. Related work has examined generative reviewer agents, automated paper reviewing, multi-agent critique, and the vulnerability of AI reviewers to adversarial manipulation. The evidence is developing, but it does not establish that AI review can replace expert methodological, statistical, editorial, or clinical judgment.

That is not what I am proposing.

The adversarial review is not a substitute for peer review.

It is a pre-submission stress test.

Its purpose is to expose defects while the authors can still correct them.

Why Ordinary AI Criticism Is Usually Too Weak

Ask a language model:

“Please review this manuscript.”

The answer will often begin with praise.

The topic is important.
The manuscript is well organized.
The methods are generally appropriate.
The findings are potentially valuable.

Then come several broad suggestions:

• clarify the methods;

• expand the limitations;

• strengthen the discussion;

• improve the conclusion;

• consider additional references.

This may sound professional, but it often has little value.

It resembles the weakest form of human peer review. It is polite, generic, and difficult to act on.

The model has interpreted its role as cooperative. It is trying to help the author complete the paper, not determine whether the paper deserves to survive scrutiny.

A useful adversarial prompt must change the model’s assignment.

It must explicitly prohibit praise, politeness padding, and unsupported reassurance. It must demand itemized findings. It must separate errors that can be proved from the manuscript itself from claims that require external verification. It must force the model to identify the exact text, table, figure, calculation, or conclusion at issue.

Most importantly, it must tell the model that coverage matters more than diplomacy.

The first pass should find as many plausible defects as possible.

The authors can then decide which findings are valid.

Share

Why the Review Should Be Independent

There is a practical reason to use a different model, or at least a fresh session with no drafting history.

The model that helped write the manuscript has participated in constructing the argument. It has repeatedly been rewarded for making the argument work. It may inherit the assumptions, definitions, and framing that shaped the draft.

An independent reviewer begins without that commitment.

This is not true methodological triangulation. Triangulation usually means examining a finding across different data sources, methods, investigators, populations, or conceptual frameworks.

A drafting model followed by an independent critic is better described as:

Independent adversarial review

or, more precisely:

Structured adversarial manuscript review

My working definition is:

A structured adversarial manuscript review is an independent pre-submission quality-control process in which an AI system is instructed to challenge the manuscript by searching systematically for factual errors, internal inconsistencies, statistical defects, citation problems, missing definitions, alternative explanations, and conclusions that exceed the evidence. All findings remain allegations until verified by the authors.

The final sentence is essential.

AI criticism is not automatically correct.

A model can misunderstand the design, request an inappropriate analysis, overlook information already present, invent a methodological rule, or incorrectly claim that a reference does not support a statement.

The output must therefore be treated as a list of potential defects for verification, not as an authoritative verdict.

What follows is not vague advice about using AI better. It is the exact adversarial-review method and the full prompt I use to pressure-test manuscripts before submission, because if AI is going to touch scientific writing, it should be forced to expose weakness, not just polish prose.

If you write, review, or submit research and want the actual workflow rather than the slogan, subscribe and keep reading.

The Complete Structured Adversarial Manuscript Review Prompt

The prompt below is designed for medical and scientific manuscripts. It can be adapted for reviews, original investigations, opinions, guidelines, policy analyses, or ethics papers.

Two Steps:

1. Copy then paste it into an AI (Claude/hatGPT/Gemini etc)

2. Paste the complete manuscript after the prompt. Include the abstract, main text, tables, figure legends, references, and supplementary material whenever possible.

→ COPY/Paste This Prompt ←

STRUCTURED ADVERSARIAL MANUSCRIPT REVIEW PROMPT

You are acting as a senior medical journal peer reviewer, statistical editor, research-methodology reviewer, and scientific-integrity auditor.

Your assignment is adversarial quality control.

Do not help the authors defend the manuscript. Do not assume that the central argument, analyses, references, or conclusions are correct. Attempt to identify every defect that could undermine validity, reproducibility, interpretation, clinical relevance, ethical acceptability, or suitability for publication.

Your role is not to be hostile. Your role is to be skeptical, precise, and evidence disciplined.

Do not begin with praise. Do not provide a summary of strengths unless a claimed strength requires correction or qualification. Do not soften important criticism with compliments.

At the finding stage, prioritize coverage over politeness. Surface every defect you can substantiate, including lower-severity defects. The authors will rank and adjudicate the findings afterward.

Core rules

1. Treat the manuscript as the sole source of truth for internal-consistency checks.

2. Do not invent facts, references, guideline requirements, statistical rules, journal policies, or missing data.

3. When a criticism depends on information outside the manuscript, label it:

   EXTERNAL VERIFICATION REQUIRED

   Explain exactly what must be checked.

4. Distinguish clearly among:

   • a demonstrated error;

   • a probable error;

   • a possible concern;

   • missing information;

   • a matter requiring external verification;

   • a matter of interpretation or editorial judgment.

5. Do not claim that a citation is inaccurate unless you can inspect the cited source. When you cannot inspect it, state:

   Citation-content match not verified.

6. Do not treat absence of evidence as evidence that something did not occur.

7. Do not recommend additional analyses merely because they are possible. Recommend them only when they are necessary to evaluate the stated research question, address a material bias, test a central assumption, or support the conclusions.

8. Do not confuse statistical significance with clinical importance.

9. Do not infer causation from observational associations unless the design and analysis justify causal inference.

10. Do not accept polished language as evidence of scientific validity.

Review the manuscript in the following domains

A. Research question and definitions

Identify:

• unclear or unstable research questions;

• hypotheses introduced after results are known;

• undefined exposures, outcomes, populations, or clinical terms;

• changes in definitions between the abstract, methods, results, tables, and discussion;

• composite outcomes that combine clinically dissimilar events;

• thresholds or categories that appear arbitrary or post hoc;

• terminology that overstates what was actually measured.

B. Study design

Evaluate:

• whether the design can answer the stated question;

• whether the sampling frame matches the target population;

• inclusion and exclusion criteria;

• selection bias;

• immortal-time bias;

• prevalent-user bias;

• indication bias;

• referral bias;

• spectrum bias;

• verification bias;

• misclassification;

• loss to follow-up;

• informative censoring;

• temporal ambiguity;

• inappropriate use of retrospective data to imply prospective inference;

• whether controls or comparators are appropriate;

• whether the study period introduces secular changes that were not addressed.

C. Data integrity and internal consistency

Cross-check all numbers across:

• abstract;

• main text;

• tables;

• figures;

• supplements;

• denominators;

• subgroup totals;

• percentages;

• confidence intervals;

• P values;

• effect estimates;

• participant-flow diagrams.

Recalculate simple percentages and totals when the necessary information is available.

Flag:

• percentages that do not match numerators and denominators;

• totals that do not reconcile;

• impossible values;

• inconsistent sample sizes;

• unexplained missingness;

• outcome counts that differ between sections;

• duplicated categories;

• table labels that contradict the text;

• figure legends that do not match the displayed analysis.

D. Statistical analysis

Assess whether:

• the statistical test matches the variable type and study design;

• paired or clustered data were treated as independent;

• repeated observations were handled appropriately;

• time-to-event data were analyzed correctly;

• competing risks were relevant;

• model assumptions were evaluated;

• covariates were selected appropriately;

• mediators or colliders may have been adjusted for;

• overadjustment may have occurred;

• confounders were omitted;

• continuous variables were inappropriately categorized;

• nonlinear relationships were ignored;

• multiple comparisons were addressed;

• subgroup analyses were prespecified and adequately powered;

• interaction was formally tested rather than inferred from separate P values;

• missing data methods were adequate;

• complete-case analysis could create bias;

• imputation methods were described and appropriate;

• propensity methods achieved balance;

• model discrimination and calibration were distinguished;

• confidence intervals support the language used;

• non-significance was incorrectly interpreted as equivalence or no effect;

• post hoc power calculations were used inappropriately;

• causal language exceeds what the analysis supports.

For every statistical concern, identify the exact analysis and explain why it may be invalid or misleading.

E. Results presentation

Identify:

• selective reporting;

• emphasis on favorable secondary outcomes over the primary outcome;

• switching between absolute and relative effects to amplify findings;

• omission of absolute risks;

• failure to provide uncertainty estimates;

• clinically trivial effects presented as important;

• subgroup findings presented as definitive;

• exploratory analyses presented as confirmatory;

• conclusions based primarily on P values rather than effect size and precision;

• discrepancies between adjusted and unadjusted findings that are not explained.

F. Interpretation and causal precision

For every major conclusion, ask:

1. What exact result supports this statement?

2. Does the result establish association, prediction, diagnostic performance, treatment effect, or causation?

3. Is the wording stronger than the design permits?

4. Has the manuscript excluded plausible alternative explanations?

5. Is the conclusion generalizable beyond the study population?

6. Does the conclusion confuse absence of statistical significance with evidence of absence?

7. Does it confuse model performance with clinical benefit?

8. Does it claim safety without adequate power to detect harm?

9. Does it describe a surrogate outcome as though it were a patient-important outcome?

Quote or identify every sentence that should be weakened, qualified, or deleted.

G. Clinical relevance

Determine whether:

• the study addresses a clinically meaningful question;

• the measured outcome matters to patients;

• the reported effect would change practice;

• benefits and harms are treated symmetrically;

• the comparator reflects current practice;

• diagnostic or predictive performance has been connected to clinical decisions;

• implementation claims are supported;

• feasibility, equity, cost, and access are addressed when relevant;

• the manuscript recommends action before clinical utility has been demonstrated.

H. Ethics and professional responsibility

Evaluate:

• informed consent;

• institutional review;

• privacy;

• vulnerable populations;

• conflicts of interest;

• industry influence;

• fairness;

• representativeness;

• algorithmic bias;

• stigmatizing language;

• unjustified exclusion;

• burdens imposed on patients;

• whether ethical claims are presented as settled when they remain contested;

• whether autonomy is confused with an unlimited entitlement to any requested intervention;

• whether professional obligations are defined and justified.

Do not invent an ethics violation. Identify only what is demonstrated or what information is missing.

I. References

Check for:

• uncited factual claims;

• references that appear irrelevant based on title or context;

• reliance on narrative reviews when primary evidence is required;

• outdated references for rapidly changing claims;

• citation clusters that do not clearly support individual assertions;

• references used to support stronger claims than the manuscript attributes to them;

• missing landmark or contradictory evidence.

Do not claim to have verified a reference unless you actually inspected it.

Create a separate list titled:

References requiring source-level verification

For each entry, state the manuscript claim that must be checked against the source.

J. Abstract and conclusion

Determine whether the abstract accurately represents:

• design;

• population;

• sample size;

• primary outcome;

• effect size;

• uncertainty;

• limitations;

• level of causal inference.

Identify anything in the abstract or conclusion that is not directly supported by the results.

Required output format

For each finding provide:

Finding number

Location: section, paragraph, table, figure, or quoted text

Problem: precise description

Why it matters: effect on validity, interpretation, reproducibility, clinical relevance, or publication suitability

Severity:

• Critical

• Major

• Moderate

• Minor

Verification status:

• Verified internally

• Probable

• Possible

• Missing information

• External verification required

Recommended correction: the minimum change needed to resolve the problem

After listing all findings, provide:

1. Five most consequential defects

2. Claims that must be weakened or deleted

3. Analyses that must be corrected or justified

4. References requiring source-level verification

5. Missing information that prevents full assessment

6. Overall editorial recommendation:

   • Accept

   • Minor revision

   • Major revision

   • Reject

7. One-paragraph confidential note to the editor

Do not recommend acceptance merely because no obvious defect was found. State what you were and were not able to verify.

→ END OF PROMPT ←

How to Use the Output

The model’s response should not be copied blindly into the manuscript.

Each finding should be adjudicated.

I divide the findings into four groups:

1. Correct and actionable

The model has found a real error or weakness. Correct it.

2. Correct concern, wrong proposed solution

The model has identified a genuine problem but suggested an inappropriate analysis or revision. Address the underlying concern using the correct method.

3. Unverifiable

The concern depends on a guideline, source, policy, or cited article that has not been inspected. Verify it independently.

4. Incorrect

The model misunderstood the manuscript or invented a requirement. Reject the criticism and document why.

This adjudication is not optional.

Without human verification, adversarial review can simply replace author overconfidence with model overconfidence.

What This Process Can and Cannot Do

A structured adversarial review can identify internal inconsistencies, unsupported language, missing definitions, numerical discrepancies, and obvious methodological problems.

It may also expose something more subtle: the difference between what the manuscript actually demonstrates and what the authors have gradually come to believe it demonstrates.

But it cannot guarantee validity.

It cannot inspect data it has not been given. It cannot verify every reference without access to the sources. It cannot replace a statistician who understands the dataset, a clinician who understands the disease, or an editor who understands the literature and the journal.

Recent research suggests that carefully designed AI feedback can make reviews more specific and actionable, but it remains uncertain whether such systems consistently strengthen the underlying science. AI-generated reviewers are also vulnerable to manipulation and can produce confident but erroneous assessments.

The correct conclusion is not that AI can perform peer review autonomously.

It is that authors should use every available method to identify errors before asking reviewers, editors, clinicians, and patients to find them later.

The Principle

AI should not be used only to make a manuscript more fluent.

It should be used to make the manuscript harder to fool ourselves about.

The drafting model asks:

How can this argument be expressed more clearly?

The adversarial reviewer asks:

Why should anyone believe it?

Scientific writing needs both questions.

Before submitting a manuscript, do not ask AI only to make it sound stronger.

Ask another AI to try to break it.

Then verify every criticism and repair everything that does not survive.

Suggested AI-use disclosure

Generative artificial intelligence was used to assist with language refinement and to conduct a structured adversarial pre-submission review. The adversarial review was designed to identify internal inconsistencies, unsupported claims, statistical concerns, citation issues, and conclusions exceeding the evidence. All AI-generated suggestions and criticisms were independently evaluated by the authors, who retained full responsibility for the manuscript’s content, accuracy, analysis, references, and conclusions.

Selected sources

1. Bougie N, Watanabe N. Generative adversarial reviews: when LLMs become the critic. arXiv. 2024;arXiv:2412.10415.

2. Zhu M, et al. DeepReview: improving LLM-based paper review with human-like deep thinking process. Proceedings of ACL. 2025.

3. Lin TL, et al. Assessing the vulnerability of large language models in automated peer review. Findings of EMNLP. 2025.

4. Thakkar N, et al. A large-scale randomized study of large language model-assisted peer-review feedback. Nature Machine Intelligence. 2026.

It is easy to warn people not to tell an artificial intelligence system too much about themselves.

• Do not enter your medical history.

• Do not describe your symptoms.

• Do not disclose your medications, pregnancy, mental health, or sexual history.

• Do not trust a chatbot with information that should remain between you and your doctor.

That advice sounds responsible.

It also assumes that you have a doctor.

For much of the world, that assumption is false.

In 2023, approximately 4.6 billion people were not fully covered by essential health services. In 2022, 2.1 billion experienced financial hardship because of health-care costs. The World Health Organization now projects a global shortage of approximately 11 million health workers by 2030, concentrated mainly in low- and lower-middle-income countries. These figures do not mean that every one of those people has literally never encountered a physician. They mean that dependable, timely, affordable access to needed care remains unavailable to more than half of humanity.

For these people, the relevant comparison is often not:

AI versus a trusted personal physician.

It is:

AI versus no appointment.
AI versus no transportation.
AI versus a clinic that is closed.
AI versus a consultation the patient cannot afford.
AI versus nobody who speaks the patient’s language.
AI versus searching alone at two o’clock in the morning.

Before we accuse people of “oversharing” with AI, we should ask a more uncomfortable question:

Who else were they supposed to ask?

What People Are Actually Asking AI

A 2026 study analyzed 617,827 health-related conversations with Microsoft Copilot from users around the world.

Approximately two in five conversations involved general health information and education. Nearly one in five involved personal symptoms or discussion of a specific medical condition. Additional conversations concerned emotional well-being, health insurance, medical paperwork, and navigating health-care systems.

Personal health questions increased during the evening and nighttime, precisely when conventional services were least available. Questions concerning emotional well-being increased especially sharply at night. Many people were also asking on behalf of children, partners, parents, or other people for whom they were providing care.

The accompanying Nature commentary reached the obvious conclusion: people are using conversational AI to fill gaps in access to health information and health-care systems.

This is not merely a story about technological enthusiasm.

It is a map of unmet need.

Subscribe now

The Privacy Debate Is Being Conducted From a Position of Privilege

A person with insurance, a primary physician, specialist access, an after-hours telephone service, reliable transportation, and a nearby emergency department can reasonably decide that the privacy risks of a chatbot are unnecessary.

That person has alternatives.

A woman in a remote community who has developed bleeding during pregnancy may not.

A patient who cannot afford another consultation may not.

A caregiver trying to understand an elderly parent’s new medication at midnight may not.

A person with a stigmatized condition who fears judgment or legal consequences may not.

A patient who waited six months for an appointment and was given seven minutes with a clinician may technically have “access,” but may still have no meaningful opportunity to ask questions.

Privacy matters. Confidentiality matters. Data security matters.

But privacy cannot be discussed as though it were the only ethical value.

People also have interests in obtaining information, understanding symptoms, recognizing danger, navigating a fragmented system, and making decisions before professional help becomes available. When these interests conflict, the answer cannot simply be to tell people with no realistic alternative to remain uninformed.

That is not protection.

It is abandonment expressed in the language of safety.

Read more

“Every new medical technology promises physicians more time with patients. Too often, it gives health systems more patients per physician.”

For two centuries, medicine has repeatedly declared that physicians were becoming obsolete.

The stethoscope would replace listening.

The laboratory would replace clinical judgment.

Radiology would replace physical examination.

Evidence-based medicine would replace experience.

Electronic health records would replace memory.

Now, according to many commentators, artificial intelligence will replace physicians.

I disagree.

What AI is replacing is something different.

It is replacing the scarcity of medical knowledge.

For most of history, physicians possessed something their patients could not access.

Knowledge.

Experience.

Interpretation.

If you wanted to understand preeclampsia, placenta accreta, or recurrent pregnancy loss, you had one realistic option.

Ask your doctor.

That monopoly has disappeared.

Today every patient carries an extraordinarily capable medical information system in a pocket.

Patients arrive already informed, sometimes correctly, sometimes incorrectly, but almost never uninformed.

That changes medicine forever.

The physician is no longer the gatekeeper of information.

The physician becomes the interpreter of information.

That is a profound difference.

A recent JAMA Perspective argues that physicians are losing their “aura,” the unique authority that historically surrounded medical expertise. The author compares today’s AI revolution with photography’s disruption of painting. When reproductions became possible, painters did not disappear. They reinvented themselves. Physicians, he argues, must now do the same.

I think that analysis is only half right.

Medicine has already been losing that aura for decades.

Evidence-based guidelines replaced individual authority with standardized recommendations.

Electronic health records transformed physicians into documentation specialists.

Quality metrics replaced professional discretion with measurable performance.

Artificial intelligence did not begin this process.

It simply exposed it.

The mistake many physicians are making is believing that the competition is AI.

It is not.

The competition is the physician who knows how to use AI well.

Throughout medical history, new technologies have never eliminated competent physicians.

They have eliminated physicians who refused to adopt better tools.

Ultrasound did not replace obstetricians.

Obstetricians who mastered ultrasound replaced those who did not.

The same happened with laparoscopy.

With fetal echocardiography.

With genomics.

Large language models are following exactly the same pattern.

For obstetricians, the implications are immediate.

Our specialty depends on enormous amounts of constantly changing information.

Guidelines.

Drug safety.

Genetics.

Ultrasound.

Ethics.

Population studies.

No physician can memorize all of it.

Nor should we try.

The new competency is no longer remembering every recommendation.

It is knowing how to retrieve, verify, interpret, and apply evidence better than anyone else.

That is not a technological competency.

It is a professional one.

This requires a different educational model.

Residency programs should no longer ask only whether trainees know the answer.

They should ask whether trainees know how to obtain the best answer.

Can they supervise an LLM?

Can they detect hallucinated references?

Can they recognize outdated guidelines?

Can they critically compare competing recommendations?

Can they explain the evidence to a frightened patient?

Those are becoming essential clinical skills.

Ironically, AI may make physicians more human.

If documentation, literature searches, and administrative writing increasingly become automated, physicians regain something far more valuable than efficiency.

Attention.

Listening.

Judgment.

Empathy.

Those were never threatened by AI.

They were threatened by paperwork.

The future obstetrician will not be remembered because they memorized the most guidelines.

They will be remembered because they asked the best questions, interpreted the evidence most wisely, recognized when the AI was wrong, and remained fully accountable for every decision.

That is why I do not believe physicians are losing their role.

Our role is becoming clearer.

Medical knowledge is no longer scarce.

Clinical judgment still is.

And that may become the most valuable skill of all.

Most doctors and clinicians use AI the same way.

Open a chat. Type a question.

Read the answer.

Close the tab.

It feels like using the tool. It is closer to wasting it. Every one of those sessions starts from zero. The AI does not know who you are, what you publish, what standards you hold, or how you write. So it gives you the average answer, written for the average person. In medicine, the average answer to your specific question is often the wrong one.

Ruben Hassid, who writes one of the largest newsletters on practical AI use, estimates that about nine out of ten users work exactly this way: one question, one answer, done. From what I see among colleagues, doctors are no exception. We may be worse.

We are busy, we are skeptical, and many of us tried an AI once in 2023, caught it inventing a reference, and filed the whole technology under “not ready.”

The technology was not the problem.

The setup was.

Nobody walks into an operating room and starts cutting.

We check the instruments, the bookings, the consent, the team. Then we start. AI rewards the same habit, and almost no one in medicine has been told this.

The gap between doctors who set up their AI and doctors who do not is now, in my view, wider than the gap between any two AI models.

What follows is the exact setup I use, written so you can copy it this weekend. Paid subscribers read on.

Read more

Better Prompts Lead to Better Medical Answers

Large language models can be powerful pregnancy education tools, but the quality of the answer depends heavily on the quality of the prompt. A vague prompt such as “Should I have an induction?” usually produces a generic response that may not apply to your pregnancy. A specific prompt that includes your situation, asks for evidence, requests absolute rather than relative risk, and separates established evidence from expert opinion is much more likely to produce a useful answer.

Subscribe now

An effective prompt tells the AI exactly what role to take, what information to analyze, what evidence to use, and how to present the answer. It should ask for current clinical guidelines, the strongest available studies, and clear explanations in plain language. It should also ask the AI to identify uncertainty, explain where evidence is limited, and distinguish facts from interpretation. The goal is not to replace your obstetrician or midwife, but to help you ask better questions and participate more effectively in shared decision-making.

For paid subscribers, here are all prompts and how to best use them.

Read more

91% of the top medical journals with an AI policy forbid reviewers from uploading your manuscript to a chatbot. That rule protects your confidentiality. It says nothing about you running your own draft through AI before you ever hit submit — which is exactly what you should do.

A reviewer opens your manuscript at eleven at night.

She is an expert.

She is…

Read more

About one in three American adults now ask an AI chatbot a health question in a given year.

That is roughly the same share that turns to social media for health information. So a quiet research question has become an urgent one: when should a medical AI keep its mouth shut?

A new review in npj Digital Medicine pulls together the young science of “abstention” — a technical word for a simple idea.

Subscribe now

It means knowing when to stop, ask for more information, or just say “I don’t know.” The finding is blunt. Today’s models almost never do it. In one study, a leading model held back on only 1 in 10 medical questions that had no safe answer. The rest of the time it answered anyway, fluent, confident, and often wrong.

So researchers are building tools to teach AI when to stay quiet. And the behavior they score as “safe,” again and again, is the same: when in doubt, the model should stop and say, “Please consult a doctor.”

That sounds right.

It is also where the thinking stops.

The whole approach rests on one assumption nobody checks: that the doctor you are sent to will do better.

Subscribe now

Start with the machine. When an AI gives a confident answer it has no real basis for, we now call it confabulation — making up a plausible story to fill a gap in what it actually knows. The field takes this seriously. There are now careful tools that catch a model in the act of inventing, by sampling its answers and flagging when its confidence is hollow. Good work, and necessary.

But confabulation is not a computer glitch. It is what any mind does, silicon or human, when it is unsure and feels pressure to produce an answer.

Daniel Kahneman spent a career proving this about people. He showed that confidence is not a readout of accuracy. It is a feeling, produced by how good a story the mind can tell itself. He called it the illusion of validity: we feel sure when the pieces fit, even when the evidence is thin. And he named the engine, the mind builds its answer from whatever is in front of it and quietly ignores what is missing.

Read that again. It describes a language model. It also describes a tired clinician at 3 a.m.

The evidence on doctors is not gentle. In a landmark 2008 review, two patient-safety researchers argued that physicians badly underrate how often their diagnoses are wrong. When one of them asked rooms full of doctors whether they had made a diagnostic error in the past year, only about 1 in 100 said yes. A 2013 study put numbers on the gap. Physicians got about 55 out of 100 easy cases right — and fewer than 6 out of 100 hard ones. Their confidence barely moved: about 7.2 out of 10 on the easy cases, 6.4 on the hard ones. Accuracy fell off a cliff. Confidence took a gentle stroll downhill. On the hard cases they were wrong more than nine times in ten and stayed nearly as sure of themselves.

What they almost never said was “I don’t know.” That phrase is in remarkably few doctors’ repertoires. The culture of medicine treats uncertainty as something close to failure, and tolerance for not-knowing measurably drops over the course of medical school. We train it out of people.

So here is the asymmetry the AI field keeps walking past. We have built detectors to catch the machine bluffing. We have built nothing of the kind for the clinician who does the same thing, undetected, and then writes it in the chart. The machine that confabulates gets a benchmark and a leaderboard. The human it defers to gets the benefit of the doubt.

In obstetrics this matters twice over, because there are two patients. And there is a second blind spot in the research: it treats silence as the safe choice. At the bedside it often is not. The harm in my field comes far more from things not done, the warning sign not named, the worrying tracing not acted on, the escalation that came too late, than from saying too much. An AI trained to go quiet whenever it is unsure can be trained straight into the passivity that hurts patients. And “consult a doctor” assumes there is a doctor to consult, and that the doctor knows. For many people typing questions at midnight, neither is true.

This is the part the silence debate misses, and it is the heart of a clinical opinion my colleagues and I just published. The danger was never that clinicians would use AI. The danger is that they will use it poorly, copying confident output without checking it, accepting an error because it reads well, treating a chatbot as an authority instead of a tool. The answer is not to avoid these tools, and it is not to make them quieter. It is competence: knowing when AI helps and when it does not, how to test its answers, how to catch a made-up drug dose before it reaches a patient, how to say plainly what is known and what is not, and who stays accountable when the screen goes dark. The clinician does.

Obstetrics has done this before. Ultrasound, electronic fetal monitoring, cell-free DNA screening, each arrived to resistance, and none of them replaced the obstetrician’s judgment. Each demanded a higher version of it. AI is next in that line.

But notice what that competence really asks for. Knowing the limits of your own knowledge. Saying “I don’t know” out loud. Matching your confidence to what you actually checked. That is exactly the calibration we are now trying to engineer into the machine — and it is the same skill medicine has never seriously required of itself.

For a patient, the practical version is small and powerful. The AI article you read before your appointment can sound certain and be wrong. So can the confident answer across the desk. The clinician worth trusting is not the one who never hesitates.

Share

It is the one who can say, “I’m not sure: let me check,” and then actually checks.

You are allowed to ask two questions out loud: How sure are you? And did you verify it? A good doctor will not flinch at either.

My take: we are studying the wrong half of the problem. The question was never silence versus speech. It is calibration — knowing what you know, and admitting what you don’t. Kahneman built his entire method on the premise that he could not trust his own confidence, so he backed everything with data. That is the discipline. It is rare in the model. It is just as rare in the physician we hold up as the safe alternative.

The arrival of AI is forcing a calibration test on medicine that medicine never volunteered for. We should take it. The goal is not a quieter machine. It is a profession — human and artificial together — that is finally honest about the edge of its own knowledge. In obstetrics, where two patients depend on that honesty, “first do no change” has to give way to “first do good.” Competence, not silence, is the safer answer.

Bottom line: don’t ask how to make AI stay quiet. Ask how to make both the machine and the clinician tell you the truth about how much they actually know.

If this is the kind of thinking you want more of, subscribe to ObGyn Intelligence. It is free, it is independent, and it says plainly what the data shows.

References

1. Presacan O, Nik A, Ojha J, Thambawita V, Ionescu B, Riegler MA. When silence is safer: a review and decision-theoretic framework for LLM abstention in healthcare. npj Digit Med. 2026. doi:10.1038/s41746-026-02882-1.

2. Grünebaum A, Dudenhausen J, Chervenak FA. Clinical artificial intelligence competence in obstetrics and gynecology: patient safety, physician accountability, and responsible use. Am J Obstet Gynecol. 2026. doi:10.1016/j.ajog.2026.06.011.

3. Berner ES, Graber ML. Overconfidence as a cause of diagnostic error in medicine. Am J Med. 2008;121(5 Suppl):S2-S23. doi:10.1016/j.amjmed.2008.01.001.

4. Meyer AND, Payne VL, Meeks DW, Rao R, Singh H. Physicians’ diagnostic accuracy, confidence, and resource requests: a vignette study. JAMA Intern Med. 2013;173(21):1952-1958. doi:10.1001/jamainternmed.2013.10081.

5. Kahneman D. Thinking, Fast and Slow. New York, NY: Farrar, Straus and Giroux; 2011.

6. Grünebaum A, Chervenak J, Pollet SL, Katz A, Chervenak FA. The exciting potential for ChatGPT in obstetrics and gynecology. Am J Obstet Gynecol. 2023;228:696-705. doi:10.1016/j.ajog.2023.03.009.

You no longer need an engineer to turn what you know into something patients can use.

“Tools will come and tools will go. Only the vibe coder remains.” A legendary music producer said it about software. It is also the best argument I know for why doctors should learn to build their own clinical tools. Here is what it means for medicine.

Andrej Karpathy, a founder of OpenAI, gave it a name in February 2025. Rick Rubin, who has produced more hit records than almost anyone alive, gave it a philosophy: “Tools will come and tools will go. Only the vibe coder remains.” Collins Dictionary named it the Word of the Year for 2025. Vibe coding has arrived.

Subscribe now

Most doctors are watching from the sidelines. They should not be.

What vibe coding actually is

Vibe coding means building software by describing what you want in plain language and letting an AI write the code. You do not memorize syntax. You do not hire an engineer. You say what the tool should do, the machine builds it, you test it, and you correct it. Karpathy’s original line was to forget that the code even exists. You see something, say something, run it, fix it.

Rubin’s point cuts deeper. He produces records without playing the instruments. His value is taste and judgment, knowing what good sounds like and steering the artist toward it. The tool in front of him keeps changing. His ear does not. That is the part that lasts.

For a doctor, the ear is clinical judgment

Fifty years at the bedside, thousands of patients, the feel for what a number means and when it matters. That is the scarce thing. Until now, turning that judgment into a working tool required someone else. You explained the idea to a programmer who did not know obstetrics, waited months, paid a great deal, and often got back something that missed the clinical point. The translation step was the bottleneck. Vibe coding removes it.

My own numbers

I have built more than 100 clinical tools this way. Risk calculators, screeners, patient education pages, decision aids, all on tools.obmd.com and liveevidence.com, all created by directing AI rather than writing code by hand. To my knowledge I am the only ObGyn in the United States with a tool library of this size built this way. I am in my seventies. I do not say this to impress you. I say it because if I can do it, the resident who grew up with a phone in her hand certainly can.

The tools are the easy part. What doctors get wrong about vibe coding is the part that actually matters.

The judgment is the product

A software engineer can build you a calculator. He cannot tell you whether the threshold is right, whether the population it was validated in looks like your patient, or whether the output will frighten a woman at two in the morning over a number that does not warrant it. You can. The clinical judgment is not a feature you add at the end. It is the whole point.

This is why doctors, not engineers, should hold the controls. The tool is only as good as the question behind it, and the question is clinical. When I build a tool, I decide what counts as a meaningful number. I write the plain-language explanation. I choose what the tool refuses to say because the evidence does not support saying it. Those are not coding decisions. They are medical ones.

Where vibe coding gets dangerous

Now the honest part. Karpathy’s advice to accept everything and forget the code exists is fine for a personal weekend project. It is reckless for a tool a patient will trust. A tool that computes a risk score is making a medical claim. It deserves the same scrutiny as a drug or a device.

So the rules are simple. Every number in the tool traces to a published source. I check the math against the original model, by hand, before anything goes live. I do not ship a black box: if a tool gives a result, it shows its reasoning and cites its evidence. And no patient data is ever stored or transmitted. My tools run entirely in the browser and keep nothing. Privacy is not a setting you switch on. It is the design.

Here is the paradox worth sitting with. The thing that makes a doctor safe at vibe coding is the same thing that makes vibe coding dangerous in untrained hands: judgment about what the numbers mean. A layperson who vibe codes a pregnancy risk tool does not know the threshold is wrong. You do. That is not a small advantage. It is the entire case for why this skill belongs to clinicians.

What this means for patients

When your doctor uses a tool to estimate your risk, you have a right to ask three things. Where do the numbers come from. Has the tool been checked against the published evidence. And what happens to the information you typed in. A good tool answers all three plainly. A tool built by a clinician who understands the evidence can answer them. A black box bought from a company often cannot.

The skill outlasts the tool

The tool I use today will be gone in two years. Karpathy himself, one year after coining the term, already calls vibe coding passe and renames the professional version agentic engineering. The label keeps changing. The brand of AI keeps changing. None of that matters. What Rubin understood, and what doctors should understand, is that the durable skill is not the tool. It is knowing what good looks like and being able to direct the machine toward it. For a doctor, that is simply clinical judgment pointed at a new instrument.

Conclusion

I am not telling you to become a programmer. I am telling you the opposite.

The era of the doctor as the person who memorizes ten thousand facts is ending, because the machine remembers better than any of us.

What it cannot do is care for the specific patient in front of you, or decide what a number should mean for her.

That is yours.

Vibe coding lets you build the tools that judgment deserves, without waiting for permission, a budget, or a programmer who has never set foot in a delivery room. Rubin reimagined the Tao Te Ching for the age of code. I will borrow from a source he would appreciate, Marcus Aurelius: the obstacle is the way. For doctors, the tool was never the obstacle. The translation was. It is gone. Build the thing only you can build.

ObGyn Intelligence is independent because readers keep it that way. If this changed how you think about the next decade of your work, subscribe. The evidence, and now the tools, should travel.

Subscribe now

References

1. Rubin R, Andreessen M, Horowitz B, Midha A, Torenberg E. Rick Rubin: Vibe Coding is the Punk Rock of Software. a16z Podcast. 2025. Available from: https://a16z.com/podcast/rick-rubin-vibe-coding-is-the-punk-rock-of-software/

2. Karpathy A. Post on X introducing the term “vibe coding.” February 2, 2025.

3. Collins Dictionary. Word of the Year 2025: vibe coding. 2025.

4. Taft DK. Vibe coding is passe. Karpathy has a new name for the future of software. The New Stack. February 10, 2026. Available from: https://thenewstack.io/vibe-coding-is-passe/

ObGyn Intelligence - Evidence Matters | obmd.com

A patient comes back for her second visit. Her symptoms have shifted. Her lab results are in. A good doctor remembers what happened last time, reads the new numbers, and adjusts the plan. That thinking over time is the hard part of medicine. The one-shot diagnosis is the easy part.

Subscribe now

A new study in Nature asked a simple question: can an artificial intelligence do the hard part?

AMIE stands for Articulate Medical Intelligence Explorer.

It is a medical AI system built by Google, running on their Gemini model. Earlier work showed it could take a patient history and suggest a diagnosis. This new paper pushes further, into what doctors call management — deciding what tests to order, what to prescribe, what to watch, and how to change course as a patient returns over weeks.

The design is clever, and worth understanding in plain terms. AMIE is built from two parts that work like the two speeds of human thinking.

• One part is a fast talker. It chats with the patient, sounds warm, and keeps track of the conversation.

• The other part is a slow thinker.

While the conversation continues, it quietly reads through hundreds of pages of clinical guidelines and writes a careful plan, with every recommendation tied to the exact guideline it came from. The idea comes straight from Daniel Kahneman’s Thinking, Fast and Slow: quick intuition on the surface, slow reasoning underneath.

To test it, the researchers ran a blinded, randomized exam. They wrote 100 made-up cases, each unfolding over three visits a day or two apart. The cases spanned five specialties, and obstetrics and gynecology made up 20 of the 100. Trained actors played the patients. For each case, the same patient was seen once by AMIE and once by one of 21 primary care doctors, all by text chat. Specialists then graded both, without knowing which one was the machine.

AMIE did at least as well as the doctors on overall management, and clearly better on two things. First, it was more precise. Instead of “prescribe an antibiotic,” it named the drug, the dose, the route, and the duration. On treatment precision it scored around 95 out of 100; the doctors scored in the 60s. Second, it stuck closer to the guidelines and showed its work, citing the source for each recommendation. When graders did prefer one over the other, they picked AMIE far more often than the doctor — though in about half the cases they saw no real difference at all.

Here is the part the headlines will skip. The doctors were set up to lose. They came from Canada and India but were judged against United Kingdom guidelines they do not use every day. They worked through a text box, not the phone or video they normally use. The cases had clean, correct answers, which real patients rarely provide. And the visits were one or two days apart, not the weeks the cases described — so human memory was tested in a way that does not match real care. AMIE, meanwhile, was built and tuned for exactly this setup. This is what I call the missing comparator. When you compare a system to people working with one hand tied, “the AI won” is not the finding. The finding is narrower, and still useful: the AI is very good at reading guidelines and writing precise plans.

Subscribe now

Why should an ObGyn care? One in five of these cases was ours. The things AMIE did best — remembering the last visit, citing the right guideline, writing a precise plan — are exactly the things that get sloppy on a busy, tired labor floor. A tool that never forgets and always cites the source has real value there. But the things AMIE cannot do are also ours: the woman in front of you whose values, fears, and body do not match the guideline. The study’s own authors say plainly that AMIE is not ready for real patients. They are right.

My take. This is careful, honest work, and I do not say that lightly about an AI paper. The two-speed design is smart, grounding the plan in guidelines is the right instinct, and the authors were more open about their limits than most. But “AI matches doctors” is not what this shows. It shows that AI beats a distracted human at the one thing computers have always been better at: holding everything in memory and quoting the manual exactly. That is useful. It is not judgment. For ObGyn the lesson is neither fear nor worship. Let the machine carry the memory and the guideline lookup, and keep the thinking about the specific woman where it belongs — with you. The real risk is not that AMIE replaces us. It is that we stop practicing the parts it does well and forget how to do them ourselves.

Share

One footnote, for the attentive reader. The study was funded by Google, and the authors may own Google stock. I mention this not as an accusation but as arithmetic.

AMIE is real, it is impressive, and it is a research demonstration, not a doctor. Read the result, not the headline.

References

1. Liévin V, Palepu A, Weng W-H, Saab K, Stutz D, Cheng Y, et al. Towards Conversational AI for Disease Management. Nature. 2026. doi:10.1038/s41586-026-10764-5.

2. Tu T, et al. Towards conversational diagnostic artificial intelligence. Nature. 2025;642:442–450.

Last week I went looking for one old document in ChatGPT. I found something I did not expect. There was a folder holding every file I had ever uploaded. Old drafts. PDFs. Documents I was sure I had deleted months ago. I had deleted the conversations. The files were still sitting there.

The folder is called the Library.

Most people do not know it exists.

It’s in your left sidebar.

In front of your eyes.

That is the whole problem in one sentence.

Here is what the Library is.

When you upload a file to ChatGPT, or when ChatGPT makes a file for you, it saves a copy. Documents, spreadsheets, pictures, PDFs, all of it. You do not turn this on. You do not click save. It happens on its own, quietly, in the background. The files pile up whether you notice or not.

Now here is the part that matters most. Deleting the chat does not delete the file. When you delete a conversation, it disappears from the list on the side of your screen. It looks gone. But the file you uploaded stays in the Library. So the thing almost all of us believe, that deleting a chat means it is gone, is simply not true. The chat is gone. The file is not.

Think about what people put into ChatGPT. A resume. A lease. A letter from a lawyer. A photo of a bill. A medical record. For my colleagues in medicine, think about every case we called de-identified that was not quite as anonymous as we told ourselves. Any of it can be sitting in storage you never knew you had, on a company’s computer, not yours.

• Opt-in means a person’s data is not used unless they actively give permission.

• Opt-out means the data may be used by default unless the person notices the setting and turns it off.

Opt-in is safer because it respects informed consent, reduces accidental data sharing, and places the burden on the organization to obtain permission before using personal information. Opt-out systems are weaker because many people never see the setting, do not understand it, or assume privacy is already protected.

Subscribe now

There is a second layer, and it is also turned on by default. ChatGPT has a setting called Improve the model for everyone. When it is on, the company can use what you type, and the files you upload, to train its future models. On by default. Most people never see the switch.

And when you finally do delete a file the right way, the company removes it from its servers within thirty days. So delete does not mean gone today. It means gone in about a month. That is better than forever. It is also not nothing.

So what do you do about it? Three things, and none of them take long.

1. First, turn off training. Click your picture in the top right corner of the screen. Open Settings, then Data Controls. Find the line that says “Improve the model for everyone” and switch it OFF. From that moment on, your new chats are not fed into the next model.

   

2. Second, clean out the Library in one go. Open the Library from the menu. At the top of the list of files you will see the word Name. Click it. A box appears that lets you select everything at once. Check it, pick what you do not want kept, and delete. You do not have to click each file one at a time, which matters when there are hundreds of them.

   Subscribe now

   

   

3. Third, use Temporary Chat for anything private. The button sits in the upper right of the chat window. Looks like a broken circle. Not easy to find. A Temporary Chat is never saved to your Library and is never used for training. It is the right tool for the one-off question, the symptom you are embarrassed about, the document you do not want kept, the thing you would not say with the door open.
   Here it is, upper right of a new chat:

What does this mean for you?

Not that ChatGPT is dangerous (or maybe it is?).

Share

Saving your files so you can find them later is a fair and useful feature. For some.

The problem is not the Library. The problem is the silence around it. The saving is on by default. The deleting does not work the way the screen makes you think it does. And almost no one is ever told. That is not a technology problem. That is a consent problem.

Here is my take. I have spent fifty years telling patients the same thing. Consent only counts when it is informed. A choice you were never told about is not a choice you made. It is something that was decided for you while you were looking somewhere else. The rule that governs a procedure should govern your data too. You have a right to know what is being kept, what is being used, and how to stop it.

Share

So open your Library today. Look at what is in there. You may be surprised, the way I was. Then decide, with your eyes open, what stays and what goes.

It took me one click to find the folder. It took one more to empty it. The second click is the one worth telling people about.

If this saved you from leaving something private on a stranger’s server, it did its job. ObGyn Intelligence is where I write about exactly this, the gap between what we are told and what is true, in medicine and in the tools we now use every day. Subscribe, and bring the friend who still thinks deleting the chat deletes the file.

Sources

OpenAI. File storage and Library in ChatGPT. OpenAI Help Center. Available from: https://help.openai.com/en/articles/20001052-file-storage-and-library-in-chatgpt

OpenAI. What if I want to keep my history on but disable model training? OpenAI Help Center. Available from: https://help.openai.com/en/articles/8983130

Two pathologists look at the same breast biopsy. One sees cancer or another anomaly. One does not. This is not a thought experiment. In a 2015 study in JAMA, 115 pathologists read the same slides. Across the full spectrum of breast findings — benign, atypia, DCIS, invasive — overall agreement with the expert panel was 75 percent, with the disagreement concentrated in the preinvasive gray zone.” The cases that mattered most — the early, borderline changes — were the ones they disagreed on most. Same slide. Same patient. A different answer depending on who held the microscope.

I keep that number in mind whenever someone tells me medical AI cannot be trusted.

A new editorial in the International Journal of Medical Informatics makes a strong claim right in its title: almost all machine learning models for medicine are wrong. The authors are serious people, and part of their argument is right.

The labels used to train these tools come from human experts who disagree.

The models are tested on the same kind of patients they learned from, not on new hospitals or new populations. They are scored with numbers that look exact but shift the moment the setting changes. And once switched on, they are rarely watched, even though medicine keeps changing underneath them.

Subscribe now

The title is not new.

It echoes a famous 2005 essay, “Why most published research findings are false.” That one was about human research — the same studies and expert opinions these models are trained to copy. Follow its logic and almost all of medicine rests on shaky findings too. Somehow the cannon gets pointed only at the machine.

Share

All true, as far as it goes. I have made versions of these points myself. Build a tool on a weak foundation, never test it in the real world, and you should not trust it. But the paper never asks the one question that matters most. Wrong compared to what?

Every flaw on their list is also a flaw in human doctors — usually a bigger one. Doctors disagree with each other, as those pathologists did.

They are trained once, tested once, and then practice for decades. Nobody runs a doctor through a second hospital to confirm she still performs. Nobody measures, year after year, whether her judgment still matches the evidence. And the knowledge she learned in training is often more than a decade behind the best current science — on average it takes about 17 years for research to reach everyday practice. A machine can be corrected overnight. A habit takes a generation.

So when a study calls a model “wrong,” I want to know: wrong next to what human standard?

The human standard is not perfect truth.

It is a group of people who disagree, carry old knowledge, and are almost never checked.

This is where the paper trips over its own feet. Its strongest section explains that there is no clean truth in medicine. Diagnoses are fuzzy. Lab tests vary. Experts split. The authors say so plainly. Then they turn around and call the machine “wrong” — wrong against the same fuzzy standard they just told us not to trust.

You cannot have it both ways. If the ruler is made of rubber, the word “wrong” loses its meaning. When a model disagrees with a label, it often disagrees with a label a second expert would have rejected too. That is not error. That is the ordinary disagreement of medicine, renamed as a machine defect.

So who is the arbiter of truth? In medicine, there isn’t one. There is a committee of fallible humans, and we have agreed to treat their majority vote as the truth. That is fine — as long as we judge the machine by the same yardstick we use on ourselves, not a stricter one we invented the day the machine walked in.

Subscribe now

I will give the authors one real point: scale. One tired doctor harms patients one at a time. A bad model, switched on across a whole hospital system, can harm thousands before lunch. That is a genuine reason to watch deployed tools closely — to test them, keep them honest, and check them over time. Their list of fixes is a good list. I would sign most of it. But that is an argument about how we deploy these tools, not about whether they are “wrong.” The methods in this paper are right. The frame is a double standard wearing a white coat. We are demanding proof from the machine that we never demanded from ourselves.

My position is simple.

Hold medical AI to a high standard.

Then hold human medicine to the exact same one.

Stop asking whether the model is perfect.

Start asking whether it is better than what we already do — and then actually measure what we already do, which we almost never have. A tool does not need to be flawless. It needs to beat the disagreeing, aging, unchecked standard we already live with. Often, it already does.

The next time you read that almost all medical AI is wrong, ask the question the headline skips: compared to whom?

Share this with a colleague who still calls the human exam the gold standard — the data on how often we agree should travel.

ObGyn Intelligence is free because the work matters. If you want to keep it independent and in your inbox, subscribe.

References

1. Cabitza F, Jurman G, Molinari F, Bellazzi R. Why almost all ML models for medicine are wrong — and what we need for evidence-based medical AI. Int J Med Inform. 2026 (in press). doi:10.1016/j.ijmedinf.2026.106538.

2. Elmore JG, Longton GM, Carney PA, et al. Diagnostic concordance among pathologists interpreting breast biopsy specimens. JAMA. 2015;313(11):1122-1132. doi:10.1001/jama.2015.1405. PMID 25781441.

3. Morris ZS, Wooding S, Grant J. The answer is 17 years, what is the question: understanding time lags in translational research. J R Soc Med. 2011;104(12):510-520. doi:10.1258/jrsm.2011.110180. PMID 22179294.

4. Ioannidis JPA. Why most published research findings are false. PLoS Med. 2005;2(8):e124. doi:10.1371/journal.pmed.0020124. PMID 16060722.

Ask any AI chatbot a medical question and you get an answer in seconds. It sounds calm, organized, and sure of itself. That last part is the problem. Confidence is not the same as reliability, and a chatbot has no way to tell you which of its sentences it would stake its life on and which it is simply guessing.

So let me show you something simple. I asked a medical AI one ordinary clinical question — the kind a pregnant woman or a first-year resident might ask. Then I asked it the same question a second time, word for word, with a single instruction added to the end. Same model. Same minute. Two very different answers.

Here is the first one.

Prompt: “Tell me what blood tests a pregnant woman should do at her first prenatal visit.”

Here is the answer:

“At her first prenatal visit, a pregnant woman should have her blood type and Rh factor checked, along with an antibody screen and a complete blood count to look for anemia. She should be tested for rubella immunity, hepatitis B, hepatitis C, HIV, and syphilis, and screened for chlamydia and gonorrhea. Her thyroid should be checked with a TSH test, and she should be offered genetic carrier screening and screening for Down syndrome. A urine test is also part of the standard first visit.”

Not wrong, exactly. A reasonable list. But notice what it does not do. It never separates the tests every pregnant woman gets from the ones that depend on her age, weight, or history. It presents a thyroid test as routine — when the main US guideline actually recommends thyroid testing only for women with specific risk factors. It cites nothing. And it never once tells you what it would need to know about the patient to answer well. It just talks.

Now the same question, with one instruction added to the end.

That instruction is the Clinical Evidence Prompt or the CEPrompt™.

PROMPT: “(Tell me what blood tests a pregnant woman should do at her first prenatal visit [PLUS THE INSTRUCTION, the CEPrompt™.”

Here is the full answer it produced — nothing trimmed.

“This is the US/ACOG framework at a routine first visit, assuming no results already documented this pregnancy.

Universal — standard of care for all (high certainty): blood type, Rh(D) and antibody screen; CBC; rubella IgG; hepatitis B (HBsAg minimum, triple panel if not previously documented); hepatitis C (every pregnancy); HIV (opt-out); syphilis (repeat in third trimester and at delivery). Urinalysis and urine culture are also routine but not blood-based.

Selective — depends on the individual patient (moderate certainty): chlamydia/gonorrhea (universal if under 25, otherwise risk-based); varicella IgG if no immunity history; hemoglobinopathy evaluation guided by CBC/MCV and ancestry; early diabetes screening if overweight/obese with risk factors or prior GDM; lead and TB testing if at risk.

Offered, not assigned: genetic carrier screening and aneuploidy screening (cell-free DNA or serum) — both offered to all regardless of age.

Evidence vs. position: The universal panel reflects guideline consensus. Thyroid (TSH) is the exception — ACOG (PB 223) recommends selective, risk-based testing, not universal. Presenting routine TSH as standard is an inference, not consensus; universal offering is a defensible minority position, not the guideline.

Missing information that changes the selective list: maternal age, BMI, ancestry, prior obstetric/GDM history, immunization and prior-testing records, STI risk, gestational age at presentation.

Certainty: high for the universal panel; patient-dependent for selective tests; the thyroid point is genuinely contested.

Citations: Viral Hepatitis in Pregnancy. Clinical Practice Guideline No. 6. ACOG. Obstet Gynecol 2023;142:745–59. SMFM Consult Series #69: Hepatitis B in pregnancy. Am J Obstet Gynecol 2024;230:B2–B11. Owens DK, et al. USPSTF hepatitis C screening. JAMA 2020;323:970–5. doi:10.1001/jama.2020.1123.”

Read that again, because every line of it is doing a job the first answer skipped. It splits the tests into three groups instead of one: the ones every pregnant woman gets, the ones that depend on her age, weight, ancestry, or history, and the ones that are offered for her to accept or decline. It draws a clear line between guideline consensus and the author’s own position — and it catches the thyroid trap, telling you plainly that routine TSH for everyone is not what the main US guideline recommends, even though the first answer presented it as routine. It tells you what it would need to know about the actual patient before the selective list can be finalized. It labels how certain it is, group by group. And it ends with real citations you can open and check, not a vague wave at “the guidelines.”

Subscribe now

That is the difference between a list and a consultation. The first answer hands you tests. The second one reasons with you — it shows what is settled, what depends on the patient, what is contested, what is missing, and where it got each claim. A clinician can act on the second answer. A patient can ask better questions because of it. Nobody should act on the first.

Same machine. The model did not get smarter between the two answers. The second answer is better because the prompt told it to show its work — to separate what the evidence supports from what it is assuming, to say how sure it is, to name what it is missing, and to admit when something is genuinely debated.

That instruction is the whole trick.

That instruction is the Clinical Evidence Prompt or the CEPrompt™.

It is one sentence.

You paste it at the end of any clinical question, and it changes the kind of answer you get back. Not the topic. The honesty.

The exact wording of the CEPrompt™ — the sentence that turns the first answer into the second — is below, with the six things it does and how to attach it to any question you ask. Paid subscribers read on.

Read more

By 2026, about one in three American adults had asked an AI chatbot for health advice. Among physicians, 81% report using artificial intelligence in their work. And fewer than 15% have had any formal training in how to use it safely. Read those three numbers again. We have adopted a clinical tool faster than almost any technology before it, and we skipped the part where someone teaches us how to use it.

A large language model — ChatGPT, Claude, Gemini — is not a database of checked facts.

It is a very good pattern engine that predicts the next likely words.

Most of the time that is helpful. Sometimes it invents a citation that does not exist, or a dose that is wrong, and it does so with total confidence. The danger is not that the tool is bad. The danger is that it is convincing, and that we are using it without knowing where it breaks.

It can make mistakes. It hallucinates. But with proper training you can get the most out of it.

So we built the course.

It is called AI Health Course. Passwords are required to get in. The course is in private preview. Paid subscribers get the access password below after the paywall.

Please message me on LinkedIn for any questions.

It is short — about three to four hours, in sittings of fifteen to thirty minutes — and it is written both for patients and clinicians, not engineers. It covers how to give the model the right context, how to spot an answer that is unsafe, how to verify what matters, how to keep patient information private, and how to stay responsible for the decision. Every idea is taught in plain language, then practiced on a real, de-identified example.

Here is the part I like. I built the course itself by “vibe coding.” That means I described what I wanted in plain English, Claude wrote the working code, and I tested and corrected it until it was right, no web developer in the room. But nothing was taken on the model’s word. Every clinical number, every citation, and every privacy rule was checked against the original source before it went in. The tool drafts; the clinician stays accountable. That is the whole message of the course, and it is also how the course was made.

The course is in private preview.

Paid subscribers get the access passwords below.

Subscribe now

Read more

The Society for Maternal-Fetal Medicine (SMFM) regularly posts an “Ultrasound Case of the Week Challenge.” It’s meant to sharpen our diagnostic instincts and celebrate the art of pattern recognition.

But the exercise now has a rival: artificial intelligence that can see what we see—and explain it better.

Subscribe now

The Case (as posted by the Society of Maternal-Fetal Medicine - SMFM)

Case Summary:
A patient is referred to Maternal-Fetal Medicine at 32 weeks’ gestation because of ultrasound findings of short long bones. On detailed scan, you confirm:

• Shortened limbs with mild femoral bowing

• Normal ossification

• No evidence of fractures

• Amniotic fluid index (AFI): 29 cm (polyhydramnios)

• No other structural anomalies noted

Question: Based on these findings, what is the most likely diagnosis?

Options:
A. Thanatophoric dysplasia
B. Achondroplasia
C. Osteogenesis imperfecta
D. Campomelic dysplasia
E. Diastrophic dysplasia

Correct Answer: B. Achondroplasia

The AI Prompt

Prompt to ChatGPT-5:
“You are a Maternal-Fetal Medicine specialist. Interpret this 32-week fetal ultrasound showing short long bones, mild femoral bowing, normal ossification, no fractures, and an AFI of 29 cm. What is the most likely diagnosis? Explain your reasoning, discuss the differential, outline recommended management, and address ethical issues related to prenatal counseling.”

ChatGPT’s Full Response (abridged for clarity)

Diagnosis:
The findings are most consistent with achondroplasia, the most common non-lethal skeletal dysplasia.

Reasoning:

1. Limb pattern: Shortening of long bones (rhizomelic pattern—proximal > distal) with mild bowing is typical.

2. Ossification: Normal echogenicity distinguishes achondroplasia from lethal dysplasias like thanatophoric dysplasia, which show poor ossification or platyspondyly.

3. Fractures: Absence of fractures argues against osteogenesis imperfecta.

4. Thorax: Normal size (not narrow) suggests adequate pulmonary development, differentiating from lethal dysplasias.

5. Polyhydramnios: Common secondary feature due to limited fetal swallowing or altered skeletal mechanics.

Differential Diagnosis:

• Thanatophoric dysplasia: More severe bowing, narrow chest, poor ossification.

• Campomelic dysplasia: Anterior bowing of tibia/fibula, ambiguous genitalia possible.

• Diastrophic dysplasia: Hitchhiker thumb, clubfoot, cystic ear swelling.

• Osteogenesis imperfecta: Multiple fractures, hypomineralization.

Management Plan:

1. Confirmatory Testing:

   • Offer FGFR3 genetic testing (dominant mutation) by amniocentesis if diagnosis uncertain.

   • Counsel parents about inheritance (often de novo, but 50% recurrence if one parent affected).

2. Monitoring:

   • Serial ultrasounds for growth, AFI, and head circumference (often macrocephalic).

   • Assess fetal well-being and plan delivery at a tertiary center with neonatal support.

3. Delivery Planning:

   • Mode of delivery: Vaginal delivery often possible, but cephalopelvic disproportion may occur; cesarean may be required if fetal head size exceeds maternal pelvic dimensions.

4. Postnatal Care:

   • Neonatal evaluation to confirm diagnosis.

   • Early referral to genetics and orthopedics.

Ethical Considerations:

• Autonomy: Parents must be given accurate, balanced information—achondroplasia is non-lethal and compatible with a normal lifespan.

• Non-maleficence: Avoid framing as “abnormal” or “lethal.” Words matter.

• Justice: Families deserve equal access to expert counseling regardless of geography; AI-assisted image interpretation could bridge that gap.

• Beneficence: Encourage connection with patient support networks (e.g., Little People of America).

Summary:

Most likely diagnosis: Achondroplasia.
Rationale: Rhizomelic limb shortening, normal ossification, no fractures, normal chest, polyhydramnios.
Next steps: Confirm via FGFR3 testing, multidisciplinary perinatal care, delivery planning, and ethical counseling.

The Meta-Lesson

In this single exchange, the AI performed the entire intellectual arc of a case challenge: observation, reasoning, differential, testing, management, and ethics. It didn’t just choose the right answer—it articulated the why and the how.

What SMFM’s quiz presents as a game of recall, AI transforms into a masterclass in reasoning. The educational format can evolve: the point isn’t whether the AI knows, but whether we can understand what it knows.

Imagine if each SMFM “Case of the Week” included a “ChatGPT-5 analysis” sidebar. Learners could study the logic, question assumptions, and compare human versus machine reasoning. Instead of multiple choice, the challenge becomes interpretive: Do you agree with the AI’s rationale? Why or why not?

That is true diagnostic education—understanding, not guessing.

The Ethical Frontier

Is it ethical to use AI in prenatal diagnosis?

Today, the stronger argument is that it is unethical not to use AI.

When a technology demonstrably enhances diagnostic accuracy, speeds interpretation, reduces disparities, and helps explain findings in language families can understand, deliberately avoiding it violates the physician’s duty to beneficence and nonmaleficence.

The core ethical obligation of medicine is to use the best available means to prevent harm. A clinician who ignores AI-assisted tools, despite knowing they outperform unaided human interpretation in certain pattern-recognition tasks, is choosing a lower standard of care. This is not professional humility; it is ethical inertia. If AI can flag skeletal dysplasia earlier, identify anomalies that could otherwise be missed, or prepare families for counseling with greater clarity, then refusing to integrate it is no different from refusing to use an improved diagnostic test because one prefers the old microscope.

AI is not a replacement for the physician’s judgment—it is a moral assistant that strengthens it. Used responsibly, it widens access to expertise, supports overworked clinicians, and makes high-quality prenatal imaging interpretation available in underserved regions where MFM specialists are scarce. In that sense, AI becomes an equalizer, not an intruder. To withhold such a tool from patients because of habit, pride, or fear of change is to violate the principle of justice.

Of course, ethical use demands oversight. Clinicians must understand when the model’s confidence is justified, when to override it, and how to explain uncertainty transparently. Accountability remains human. But to reject AI altogether when it can reduce diagnostic errors, improve counseling, and spare families unnecessary suffering would be a failure of professional duty.

Subscribe now

The true ethical risk today lies not in using AI, but in refusing to—thereby accepting preventable errors, slower recognition, and less informed care. History has seen this pattern before: skepticism toward antisepsis, ultrasound, and even fetal monitoring delayed their adoption and cost lives. In the same way, ignoring validated AI assistance in obstetric imaging is no longer a neutral choice. It is an ethically indefensible one.

There are many precedents where not using available technology became ethically unacceptable once its benefits were proven.

Refusing to perform ultrasound in pregnancy, when it can detect life-threatening conditions like placenta previa or anencephaly, would be indefensible.

Declining electronic fetal monitoring during labor in high-risk cases would ignore decades of evidence linking timely intervention to improved neonatal outcomes.

Omitting Rho(D) immune globulin after an Rh-negative pregnancy, or refusing to use antibiotic prophylaxis for cesarean delivery, would likewise violate the duty to prevent avoidable harm. In the same way, neglecting validated AI diagnostic assistance when it can enhance accuracy and patient understanding now falls within that same ethical category—a failure to use the best tools medicine has to protect mothers and babies.

Share

Reflection / Closing

The “Ultrasound Case of the Week” once tested what we could see. Soon, it may test how we think alongside AI. The ultimate goal is not to outguess the machine but to outgrow our dependence on guessing. The next generation of maternal-fetal medicine will belong to those who can merge evidence, empathy, and ethics—whether the first diagnosis comes from a human hand or a silicon brain.

What vibe coding is, why it matters, and why an ObGyn professor is further along than most health systems.

A resident came to me last year with a clinical question she had already run through a large language model. The answer was good. Her instinct to use AI was correct. What surprised me was not her question but her hesitation to admit where she had looked first.

She was using the technology but treating it like a secret.

I understand the hesitation. Medicine moves slowly, and the slowest part moves fastest right now. Something has changed in the last year that most clinicians have not yet named. The name is vibe coding.

What Vibe Coding Is

The rest of this post is for paid subscribers.

Read more

If I can see the file, it exists.

If it is in iCloud, it is protected.

If it appears on more than one computer, it must be backed up.

That is what I thought.

After more than 40 years of using Apple computers, my MacBook crashed. Only then did I discover that iCloud had silently stopped updating some of my files for about a month. I had assumed that because my work was in iCloud, it was safe. I saw them in my iCloud folders. I saved them there. All my folders were in iCloud. Or so I thought.

They were not.

The problem was not that I had ignored computers. The problem was that the system uses familiar words in confusing ways.

It’s what I teach about patient safety. Rarely the individual, mostly the system.

Most people hear “sync” and think “backup.” That is understandable. If a file moves from a laptop to iCloud and then appears on another laptop, it feels protected. It feels as if there is another safe copy. Big Mistake!

But the real test is not what the file looks like on a normal day.

The real test is what happens when something goes wrong.

If I accidentally delete the file, is there a separate recovery copy?

If I overwrite the file, can I get the older version back?

If my laptop stops sending files to iCloud, will I know? (Hint: You won’t necessarily know)

If a file appears in Finder but is not physically downloaded to my computer, will my backup service copy it?

This is where the difference becomes painfully real.

iCloud is excellent at making files available across devices. That is its main job. A true backup has a different job: it keeps a separate recoverable copy when the current file is lost, damaged, deleted, overwritten, or never uploaded.

A simple way to say it is this:

iCloud helps your files appear in more places.

A backup helps your files survive when something goes wrong.

That difference should be obvious. It is not. Apple does not make it obvious. The icons are small. The settings are confusing. The word sync sounds reassuring. The word cloud sounds even more reassuring. But reassurance is not the same as recovery.

For physicians, professors, researchers, editors, writers, and students, this matters. Our files are not just files. They are manuscripts, lectures, peer reviews, grant drafts, response letters, legal documents, patient safety projects, and years of intellectual work.

This article is for people like me: educated, busy, careful, and still surprised to learn that seeing a file, storing a file in iCloud, and having a true backup are not the same thing. Trust me, I know it. Following the explanations and instructions below will save you many times the price you pay for a yearly ObGyn Intelligence subscription.

I paid many more times the price.

Read more

That is not because we are careless. It is because the system is designed to make us stop thinking about it.

We open a laptop. We see a folder. We see a file. The same file may appear on another computer. It may be visible from a phone. It may be in OneDrive, Google Drive, Dropbox, Box, or another cloud service. The natural conclusion is simple:

The file is safe.

That conclusion may be wrong.

This is not only an Apple or iCloud problem. The same confusion exists on Windows computers and in nearly every cloud storage system. The names change, but the risk remains the same.

On a Mac, the file appears in Finder.

On a Windows computer, the file appears in File Explorer.

In both cases, the file may look like it is on the computer. But sometimes it is not fully stored there. It may be mostly in the cloud. It may be online-only. It may be waiting to download. It may be visible as a placeholder. It may be available only if the internet connection works.

Subscribe now

That distinction is easy to miss.

Most people hear the word “sync” and think “backup.” That is understandable. If a file moves from my computer to the cloud and then appears on another computer, it feels backed up. It feels as if there is now another safe copy.

But the real test is not what happens on a normal day.

The real test is what happens on a bad day.

• If I accidentally delete the file, is there a separate recovery copy?

• If I overwrite the file, can I recover the older version?

• If my computer stops uploading files to the cloud, will I know?

• If a file appears in File Explorer but is not physically downloaded, will my backup service copy it?

• If the cloud service changes a file everywhere, do I still have an independent copy from before the change?

This is where the difference becomes real.

Cloud services are excellent at making files available across devices. That is their main job. They let us work from a desktop, laptop, tablet, phone, office computer, or browser. That convenience is real.

But convenience is not the same as protection.

A simple way to say it is this:

Cloud sync helps your files appear in more places.

A backup helps your files survive when something goes wrong.

Those two things overlap in the mind of the user. They do not always overlap in real life.

That is the problem.

This article is for people who use non-Apple computers, Windows laptops, office desktops, hospital computers, university computers, or cloud systems such as OneDrive, Google Drive, Dropbox, and Box. It is also for people who think they are “backed up” because their files are “in the cloud.”

That may be partly true.

But it may not be safe enough.

If your professional life depends on documents, the question is not whether your files appear in the cloud. The question is whether you can recover them when the cloud, the computer, or your own assumptions fail.

The rest of this article is the practical instruction and checklist I wish I had used before losing a month of work.

The yearly subscription pays off many times more of what you could lose without these instructions.

It may save you from learning the same lesson after the crash.

Read more

Subscribe now

In 30 years of clinical practice, I will encounter perhaps 6 of the 47 causes of lymphadenopathy I memorized in medical school. The other 41 occupy space where something more useful could live.

That is not a flaw in my memory. It is a design flaw in how we trained a profession.

For more than a century, medical education has rested on a single assumption: the doctor is the database. Every diagnosis, every drug interaction, every rare syndrome stored in one human brain. Exams rewarded recall. Residency rewarded the speed of recall.

The best doctor was the one who remembered the most.

For most of that century, this made sense. If you did not know it, you could not look it up at the bedside. The encyclopedia model was the only model available.

That era is over. And medicine has been slow to notice.

The Scale of the Problem

There are now more than 40,000 active biomedical journals. PubMed adds more than one million new citations every year. No physician has ever been able to keep up with this volume. We have known this for decades. We just pretended otherwise, and patients paid for the pretense with care shaped by what their doctor happened to have memorized last.

Share

Daniel Kahneman spent a career documenting how human judgment fails under conditions exactly like these. Physicians over-rely on what they have seen most recently. They anchor on the first diagnosis they consider. They remember dramatic cases and forget base rates. This is not a moral failure. It is how human cognition works under cognitive load and information overload.

AI does not have these biases. Used properly, it is not a replacement for physician judgment. It is a check on the ways physician judgment goes wrong.

What AI Actually Changes

Tools like Claude and ChatGPT can retrieve drug interactions, diagnostic criteria, guideline updates, and rare disease presentations in seconds, with citations. This does not replace the physician. It replaces the part of the job that was never really medicine: storage and retrieval of facts.

What medicine actually requires at its best is something different. A physician listens to a patient’s story. She notices what does not fit. She weighs risks against benefits for this specific patient, not the average patient in a trial population. She explains a difficult diagnosis to a frightened family. She makes a judgment call when the evidence is incomplete.

None of that requires memorizing 47 causes of lymphadenopathy. All of it requires presence, reasoning, and relationship. Those are the capacities medical training has historically crowded out in favor of encyclopedic recall.

When a physician does not have to spend half the appointment recalling whether a medication is safe in pregnancy, she can ask how the patient is actually doing. The doctors who embrace AI will not be lazier. They will be more present.

The Predictable Objection

Every time I raise this argument, someone says: doctors need to know things. You cannot just look everything up.

My father, who had to leave school at age 14, told me when I entered medical school: you do not need to remember everything, just know where to find the answer. That was in 1968, more than 50 years before ChatGPT.

He was right then. He is more right now.

Of course, core knowledge matters. A surgeon needs anatomy. An obstetrician needs fetal physiology. I am not arguing for ignorance. I am arguing for a different kind of knowledge: frameworks for thinking, not warehouses of facts. The ability to reason through a problem is not the same as the ability to recall a list.

What This Means for You

If your doctor uses AI-assisted tools to check a drug interaction or look up a rare presentation during your visit, that is not a sign of incompetence. It is a sign of good judgment. It means she is not relying on a memory formed years ago. It means she is checking current evidence for your specific situation.

The physicians who resist these tools are not protecting their patients. They are protecting the identity they built around memorizing things. These are not the same thing. Ask your doctor whether they use AI tools in your care. The answer will tell you something important.

Conclusion

I have practiced medicine for more than 50 years. I have forgotten more facts than most young physicians have yet learned. I am a better clinician now than I was at 40, not because I remember more, but because I know what questions to ask and how to reason toward an answer. AI handles the memory. The physician handles the medicine. These are not competing claims. They are a division of labor that medicine has needed for decades and that education has refused to acknowledge. The profession will come around. The technology does not wait for permission.

Bottom Line

Medical training built doctors into hard drives when it should have built them into processors. AI now gives us a better hard drive. The question is whether medical education has the courage to retrain for the job that has always actually mattered.