AI Medical Diagnosis Gains in Imaging as FDA Loosens Oversight

Narrow imaging tools are improving diagnosis. General chatbots are not there yet.

Whether AI can make medical diagnosis more accurate has no single answer in 2026. The research this year splits into two groups. Purpose-built imaging tools now have solid evidence behind them, some of it from randomized trials. General-purpose language models score well in some tests and fail badly in others. The gap between those two groups matters more than any one benchmark.

The regulatory approach in the United States is moving faster than the evidence. Federal oversight of AI decision-support software has been loosened. As a result, hospitals now carry much of the work of deciding which tools to trust.

The scale of FDA-authorized AI devices

The FDA says it had authorized more than 1,600 AI-enabled medical devices for U.S. marketing as of September 2026. Its examples include skin-cancer imaging systems, diabetic retinopathy detectors and algorithms that automate insulin dosing.31 Published counts vary because they were taken at different times and the FDA's list is updated only periodically. One analysis counted 1,451 devices in March 2026, up from 1,250 a year earlier. Radiology accounted for 1,104 of those, far more than cardiovascular at 141 and neurology at 67.32 The agency itself describes the list as incomplete. One review found 92 entries dated between January 5 and March 30, 2026, and warned that this was not a full count for the year.35

Two details from that tally matter for the accuracy question. First, nearly all of these devices reached the market through the 510(k) pathway. That route requires showing a device is substantially equivalent to one already on the market. It accounted for 1,396 authorizations, compared with 37 De Novo grants and 18 premarket approvals.32 Second, as of March 2026 the FDA had not authorized any device built on generative AI or large language models.32 So most of what hospitals use as "FDA-cleared AI" is narrow, task-specific software reviewed under the lighter pathway. It is not the chatbot-style technology behind most headlines.

Where the accuracy gains are real

The strongest evidence comes from screening and detection. A narrative review published this week in Frontiers in Digital Health looked at a population-based mammography trial with more than 105,000 women. In the full population, AI-supported reading found 6.4 cancers per 1,000 women screened, against 5.0 with standard reading. Radiologist workload fell by about 44%.19 In a multicenter colonoscopy trial, computer-aided detection raised the adenoma detection rate from 40.4% to 54.8%.19 Both are Level 1 randomized evidence. Most AI marketing claims have nothing comparable behind them.

Results from vendors point the same way, with the usual caution about company-run studies. In January 2026, Aidoc received clearance for what coverage calls the first foundation-model-powered clinical AI device. It is a body-CT triage tool covering 14 conditions, with a mean sensitivity of 97% and a mean specificity of 98% in its pivotal study.9 Aidoc says it runs in close to 2,000 hospitals and processes about 60 million patient cases a year.9 Outside radiology, a scoping review found that AI-supported digital microscopy matched reference standards for tests such as malaria detection and complete blood counts. It was more sensitive than manual microscopy in six of seven head-to-head comparisons.20

Generalist models: mixed results

Large language models show a much less consistent picture. A Harvard Medical School and Beth Israel Deaconess study published in Science in May looked at 76 real emergency-room cases. OpenAI's o1 model gave an exact or very close diagnosis at triage 67% of the time. The two attending physicians it was compared with managed 55% and 50%.17 The authors did not claim the model was ready for the emergency room. They said the results showed an urgent need for prospective trials in real patient care.17

A month earlier, a Mass General Brigham study in JAMA Network Open found something different. It tested models step by step as a case unfolded. Every model failed to produce an appropriate differential diagnosis more than 80% of the time, even though final-diagnosis accuracy ranged from about 60% to over 90%.11 The two studies seem to contradict each other on early-stage reasoning, when little information is available. Harvard found o1 strongest at the initial triage point.17 Mass General Brigham found models weakest at the open-ended start of a case.11

The likely explanation is that the two studies measured different things. One scored the single best diagnosis. The other scored the quality of the full list of possibilities a clinician needs to work through. Neither result should be taken as settled.

Broader data supports caution. A 2025 meta-analysis of 83 studies put generative AI's overall diagnostic accuracy at 52.1%, about 15.8 percentage points below expert specialists.13 A 2026 Oxford study in Nature Medicine had nearly 1,300 members of the public work through medical scenarios. Those using chatbots did no better than those using a Google search or their own judgment.13 A Stanford and Harvard safety benchmark found that top models still gave severely harmful recommendations in up to 22.2% of cases.14

The weak point is the doctor-plus-AI combination

The most important finding this year is not about how accurate AI is on its own. It is about what happens when clinicians use it. In the only randomized trial of open-ended diagnostic reasoning, physicians with GPT-4 scored 76%, compared with 74% for those without it. That is not a meaningful improvement, even though GPT-4 working alone scored 92%.19

The review also found evidence of harm. In one mammography reader experiment, wrong AI suggestions cut correct assessments sharply. Among very experienced readers, accuracy fell from 82.3% to 45.5%.19 Physicians who had completed 20 hours of AI-literacy training were still misled by incorrect model advice.19 In a pre-post study, endoscopists' unassisted adenoma detection fell from 28.4% to 22.4% after they began using AI routinely. That suggests their own skills declined.19 A commercial sepsis model in clinical use performed poorly on external validation, with an area under the curve of 0.63.19

Performance often drops after deployment for simpler reasons too. One analysis reports declines of 15% to 30% from benchmark results once tools meet different patient populations and hospital systems.18 Practitioner accounts describe stroke-alert tools being switched off within weeks when false alarms swamped staff.4 One assessment also warns that leading vision-language models underdiagnose some groups, including Black female patients, at higher rates than board-certified radiologists.8

The conclusion is that an accurate tool does not automatically improve diagnosis. How the tool is put into everyday practice largely decides whether patients benefit.

Regulation is moving the other way

That makes the direction of U.S. policy especially important. On January 6, 2026, the FDA loosened its clinical decision support guidance. Software that gives a single recommendation can now reach the market without FDA review, as long as it meets the agency's other criteria.24 Coverage describes the change as part of the administration's push to deregulate AI. It also notes that it could let unregulated generative AI tools into clinical workflows.24 Sources disagree slightly on the date: some describe the guidance as issued on January 6 and reissued on January 29.2625

The change has clear limits. Software that analyzes medical images to make diagnostic recommendations is still a regulated device, and so are tools that interpret continuous physiological signals.26 At a March town hall, the FDA said tools built on large language models can qualify as non-device decision support. The condition is that a clinician must be able to independently review the basis for each recommendation.27 The agency also withdrew its separate guidance on clinical evaluation of software as a medical device.29 In August, it published a non-binding discussion paper on generative-AI-enabled devices, with comments due October 19.25

The practical result, as one analysis puts it, is that the job of vetting tools moves from the regulator to the hospitals buying them.26 Professional bodies are starting to fill that gap. In May, the American College of Radiology and the Society for Imaging Informatics in Medicine approved the first practice parameter for imaging AI. It asks facilities to set up AI governance, keep an inventory of their tools, test locally before go-live, and monitor performance on an ongoing basis.1 States are adding their own rules, mainly requiring disclosure and human review. Texas, for example, requires disclosure to patients and practitioner review of AI-created diagnostic records.25

Analysis: limited gains so far, with real risks

Some hospital leaders are getting ahead of the evidence. NYC Health + Hospitals CEO Mitchell Katz said in March that his system could replace "a great deal of radiologists" with AI if regulations allowed.10 No radiology algorithm currently has FDA or CE authorization to produce final reports on its own without review.1

Overall, AI is already improving diagnostic accuracy in narrow, well-defined tasks: screening mammography, polyp detection and imaging triage. In those areas, randomized trials and regulatory review back the claims. Generalist models are improving quickly on benchmarks, but they have not yet shown that they make clinicians better in real care. The research on automation bias suggests that poorly integrated tools can make clinicians worse.

The FDA's lighter approach to decision support applies to exactly the category of tools with the least evidence. That makes hospitals' own testing and monitoring the main safeguard. The open question for the rest of the decade is less about whether AI can diagnose well, and more about whether hospitals can reliably check that it does.