Why did doctors check only about three-quarters of AI-generated medical transcripts, and what does that say about human oversight?

This explores why physicians placed in a review role over AI-generated clinical text checked only part of it. The collection does not contain the 73% figure or a study that explains it, so it cannot tell you why those physicians skipped about a quarter of the transcripts. What it does hold is a cluster of findings that point to a likely pattern: human review of AI output usually falls off, and how the review fits into the workflow matters more than whether people intend to review.

The closest parallel comes from outside medicine. When writers were given AI-drafted paragraphs, they edited them only 23% of the time, and their edits left the text about 96% unchanged Do writers actually edit AI-generated text before publishing?. Your 73% review rate looks diligent by comparison. The lesson is that a human 'in the loop' is often a light touch, and an unchecked AI draft usually goes out as written.

The clinical studies in the collection suggest two reasons review might thin out. The first is that the output looks fine. In blinded ratings, clinicians could not tell GPT-4's advice from expert advice (they identified the source correctly only 45% of the time) and rated it equally sound and more emotionally empathetic Can clinicians tell GPT-4 advice apart from expert advice?. When AMIE took histories from 100 real urgent-care patients, human supervisors never had to step in Can conversational AI safely take patient histories without supervision?. A reviewer who keeps finding nothing to fix has less reason to keep looking. The second is that what reviewers say about AI and what they do with it can differ. Radiologists rated advice lower when it was labeled as AI, yet their diagnostic accuracy depended only on whether the advice was correct Does labeling advice as AI change how clinicians use it?. Doubting AI is not the same as checking it.

The most useful contrast is the Nairobi primary-care study. Across nearly 40,000 visits, an LLM safety net cut diagnostic errors by 16% and treatment errors by 13%. The authors credit the gains to an asynchronous design that fit the clinicians' workflow and to active rollout, not to the model alone Can AI safety nets reduce errors in live clinical practice?. This reverses the usual setup. Instead of asking busy doctors to review the AI, the AI reviewed the doctors at a moment that suited them. That suggests a missing quarter of reviews is less a story about careless physicians and more a sign that the review step was not designed around their actual workday.

The broader lesson is that 'a physician will review it' gets treated as a safety guarantee, but every study here that measures oversight finds it is partial, uneven, and shaped by design. The AI may also be strongest exactly where review is weakest: AMIE beat primary care physicians mainly at reasoning from the information gathered, not at gathering it Can an AI system diagnose better than primary care doctors?. A busy reviewer is least likely to catch errors in that kind of reasoning.