The Hallucinating Physician Is Confabulating
There is a hallucination epidemic far more prevalent and far more dangerous that no one is racing to fix: the physician who, faced with a question s/he cannot answer, invents one.
We have been warned relentlessly about AI hallucinations. Large language models, we are told, generate confident, authoritative text that is entirely fabricated—invented citations, fictional studies, plausible-sounding nonsense delivered without hesitation. This is a legitimate concern. OpenAI, Anthropic, and Google have invested billions in guardrails, fact-checking layers, and uncertainty quantification to mitigate the problem.
But there is a hallucination epidemic far more prevalent and far more dangerous that no one is racing to fix: the physician who, faced with a question she cannot answer, invents one. That is called ‘Confabulation’.
Confabulation in the White Coat
It happens countless times every day in every hospital and clinic. A patient asks a question. The physician does not know the answer—or knows the honest answer is “I’m not sure” or “the evidence is unclear.” But instead of saying so, she constructs a response on the spot. It sounds authoritative. It is delivered with the confidence her training taught her to project. And it is completely untethered from evidence.
This is not lying in the traditional sense. The physician is not deliberately deceiving. She is confabulating—a term borrowed from neurology describing how the brain fills gaps in memory or knowledge with fabricated content, often without conscious awareness. The confabulator believes what she is saying. That is what makes it so dangerous.
In obstetrics, I have heard colleagues explain with complete certainty why continuous fetal monitoring prevents cerebral palsy (it does not), why episiotomy speeds healing (the opposite is true), why elective induction at 39 weeks lowers cesarean risk for all women (population data show the opposite effect), and why a particular cesarean was “necessary” when the indication was never documented and the tracing was category I.
These are not malicious fabrications. They are the brain’s reflexive response to an intolerable situation: a physician who is supposed to know, standing before a patient who expects her to know, unable to admit that she does not.
Why Physicians Hallucinate
Medical training creates the conditions for confabulation. From the first day of medical school, uncertainty is treated as weakness. The attending asks a question on rounds; the correct response is an answer, not “I don’t know.” The student who says “I’m not sure, let me look it up” is marked as underprepared. The resident who hesitates is told she lacks confidence. The physician who admits ignorance to a patient risks losing trust—or so we are taught.
The result is a profession that systematically selects for and reinforces false certainty. We learn to project confidence independent of actual knowledge. We learn that an authoritative wrong answer is professionally safer than an honest admission of uncertainty. We learn, in short, to hallucinate.
The electronic health record compounds the problem. Documentation demands certainty. Dropdown menus do not include “unclear” or “uncertain.” Billing codes require diagnoses. The system is architecturally hostile to epistemic humility. So physicians document with precision that their actual knowledge does not support, and over time, the documented fiction becomes the remembered reality.
More Dangerous Than Any AI
The hallucinating physician is more dangerous than any hallucinating AI for three reasons.
First, there is no audit trail. When ChatGPT fabricates a citation, you can check PubMed. When a physician fabricates a rationale, the patient has no recourse. She cannot Google her doctor’s confidence. She trusts the white coat. She assumes the certainty reflects knowledge. She makes decisions about her body and her baby based on explanations that were manufactured in the moment.
Second, there are no guardrails. AI companies are building elaborate systems to detect and flag uncertain outputs, to cite sources, to say “I don’t know” when appropriate. Medicine has no equivalent infrastructure. There is no real-time fact-checking of physician statements. There is no uncertainty quantification attached to clinical recommendations. There is only the assumption—increasingly unjustified—that the physician’s confidence correlates with the evidence.
Third, the hallucinating physician cannot be updated. A language model can be retrained, fine-tuned, corrected. A physician who has confabulated an explanation and repeated it for years has converted fabrication into conviction. She now genuinely believes the false explanation. She will teach it to residents. She will defend it against contradicting evidence. The hallucination has become doctrine.
The Obstetric Confabulation Epidemic
Obstetrics is particularly fertile ground for physician hallucination. We practice in a field where the evidence base is sometimes thin, where defensive medicine incentivizes intervention, and where tradition substitutes for data.
Ask an obstetrician why she recommends induction for a patient with an “aging placenta” at 39 weeks and you will receive a confident answer. Ask for the evidence that placentas deteriorate meaningfully between 39 and 40 weeks, and confidence will give way to bluster. The concept of placental senescence driving routine early delivery is a hallucination—widely shared, deeply believed, and largely unsupported by rigorous evidence.
Ask why a patient “needed” a cesarean for failure to progress after six hours of labor, and you will hear explanations about maternal and fetal safety. Ask for the evidence that six hours represents a meaningful threshold, and you will discover the number was invented—a convenient fiction that became policy because someone said it with authority and no one checked.
Ask why continuous electronic fetal monitoring is standard for low-risk patients despite five decades of randomized trials showing no benefit over intermittent auscultation, and you will hear about safety, liability, reassurance. You will not hear an honest answer: we do it because we have always done it, and stopping would require admitting that “always” was never justified.
The Cure for Confabulation
The cure is not technological. It is cultural.
We need to create medical environments where “I don’t know” is not weakness but intellectual honesty. Where “let me look that up” is not incompetence but rigor. Where admitting uncertainty to a patient is not a failure of confidence but an act of respect.
This requires unlearning what training taught us. It requires attending physicians who model uncertainty on rounds rather than punishing it. It requires documentation systems that permit epistemic humility rather than forcing false precision. It requires a professional culture that values being right over sounding right.
Most importantly, it requires remembering that the patient across from us is making irreversible decisions about her body based on what we say. She deserves our honesty more than our confidence. If we do not know, she has a right to know that we do not know. If the evidence is uncertain, she has a right to that uncertainty.
The hallucinating physician believes s/he is protecting the patient from confusion. In reality, she is protecting herself from discomfort, and transferring the cost of that comfort onto someone who trusted her.
Look it up before opening your mouth
We frequently warn about AI “hallucinations,” but we must be equally vigilant about the hallucinations of medical professionals who prioritize anecdote over evidence. Whether counseling a patient or testifying in court regarding fetal tracings, a doctor who relies on intuition or outdated memory rather than verifying current data is acting less than professionally. This type of “System 1” thinking, fast, emotional, and unchecked, creates a dangerous fog of misinformation that can alter patient choices and legal outcomes.
True expertise requires the humility to pause and verify; if you don’t look it up before you speak, you aren’t providing medical opinion, you are simply making things up.
I often told my patient to send in important questions at least 48-72 hours before the visit so I could thoroughly research them and answer them fully evidence-based.
The Mirror Is Coming
AI-empowered patients are coming. They will arrive with citations. They will ask questions we cannot bluff our way through. They will expose every hallucination we have been delivering as fact.
We can treat this as a threat and build better defenses—more sophisticated documentation, more elaborate justifications, more confident confabulation. Or we can treat it as an opportunity to finally become the physicians we should have been all along: honest about what we know, humble about what we do not, and rigorous enough to tell the difference.
Wisdom begins not with answers but with knowing how to ask. The first question every physician should ask, before opening her mouth to answer a patient, is simple: Do I actually know this, or am I about to make it up?
The answer to that question is the beginning of trustworthy medicine.
Have you witnessed—or caught yourself in—medical confabulation? How do we build a culture that rewards intellectual honesty over performed certainty? I welcome your reflections.