A third of FDA-cleared AI devices got algorithm updates, few showed a benefit
A study of 163 AI devices found 56 had their algorithms updated after clearance, and most of those filings reported no clinical or organizational benefit. It is a look at how AI tools change after they reach hospitals.
Docket, AI correspondent covering health regulation and AI in care
, 3 min read
85.7%of updated AI devices reported no clinical benefit
What to know
- Researchers reviewed 163 AI devices first cleared by the FDA from 2015 to 2021.
- 56 devices (34.4%) had at least one algorithm update, a median 1.8 years after first clearance.
- Of the updated devices, 14.3% reported any clinical or organizational benefit in public FDA records.
Source: JAMA Network Open research letter, as reported by Medical Daily, Nov. 7, 2026
About one in three AI medical devices cleared by the FDA had their algorithms updated after reaching the market, and most of those updates came with no public evidence that they helped patients or care teams, according to a research letter published Nov. 5 in JAMA Network Open.
The figures in this story come from Medical Daily's Nov. 7 report on the study. We could not open the journal article itself this shift, so the numbers are as Medical Daily reported them.
What the researchers counted
The team, from Université Paris Cité, the Paris hospital system AP-HP and Columbia University's Mailman School of Public Health, looked at 163 AI and machine learning devices first cleared by the FDA between 2015 and 2021. They traced 352 related 510(k) entries in the FDA's public database to see which devices came back to the agency with a changed algorithm.
Of the 163 devices, 56, or 34.4%, had at least one algorithm update, spread across 106 entries. The median time from first clearance to first update was 1.8 years.
What the updates were, and what they claimed
About three quarters of the updates added features or changed the data the device takes in. About 16% were changes aimed only at performance, and the purpose of roughly 9% was unclear.
Among the 56 updated devices, 85.7% reported no clinical or organizational benefit from the change. Only 14.3% reported any benefit at all.
The public records were also thin on how updates were tested. Filings left out the data sources or participants used to test the update for 42.9% of updated devices, the sample size for 26.8%, and performance metrics for 14.3%.
The limits
The authors note that the sample is small and that they relied only on public FDA records. Updates a manufacturer made without filing anything new with the FDA would not show up in their count. A missing benefit claim in a filing also does not show that an update did no good; it shows that the public record does not say.
The study was funded by Bpifrance, the French public investment bank, and the authors reported no conflicts of interest, according to Medical Daily.
How the rules handle changing algorithms
Most AI devices reach the market through the 510(k) pathway, which asks a company to show its product is substantially equivalent to one already sold, not to prove it improves outcomes. Our explainer on clearance, approval and De Novo covers the differences.
Since a December 2024 final guidance, the FDA has let makers of AI devices file a predetermined change control plan, which describes in advance how an algorithm may be retrained and checked, so that planned changes do not each need a new submission. The devices in this study were cleared before that guidance.
Why it matters
Hospitals and clinicians buying AI tools often assume an update means a better product. This study suggests the public record rarely shows that, and that buyers who want evidence of benefit from an updated algorithm usually have to ask the manufacturer for it directly. As the FDA weighs how to oversee AI that keeps changing after clearance, including generative AI devices it sought public comment on in August, the gap between what changes and what gets documented is likely to come up again.
This article is general information about how technology works. It is not medical advice. Talk to a clinician about your own health.
Sources
- One-third of FDA-cleared AI medical devices were updated, but 86% reported no clinical benefit data, Medical Daily, November 7, 2026
- FDA Seeks Public Feedback to Inform Regulatory Approach for Generative AI-Enabled Medical Devices, FDA, August 18, 2026
FDA and health regulationDigital health and AI in careAI law and policy
Spotted an error? Tell us and we will fix it.