AI has quietly become a standard part of modern medicine — over 1,450 AI-enabled devices have been authorized by the FDA since tracking began, the overwhelming majority in radiology — and after going through the 2026 clinical evidence on diagnosis, treatment recommendations, and the documented harms, the picture is sharper and more uncomfortable than either the hype or the panic suggests. AI genuinely catches things human clinicians miss. AI also generates medical errors with the same confident tone whether it's right or catastrophically wrong, and most safeguards haven't caught up to that fact yet.
This is part of an ongoing series looking past the marketing at what the actual research says. As with the previous piece on work, education, and social connection, treat these numbers as where the evidence stands in 2026, not a final verdict — this field is moving fast enough that next year's data could shift the picture meaningfully.
Quick Comparison: Where the Evidence Actually Stands
| Domain | The Evidence For | The Documented Risk | What the Data Actually Shows |
|---|---|---|---|
| Diagnosis | AI-assisted radiologists significantly outperform either humans or AI alone in detecting hemorrhage on brain CT | Standalone AI accuracy drops when moving from retrospective testing to real prospective clinical use | AI as a second opinion beats AI as a sole decision-maker |
| Treatment Recommendations | Correct AI-suggested diagnoses meaningfully raise physician accuracy | Physicians also adopt incorrect AI suggestions at a real, measured rate (automation bias) | AI recommendations help only when clinicians retain genuine scrutiny |
| Prescriptions & Clinical Notes | Reduces documentation burden and administrative workload for clinicians | Hallucinated dosages, fabricated differentials, and invented citations appear with alarming frequency | Useful for drafting, dangerous if trusted without verification |
Diagnosis: AI Is a Genuinely Strong Second Opinion, a Riskier Solo Act
The clearest, best-supported finding in medical AI research right now is also the most boring one: AI works best paired with a human, not instead of one. A prospective, multicenter study across 67 medical organizations in Moscow, covering 3,409 brain CT scans with over 1,100 confirmed hemorrhage cases, found that radiologists assisted by AI significantly outperformed standalone AI services across every diagnostic metric measured — sensitivity reached 98.91% for AI-assisted radiologists versus 95.91% for AI alone.
That gap matters more than it might look on paper, because commercial standalone AI systems already post genuinely impressive numbers in controlled settings — 85% to 93% sensitivity and 93% to 99% specificity for hemorrhage detection under ideal conditions. The problem is that "ideal conditions" and "real hospital workflow" aren't the same thing. The same Moscow research program documented a measurable drop in AI diagnostic accuracy specifically when transitioning from retrospective validation to live, prospective clinical use — a gap that systematic reviews of emergency neuroimaging AI have flagged as a recurring, unresolved issue.
The FDA's own numbers reflect just how concentrated this technology still is: of more than 1,450 AI-enabled devices authorized since the agency started tracking them, roughly three out of every four are radiology tools, according to the FDA's AI-Enabled Medical Device List. That's not a coincidence — imaging is the use case where AI's pattern-recognition strengths line up most cleanly with a well-defined task, and it's also the specialty with the most mature track record of clinical validation.
Treatment Recommendations: Helpful Until the AI Is Confidently Wrong
A randomized controlled study of 22 physicians at a university hospital tested something specific: does giving doctors an AI-generated differential diagnosis list actually improve their diagnostic accuracy? The headline result was almost a wash — 57.4% accuracy with the AI list versus 56.3% without, not a statistically meaningful difference. But buried inside that null result is the real finding: when the AI's suggested list happened to contain the correct diagnosis, physician accuracy jumped dramatically (adjusted odds ratio of 7.68). The catch is what happened on the other side of that coin — 15.9% of cases produced omission errors, where physicians wrongly rejected a correct AI suggestion, and 14.8% produced commission errors, where physicians wrongly accepted an incorrect one.
This pattern — called automation bias — shows up consistently across newer studies too. Research on LLM-assisted diagnostic reasoning found that when test vignettes contained even a single incorrect detail, leading AI models produced hallucinated reasoning in 50% to 82% of cases, and physicians using those tools didn't reliably catch the error. Unlike older AI systems that gave a discrete answer with a confidence score attached, newer language-model-based tools generate fluent, narrative-style recommendations that read as authoritative whether or not they're correct — which makes the underlying error harder to spot, not easier.
Prescriptions & Clinical Documentation: Where Hallucination Risk Is Highest
This is the domain where the gap between "useful tool" and "genuine patient safety risk" is widest. A 2025 global survey of 70 clinicians across 15 specialties found that 91.8% had personally encountered a medical AI hallucination while using these tools in their own work, and 84.7% of those clinicians believed the hallucination they witnessed was capable of causing direct patient harm. One frequently cited real-world example: a model confidently stating a methotrexate dose of 25mg daily for rheumatoid arthritis, when the correct regimen is 7.5–25mg weekly — a dosing error serious enough to cause real harm, delivered with exactly the same fluent tone the model uses when it's correct.
Counterintuitively, research evaluating eleven foundation models across medical hallucination tasks found that general-purpose AI models actually produced hallucination-free responses more often than models specifically fine-tuned for medical use, suggesting narrow medical training can sometimes introduce its own failure modes rather than eliminate them. The downstream effects are already visible in the published scientific record — a Lancet systematic review of 97.1 million verified references in biomedical literature found fabricated citation rates climbing more than twelvefold between 2023 and 2025, and a separate analysis reported by Fortune found that in the first seven weeks of 2026 alone, 1 in every 277 papers contained at least one non-existent reference.
There's also a documentation-specific risk that doesn't get discussed enough: AI scribes and note-generation tools can quietly fabricate the appearance of clinical thoroughness — auto-generating a tidy differential diagnosis list that reads as a careful workup the clinician never actually performed. Legal and risk-management analysis has flagged this directly: physicians remain ethically and legally responsible for everything documented under their name, whether they typed it or the AI did, and "the documentation tool added that" is not a defense that holds up under scrutiny.
So What Should Patients and Clinicians Actually Take From This?
Across all three domains, the same structural pattern repeats: AI adds real value as a second check on human judgment and becomes genuinely dangerous the moment it's treated as a replacement for that judgment.
- For diagnosis, the evidence supports AI-assisted review, not AI-only review — the accuracy gap between the two is large and well-documented, particularly in real-world prospective use rather than retrospective testing.
- For treatment recommendations, AI lists help most when a clinician still independently evaluates each suggestion rather than defaulting to it — automation bias is a measured, real phenomenon, not a theoretical worry.
- For prescriptions and documentation, verification isn't optional. With hallucination rates reaching 50–82% on vignettes containing even one wrong detail, and the overwhelming majority of practicing clinicians reporting they've personally seen a harmful hallucination, this is the domain demanding the most active human oversight, not the least.
None of this argues for abandoning AI in medicine — the radiology data alone makes a strong case against that. It argues for treating "AI-assisted" and "AI-replaced" as fundamentally different categories of risk, and for being honest that right now, regulation, training, and verification habits are still catching up to how fast these tools have moved into actual patient care.
Frequently Asked Questions
Is AI more accurate than doctors at diagnosis?
Not on its own, according to the strongest available evidence — a large multicenter study found AI-assisted radiologists significantly outperformed standalone AI across every diagnostic metric measured, meaning the combination of AI plus a human reviewer beats either one alone.
How often does medical AI hallucinate incorrect information?
Rates vary by task and model, but research testing leading AI models on clinical vignettes containing a single incorrect detail found hallucinated reasoning in 50% to 82% of cases, and a separate clinician survey found over 90% of respondents had personally encountered a medical AI hallucination in their own practice.
Are AI medical devices actually approved by regulators?
Yes — the FDA has authorized over 1,450 AI-enabled medical devices since it began tracking them, with roughly three-quarters concentrated in radiology, though the vast majority go through the lighter 510(k) clearance pathway rather than more rigorous premarket approval.
Can doctors be held legally responsible for AI-generated errors?
Yes. Legal analysis is clear that physicians remain accountable for everything documented under their name regardless of whether an AI tool generated it, and several states have already enacted laws specifically addressing AI use in clinical communication and care.
Should patients worry about AI being used in their diagnosis?
The evidence suggests AI used to support a clinician's judgment is a net positive, particularly in imaging-heavy specialties like radiology, while AI used as an unsupervised replacement for clinical judgment carries documented, measurable risk — the distinction between those two uses matters more than whether AI is involved at all.