AI diagnostic software reaches patients as a regulated medical device, and the U.S. Food and Drug Administration (FDA) publishes a list of AI-enabled devices that held 1,614 rows, about three-quarters of them in radiology, when counted for this article on October 24, 2026 (the FDA gives no total itself). The best randomized evidence so far comes from breast cancer screening: a Swedish trial of 105,934 women found that AI-supported screening kept the interval cancer rate non-inferior to standard double reading (1.55 versus 1.76 per 1,000) and raised sensitivity to 80.5 percent from 73.8 percent. An FDA listing shows that a device was authorized, not how well it works in a given clinic.
Key takeaways
- A row count of the FDA list on October 24, 2026 gave 1,614 entries, and 1,230 of them were in the radiology panel; the FDA does not publish these totals.
- The FDA says the list is not a comprehensive resource of AI-enabled medical devices.
- In the MASAI mammography trial, AI-supported screening matched standard double reading on interval cancers and specificity (98.5 percent in both groups) while reducing screen-reading workload.
- In a 27-reader experiment, correct mammogram ratings fell sharply when a purported AI suggested the wrong category, even for very experienced radiologists (82.3 to 45.5 percent).
- A trial in one national screening program does not settle how an AI tool performs in other populations, equipment, or workflows.
What the FDA list shows and what it leaves out
The FDA publishes a List of Artificial Intelligence-Enabled Medical Devices. Counting the table rows on October 24, 2026 gives 1,614 entries, of which 1,230 sit in the radiology panel, 154 in cardiovascular, and 73 in neurology. The most recent date of final decision on the page is June 29, 2026. The agency states that the list is not a comprehensive resource: devices were identified mainly through AI-related terms in their marketing authorization summaries or classification, and devices whose summaries are not yet published are added later.
| Year of final decision | Entries on the list |
|---|---|
| 2023 | 226 |
| 2024 | 235 |
| 2025 | 335 |
| 2026 (through June 29) | 181 |
The yearly counts are the author's tally of the page's dates and should be read as approximate, since the list is revised from time to time. They measure authorizations. They do not measure how many patients were scanned with these tools or whether outcomes improved.
How a cleared model can change afterward
Software that learns from data can be updated, which raises the question of what happens to its clearance. The FDA's final guidance on predetermined change control plans for AI-enabled device software functions (the page is dated August 2025) addresses this. It gives recommendations for manufacturers to describe planned modifications in advance, and it applies to AI-enabled devices reviewed through the 510(k), De Novo, and premarket approval pathways. Health Canada has a comparable pre-market guidance for machine learning medical devices, covered on the site's article on AI in Canadian health care.
Mammography screening: the best randomized evidence
The Mammography Screening with Artificial Intelligence (MASAI) trial randomly assigned 105,934 women attending routine screening in Sweden, between April 12, 2021, and December 7, 2022, to AI-supported screening or standard double reading without AI. In the AI arm the software triaged examinations to single or double reading by radiologists and supported detection. Nineteen women were excluded from the analysis. The results were published in The Lancet in 2026 by Gommers and colleagues, and the paper reports the following.
- Interval cancers, which are cancers diagnosed between screening rounds, occurred at 1.55 per 1,000 participants with AI and 1.76 per 1,000 with standard reading, a non-inferior ratio of 0.88 (95 percent confidence interval 0.65 to 1.18).
- Sensitivity was 80.5 percent with AI and 73.8 percent without; specificity was 98.5 percent in both groups.
- The AI arm had fewer invasive interval cancers, 75 versus 89, but the authors present this as a descriptive finding.
- Screen-reading workload fell, according to the authors' interpretation.
The authors conclude that the approach may be considered for clinical implementation. The trial ran within one national screening setting over one screening round, so it does not show results in other countries, with other mammography equipment, or with different AI software.
The human side: automation bias
Accuracy depends on how readers respond to the output. In a prospective experiment published in Radiology in 2023, Dratsch and colleagues had 27 radiologists rate 50 mammograms with the help of a purported AI system. For 12 of the 40 test cases the suggested category was wrong. Inexperienced readers rated 79.7 percent of cases correctly when the suggestion was right and 19.8 percent when it was wrong. The pairs were 81.3 and 24.8 percent for moderately experienced readers and 82.3 and 45.5 percent for very experienced readers. The suggestions were presented as AI output rather than coming from a validated live system, and the setting was an experiment, so the figures show a risk and not a measured clinical error rate. The World Health Organization named automation bias among the risks in its January 18, 2024 guidance on large multi-modal models, along with false, biased, or incomplete outputs.
Questions to ask when reading a diagnostic AI claim
- Was the tool tested in a randomized trial, a retrospective dataset, or a lab experiment?
- Were patients, scanners, and sites similar to the setting where it will be used?
- Did the study measure outcomes such as missed cancers, or only agreement with experts?
- Who funded the work, and do the authors hold a stake in the product?
- Did a clinician remain responsible for the final read?
This article summarizes regulatory records and published research. It is not medical advice, it does not assess any specific product, and it cannot say whether a test or screening approach suits a particular person. Those questions belong with a physician or the clinic that provides the service.
FAQ
Does FDA authorization mean an AI tool diagnoses better than doctors?
No. Authorization is a regulatory decision about a specific device and intended use. It does not rank the tool against clinicians, and the FDA list does not carry performance comparisons.
In the Swedish trial, did AI make the diagnosis alone?
No. The software sorted examinations into single or double reading and supported detection, and radiologists did the reading. The trial tested a workflow, not an unsupervised system.
Why does the list not equal how often AI is used?
The list counts authorized devices. A hospital may own several tools, use them on a fraction of cases, or not switch them on, and none of that appears in the record.
Has the mammography result been repeated elsewhere?
This article did not verify a second randomized trial of the same design. The Swedish authors themselves describe one setting and one screening round, so wider confirmation remains an open question.
Related reading
- Where AI stands in organ transplantation
- AI in Canadian health care
- How Canada regulates AI
- A beginner's path into AI
More in AI in health.
Last verified: October 2026
Sources
- List of Artificial Intelligence-Enabled Medical Devices, U.S. Food and Drug Administration, page content as of October 2026
- Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions, U.S. Food and Drug Administration, 2025
- Gommers J et al., Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading without AI in the MASAI study, The Lancet, 2026
- Dratsch T et al., Automation bias in mammography: the impact of artificial intelligence BI-RADS suggestions on reader performance, Radiology, 2023
- WHO releases AI ethics and governance guidance for large multi-modal models, World Health Organization, January 18, 2024