How common are medical AI errors in actual practice?

It depends on the tool and the task. In a real ambulatory deployment, 18% of audited notes had an omission, 11.5% had a hallucination, and 5.3% contained an error judged capable of serious harm [2]. In a controlled summarization benchmark with a purpose built pipeline, hallucination was 1.47% and omission 3.45% at the sentence level [5]. Neither range is high. Neither is zero.

Which is more dangerous, fabrication or omission?

Omission, for most clinicians. Fabricated content is alarming to read but easier to catch, because it is foreign to your own recollection. A dropped finding produces a note that reads cleanly and is simply incomplete, and nothing on the screen signals the gap. Omission also dominates numerically in every scribe evaluation cited here [3][4].

Does retrieval augmented generation eliminate hallucinations?

No. Published reductions for RAG in healthcare settings cluster at 30% to 50% compared with ungrounded models [11]. That is a meaningful improvement and a poor substitute for review. Grounding constrains what a model draws on; it does not force it to reason correctly about what it retrieved.

If the AI shows me real sources, can I trust the answer?

Not automatically. A 2026 comparison found the model with 100% verifiable citations also carried the highest severe safety flag burden in that study [10]. Citation accuracy and clinical accuracy are separate axes. Open the source and check that it says what the summary claims it says.

Does telling the model to be careful help?

Partially. A prompt warning that some input may be inaccurate dropped hallucination rates from 66% to 44% overall in one controlled study, and from 53% to 23% for the strongest model tested [6]. Setting temperature to zero did almost nothing [6]. Better prompting is worth the thirty seconds and is not a safety control.

Who is responsible for an error that reaches the chart?

The signature on the note is yours. Whatever the tool did, attestation is a clinical act, which means the review step is not optional even when the tool has been right a hundred times in a row.

Back to blogs

AI Hallucinations in Medicine: Catch Them Early

Medical AI errors split into fabrication and omission. Here is what the published rates show, plus a 60 second checklist to catch them before you sign.

Reviewed by O. Daniel Odutola, MD, MBA, MPHNovember 5, 202614 min read

Before They Reach the Chart

AI hallucinations in medicine are outputs that read like clinical fact but have no basis in the source: a value that was never drawn, a symptom the patient specifically denied, a citation to a paper that does not exist. In real ambulatory deployments the rate is low but never zero, and the error most likely to hurt someone is usually the one that is missing rather than the one that is wrong. Below is where the published numbers actually sit, why the failure happens, and a verification pass short enough to run before you hit sign.

The word "hallucination" is doing you a disservice

Nothing is being perceived. A closer clinical analogue is confabulation, the fluent and unbothered account of events that did not occur. The output is not hedged and does not look different from the parts that are correct. A wrong lab value in a note does not arrive in red.

Three separate failures get filed under the same word, and they fail in different places:

Transcription error. The speech model produces words nobody said. In the FAccT 2024 "Careless Whisper" analysis, roughly 1.4% of audio segments yielded a hallucinated sequence, and of 312 hallucinated transcriptions, 38% contained an identifiable harm, including invented violence (19%) and false authority such as fabricated clinical instruction (8%) [1]. The rate was higher for speakers with aphasia (1.7%) than for controls (1.2%) [1]. Read that clinically: disfluent speech, dysarthria, and heavy background noise raise your risk of a fabricated line.

Summarization error. The transcript was fine, the note is not. This is where most ambient scribe errors live.

Knowledge and reasoning error. You asked a clinical question and the model asserted something false with full confidence, sometimes with a citation attached.

If you are still deciding how these tools fit into your week at all, start with the broader picture in our guide to AI in clinical medicine for working physicians, then come back here for the error side.

How often AI hallucinations in medicine actually reach the chart

The best data comes from health systems that audited their own deployments, not from vendor benchmarks.

In a UC Davis Health pilot, 31 physicians across multiple specialties generated 7,545 clinic notes with an ambient AI scribe over two months in 2024, and physicians formally evaluated 356 of them [2]. Accidental omissions were the most common error at 18% of evaluated notes, followed by hallucinations at 11.5% and accidental inclusions at 9.3% [2]. Most errors were mild to moderate. But 19 notes, 5.3% of those reviewed, contained an error the reviewing physician judged capable of serious or imminent harm if left uncorrected [2].

The number from that study that should actually change your behavior is a different one. Across the editing metrics, the median note had 9.0% of its AI generated words changed, and 14.9% of notes (143 of 960) were signed with no words changed at all [2].

Roughly one in seven notes went into the record untouched. If the serious error rate is 5% and the untouched rate is 15%, those two numbers eventually intersect on somebody's chart.

A four vendor comparison presented in 2025 found the same shape. Across commercial AI scribes evaluated on simulated internal medicine and surgical encounters, omission errors made up 71% of all errors identified, additions 19.4%, and outright incorrect facts 6.5% [3]. The vendor producing the shortest notes had the highest omission rate at 24.4% [3].

General purpose models used raw, without scribe engineering, do considerably worse. Asked to build SOAP notes from transcribed encounters, GPT-4 averaged 23.6 errors per case, 86% of them omissions [4]. Running each case three times, only 52.9% of scorable data elements were correct in all three replicates [4]. Same input, same prompt, different note. A tool that behaved well in your pilot is not making a promise about tomorrow.

Purpose built summarization pipelines do much better. A 2025 npj Digital Medicine framework evaluated GPT-4 against 12,999 clinician annotated sentences from 450 consultation transcript and note pairs, and found an overall hallucination rate of 1.47%, with 44% of those classified as major, alongside an omission rate of 3.45% [5].

Two things are true at once. These systems are good. They are not good enough to sign unread.

Omission is the error you are least likely to notice

Every data set above points the same direction: what the note leaves out beats what the note invents, usually by a wide margin.

This is a perception problem more than a technology problem. A fabricated sentence is a foreign object in your own note and your eye snags on it. A missing sentence is invisible. Nothing on the screen tells you the patient mentioned two weeks of exertional dyspnea somewhere in minute nine and the summary dropped it.

There is a quieter version of this. In the npj Digital Medicine breakdown, negation errors made up 30% of all hallucinations, second only to outright fabrication at 43% [5]. The polarity flipped. "Denies chest pain" becomes "reports chest pain," or worse, the reverse. Every review of systems you dictate is a string of negatives, and a flip on any one of them is a single word change that reads perfectly.

We cover the deployment side of this in more depth in our guide to AI medical scribe safety. The short version for the individual clinician is that you cannot audit for omissions by reading the note. You audit for omissions by remembering the visit.

Three failure modes, three different catches

Failure mode	What it looks like in the chart	What the evidence shows	Fastest way to catch it
Fabrication	A finding, value, or exam element that was never generated	43% of hallucinations in GPT-4 summarization [5]; 11.5% of audited scribe notes contained a hallucination [2]	Scan for any objective data point you do not personally remember producing
Negation flip	A denied symptom recorded as present, or the reverse	30% of hallucinations in the same analysis [5]	Read the review of systems and the pertinent negatives first, before the narrative
Omission	A real finding absent from the note	71% of all errors across four commercial scribes [3]; 86% of errors in raw GPT-4 notes [4]	Recall the visit and check for the two or three things that drove your plan
Invented entity	A test, sign, or syndrome that does not exist	Models echoed a fabricated clinical term in 50% to 82% of cases at default settings [6]	Stop on any proper noun you do not recognize and look it up before you repeat it
Fabricated citation	A plausible reference that resolves to nothing	16% of references in one evaluation were entirely fabricated, with another 48% containing errors [7]; two major models "hallucinated at least half of their references" on otolaryngology questions [8]	Open one citation, not all of them, and see whether it exists and says what was claimed

The invented entity problem, and why polite prompting only half works

A Mount Sinai group ran a clean experiment on this in 2025. They built 300 physician validated clinical vignettes, each seeded with exactly one fabricated element: a fake lab ("Serum Neurostatin"), a fake physical or radiologic sign ("Cardiac Spiral Sign"), a fake syndrome ("Faulkenstein Syndrome") [6]. Then they watched what six large language models did with it.

At default settings, models elaborated on the fake finding in 50% to 82% of cases [6]. They did not push back. They explained the pathophysiology of a syndrome that does not exist.

Two mitigations were tested. A prompt explicitly warning that some information might be inaccurate dropped the overall rate from 66% to 44%, and for the strongest model from 53% to 23% [6]. Setting temperature to zero, the thing engineers reach for first, left the rate near 66.5% [6].

The practical read: a well built prompt roughly halves the problem and does not come close to solving it. That is worth doing anyway, and we walk through the specific phrasing in our piece on clinical AI prompting for physicians. It is not a substitute for a human reading the output.

The mechanism is not mysterious. A team including OpenAI and Georgia Tech researchers argued in September 2025 that models hallucinate largely because standard training and evaluation reward guessing over admitting uncertainty, the way a multiple choice exam rewards a guess over a blank [9]. The model is not malfunctioning when it invents a plausible syndrome. It is doing exactly what it was scored to do.

See how Physicians' Copilot reduces error risk.

The 60 second pass: how to verify AI output before you sign

You will not re-derive the note, and a tool that required it would save you nothing. What you need is a short fixed sequence aimed at the error types the literature says are actually happening.

Read the negatives first. Go to the review of systems and the pertinent negatives before you read the narrative. Polarity flips are a third of the hallucination burden and they are invisible once you have already read the story the note is telling [5].
Check numbers against your own memory, not the note's internal logic. A summary can be perfectly self consistent and wrong. Doses, vitals, durations, and pain scores are the ones that propagate.
Ask what is missing, not just what is wrong. Before reading, name the two or three findings that actually drove your assessment and plan. Then confirm each one appears. This is the single highest yield step, because omission is the dominant error in every data set cited here [2][3][4].
Stop on any proper noun you do not recognize. A test you have never ordered, an eponymous sign you have never heard of, a syndrome that sounds almost familiar. Models will confidently build on entities that do not exist [6].
Open one citation. One. If a tool gives you references, spot check a single one every time rather than all of them occasionally. A reference that does not resolve tells you something about the whole response.
For anything that will change management, re-ask cold. New session, no prior context, same question. In one evaluation only 52.9% of data elements were correct across three runs of the same case [4].
Never let AI generated text be the first record of something you did not say out loud. If the note contains an exam finding, and you cannot remember examining that, it does not go in. This is the rule that holds when the other six get skipped on a busy day.

Steps 1, 3, and 4 are the ones to keep when you are behind. They map to negation, omission, and invented entities, which together account for most of what will actually hurt a patient.

AI fabrication in clinical answers: the citation problem is not what you think

The obvious version is well documented. Of 25 references in one AI drafted review article, 36% were accurate, 48% had a correct title but other errors, and 16% were completely fabricated [7]. A 2026 study putting 36 common otolaryngology questions to two leading models found both "hallucinated at least half of their references" [8].

The less obvious version matters more. A 2026 cross sectional study compared five models on pediatric perioperative questions and found that the tool with perfect citation performance, 100% verifiable references and zero fabrications, also carried the largest burden of severe clinical safety flags, five of the nine identified in the study [10]. The authors' conclusion is the one to carry with you: "High citation accuracy does not guarantee clinical safety" [10]. This was a single center evaluation of one condition conducted in April 2025, so treat it as a caution rather than a verdict on any product. It still undercuts the heuristic most of us use, which is that a tool showing sources must be trustworthy.

Citations tell you the model retrieved something. They do not tell you it reasoned correctly about it.

What actually improves AI accuracy in healthcare settings

A 2026 systematic review of 44 studies on hallucination mitigation gives the honest ranking [11]:

Strategy	Reported reduction in hallucinations	Practical note
Human in the loop review	Up to 95% [11]	Most effective, least scalable. This is you.
Retrieval augmented generation (RAG)	30% to 50% [11]	Grounds output in a document set. Does not eliminate the problem.
Knowledge graph integration	25% to 45% [11]	Strongest in structured question answering
Self reflection and chain of medical thought	38% reduction in catastrophic errors [11]	Model checks its own work, with the obvious limitation
Specialized training methods	15% to 25% [11]	Vendor side, invisible to you
Red teaming	20% to 40% better detection [11]	Improves detection, not generation

Two things jump out of that table. Nothing reaches zero. And the single most effective control in the published literature is a clinician reading the output, which is the one control that gets quietly removed the moment a tool starts working well enough to trust.

That is the trap. Reliability breeds the inattention that reliability cannot survive.

The AMA's 2026 survey of nearly 1,700 physicians found 81% now use AI professionally, more than double the 2023 figure, and 88% reported at least some concern about AI related skill loss [12]. That concern is well placed, and it applies to the verification skill itself.

Key takeaways

Omission, not fabrication, is the dominant error mode. It made up 71% of all errors across four commercial AI scribes [3] and 86% in raw GPT-4 generated notes [4].
About 5.3% of audited scribe notes in one deployment contained an error capable of serious harm, while roughly 15% were signed with zero words changed [2].
Negation flips, where a denied symptom becomes a reported one, account for 30% of hallucinations in summarization and are nearly invisible on a normal read [5].
Models elaborate confidently on entities that do not exist, in 50% to 82% of seeded cases; a warning prompt cuts that roughly in half but does not fix it [6].
Verifiable citations are not clinical safety. The model with perfect reference accuracy in one 2026 evaluation carried the heaviest severe safety flag burden [10].

No AI tool replaces clinical judgment. Verify all outputs against primary sources and your own assessment.

If you want tooling that makes the verification step faster instead of quietly removing it:

See how Physicians' Copilot reduces error risk.

REFERENCES

Koenecke A, Choi ASG, Mei KX, Schellmann H, Sloane M. "Careless Whisper: Speech-to-Text Hallucination Harms." Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT '24). June 2024. https://dl.acm.org/doi/10.1145/3630106.3658996 (accessed September 23, 2026)
Taylor SL, Jost M, MacDonald S, Ren Y, Hilton S, Davenport S, Aizenberg D, Hall B, Lyles CR, Adams JY. "Quality of Clinical Notes Created by Ambient Listening Generative AI: Pragmatic Prospective Pilot Study." JMIR Medical Informatics, vol 14, e86474. Published April 17, 2026. DOI: 10.2196/86474. https://doi.org/10.2196/86474 (accessed September 23, 2026). Retrieved via PubMed, PMID 41996389.
Arko L IV, Hudelson C, Kumar J, Badlani S, Stoffel M, Markowitz R, Melton GB. "Documenting Care with AI: A Comparative Analysis of Commercial Scribe Tools." Studies in Health Technology and Informatics, vol 329, pp 337 to 341. Published August 7, 2025. DOI: 10.3233/SHTI250857. https://doi.org/10.3233/SHTI250857 (accessed September 23, 2026). Retrieved via PubMed, PMID 40775875.
Kernberg A, Gold JA, Mohan V. "Using ChatGPT-4 to Create Structured Medical Notes From Audio Recordings of Physician-Patient Encounters: Comparative Study." Journal of Medical Internet Research, vol 26, e54419. Published April 22, 2024. DOI: 10.2196/54419. https://doi.org/10.2196/54419 (accessed September 23, 2026). Retrieved via PubMed, PMID 38648636.
"A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation." npj Digital Medicine. Published May 13, 2025. https://www.nature.com/articles/s41746-025-01670-7 (accessed September 23, 2026)
Omar M, et al. "Multi-model assurance analysis showing large language models are highly vulnerable to adversarial hallucination attacks during clinical decision support." Communications Medicine, Hasso Plattner Institute for Digital Health, Icahn School of Medicine at Mount Sinai. Published August 2, 2025. https://www.nature.com/articles/s43856-025-01021-3 (accessed September 23, 2026)
Safrai M, Orwig KE. "Utilizing artificial intelligence in academic writing: an in-depth evaluation of a scientific review on fertility preservation written by ChatGPT-4." Journal of Assisted Reproduction and Genetics, vol 41(7), pp 1871 to 1880. Published April 15, 2024. DOI: 10.1007/s10815-024-03089-7. https://doi.org/10.1007/s10815-024-03089-7 (accessed September 23, 2026). Retrieved via PubMed, PMID 38619763.
Sakharkar M, Jalihal P, Chang S, Patel P, Chavez M, Levi J. "Large Language Model Responses to Common Otolaryngological Questions: Evaluating Accuracy, Appropriateness, Readability, and Hallucinations." Otolaryngology, Head and Neck Surgery, vol 175(3), pp 594 to 602. Published May 29, 2026. DOI: 10.1002/ohn.70309. https://doi.org/10.1002/ohn.70309 (accessed September 23, 2026). Retrieved via PubMed, PMID 42212563.
Kalai AT, Nachum O, Vempala S, Zhang E, Robinson D, Jain S, Mitchell E, Beutel A, Heidecke J. "Why Language Models Hallucinate." OpenAI and Georgia Institute of Technology. Published September 5, 2025. https://openai.com/index/why-language-models-hallucinate/ (accessed September 23, 2026)
Kang T, Yuan C, Hu X, Huang W. "Performance of 5 Large Language Models in Perioperative Consultation for Pediatric Hypospadias: Cross-Sectional Comparative Study." Journal of Medical Internet Research, vol 28, e93393. Published July 29, 2026. DOI: 10.2196/93393. https://doi.org/10.2196/93393 (accessed September 23, 2026). Retrieved via PubMed, PMID 42525870.
"Mitigating hallucinations in healthcare AI: a systematic review of evidence-based strategies." BMC Health Services Research. DOI: 10.1186/s12913-026-14851-1. https://link.springer.com/article/10.1186/s12913-026-14851-1 (accessed September 23, 2026). Retrieved via PubMed, PMID 42251377.
American Medical Association. "More than 80% of physicians use AI professionally: AMA survey." AMA Augmented Intelligence Research, 2026 Physician Survey on Augmented Intelligence. Published March 12, 2026. https://www.ama-assn.org/practice-management/digital-health/more-80-physicians-use-ai-professionally-ama-survey (accessed September 23, 2026)

See how your offer compares

Compare real pay data before you sign anything

Physicians' Copilot Clinical AI

FAQs

It is output that is fluent, confident, and unsupported by the source material or by fact. In clinical use it takes three forms: transcription of words nobody said, note content that is not in the encounter, and assertion of a clinical fact or reference that does not exist. None of it arrives flagged as uncertain.

FAQs

Related articles