If you or a family member has ever been hospitalized, you’ve likely received a discharge summary. This single sheet of paper determines whether the next clinician can seamlessly take over care: what medications were changed, what tests remain pending, and what precautions patients should follow at home. If the summary is disorganized or omits critical information, patients risk redundant testing, medication errors, or even readmission.
The problem is that physicians have less and less time to write this document.
In mid-November, JAMA Network Open published a study that could change this—and it’s no isolated case. Across November alone, at least eight studies on large language models (LLMs) in healthcare were published in rapid succession. Taken together, they clearly answer one question: At which layers of clinical practice is AI already trustworthy—and where does it still fall far short?
## Blind Evaluation of Over 1,200 Discharge Summaries
The study methodology was robust: 1,294 real-world discharge summaries from inpatients across five London hospitals—645 generated by GPT-4o mini (using prompts co-designed by clinicians), and 649 written by physicians—were anonymized and randomized. Twelve practicing physicians then reviewed each summary against full admission records. The evaluation endpoint wasn’t technical metrics like text similarity; instead, it asked a practical, high-stakes question: **Would you sign your name to this summary?** The proportion of summaries deemed acceptable with only minor edits—or no edits required—was 22.2 percentage points higher in the LLM group than in the physician-written group.
Even more revealing were the error patterns. Among sections flagged as problematic by reviewers, the share rated as having “moderate or greater potential harm” was ~26% for the LLM group—but 45%–52% for the physician group. Critical information omission rates: 23.5% for LLMs vs. 39.8% for physicians. And regarding the most feared concern—“hallucination,” i.e., content appearing in the summary but absent from the medical record—the rates were nearly identical: 6.0% for LLMs versus 6.9% for physicians. In other words, physicians’ handwritten summaries also contain sentences unsupported by the chart—hallucination is not uniquely an AI flaw.
Here lies a fact easily obscured by headlines: the control group’s performance fell far below industry assumptions. Nearly half of the physician-written summaries flagged by reviewers carried moderate-or-greater potential harm—not because those physicians are incompetent, but because in today’s overburdened documentation workflow, the “current standard” itself has substantial room for improvement. So the correct interpretation of this study isn’t “AI outperforms physicians,” but rather: **In the most tedious, time-consuming documentation tasks, AI has reached a level of reliability warranting serious deployment—and incidentally exposed longstanding quality issues in human-written documentation.**
The authors themselves remained restrained, concluding only that LLM-generated summaries “meet, and in some aspects exceed, current standards of care”—with the crucial caveat of close clinical supervision.
## But When It Comes to Diagnosis, the Numbers Shift Dramatically
That same month, JMIR published a multicenter evaluation across ten Chinese hospitals: seven mainstream models were asked to diagnose 50 real patient cases. The top two models achieved diagnosis accuracy rates of just 64%. Yet when performing information extraction—e.g., pulling lab results from notes—the same models achieved 97.7% accuracy. The gap is stark.
A vascular surgery evaluation painted a similar picture: GPT-5, Gemini 2.5, and Claude Sonnet 4.5 scored above 83% overall across 42 cases—impressive at first glance. But when judged solely on “treatment safety,” unsafe recommendations ranged from 7% to 24%. The study also uncovered a counterintuitive signal: **the longer the model’s response, the higher the probability of an unsafe recommendation**—verbosity does not equal reliability.
Why? A benchmark called Med Evidence Bench offers a mechanistic explanation: when asked to recite clinical guidelines verbatim, all ten models scored above 0.97—near perfect. But when tested on real-world case scenarios requiring them to explicitly link claims to supporting evidence, scores plummeted to 0.18–0.45. Breaking it down further: identifying key information in notes? Models performed well (0.77). Matching that information to evidence (0.165) and formulating recommendations based on it (0.16)? Performance collapsed. In short: **models memorize guidelines fluently—but fail catastrophically when applying them to individual patients.**
There’s another shortfall few wish to confront: adversarial robustness. That same JMIR study conducted four rounds of adversarial “inductive attacks,” attempting to elicit medically incorrect responses. DeepSeek-R1 complied in 98% of attack attempts, refusing only 2%; even the best-performing model refused only 30%. This means in real-world deployment, attackers need not wait for the model to err spontaneously—they simply need to ask, and the model will comply. Already in August, the U.S. FDA’s discussion paper on generative AI medical devices listed “resistance to prompt injection” among core capabilities expected of intelligent agents—regulatory awareness, this time, is not lagging.
## Even the Judges Aren’t Fully Reliable
Another study—easily overlooked in November, yet especially relevant for writers like me—appeared in npj Digital Medicine. It tasked six open-source models with acting as “AI judges” evaluating German-language clinical texts, then compared their ratings to those of expert human panels. Results: no configuration proved consistently reliable; all AI judges exhibited positional bias (favoring items seen first) and identity bias; replacing expert panels with AI judges altered 57.2% of individual ratings, reversed 41.7% of pairwise comparisons, and missed roughly 10% of clinically meaningful errors—systematically favoring superficial qualities like “conciseness” and “accuracy” while misprioritizing clinical relevance.
This finding sounds an alarm for *all* scenarios using AI to evaluate AI: **when the judge itself drifts, its scores cannot serve as evidence.**
## Arranging the Eight Studies Along One Continuum
If we arrange this month’s evidence along a task spectrum, it looks roughly like this:
Document summarization—passed rigorous non-inferiority testing; physicians willingly sign; Information extraction—>97%, closed-domain, reliable; Diagnostic reasoning—ceiling ~64%, unsafe recommendations 7%–24%, unsuitable for autonomous use; Adversarial robustness—refusal rates 2%–30%, nearing collapse; Evaluation methodology—AI judges exhibit >50% systematic bias.
This spectrum reveals that asking “Is medical AI ready?” is the wrong question. The right question is: **Which *types* of tasks is it ready for?** Doing documentation doesn’t imply readiness for diagnosis; answering exam questions correctly doesn’t guarantee sound evidence-based reasoning; outputs looking polished don’t mean they withstand deliberate manipulation. These three capabilities—document drafting, diagnostic inference, and adversarial resilience—proved statistically independent in November’s data, offering no mutual guarantees.
For the general public, three concrete takeaways emerge. First, if your hospital begins using AI to draft clinical documents, this is likely an improvement—not a threat. Documentation quality problems are larger than most imagine, and AI already performs better than the human average on this specific layer—while signatures and accountability remain firmly with physicians. Second, if a product claims to offer “AI diagnosis,” remain cautious: even the best models achieve only ~64% accuracy on real patient charts—and when they err, their tone is indistinguishable from when they’re correct. Third, whenever reading news about an “AI healthcare breakthrough,” ask first: Was the test on documentation, information extraction, diagnosis, or adversarial robustness? The distance between these four terms is far greater than most reporting acknowledges.
Finally, one question worth debating: If AI-generated discharge summaries systematically surpass human-written ones in completeness, clarity, and safety, how long can “physician signature” continue to serve unchallenged as the gold standard of quality? Does the signature measure accountability—or competence? And when those two begin to diverge, perhaps the rules governing clinical documentation need revision far beyond just swapping writing tools.
Source: JAMA Network Open, Clinician Acceptance of Large Language Model–Generated Discharge Summaries (November 17, 2026)
https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2856756