The FDA Has Authorized Over 1,250 AI Medical Devices. In 2026 It Loosened the Rules for Some and Refused for Others.

How much is already in use

This is not a hypothetical technology. The FDA's public database listed more than 1,250 AI-enabled medical devices authorized for marketing as of July 2025, up from roughly 950 a year earlier. By early 2026 the agency had cleared or approved over a thousand AI and machine learning devices by another count, with the great majority in radiology and imaging.

Despite the attention generative AI receives, most authorized devices rely on predictive models rather than generative ones. Researchers are testing generative applications, including systems that produce synthetic imaging for training or integrate imaging with clinical text, but those remain largely experimental rather than deployed.

The pathway most of them take

Roughly 95 to 97 percent of AI and machine learning devices reached the market through the 510(k) pathway rather than through De Novo classification or premarket approval. The 510(k) route rests on demonstrating substantial equivalence to a device already on the market, which is a meaningfully different standard from independently proving benefit.

Supporters of the current system argue the volume shows a functioning process with no pattern of systemic failure. Critics argue that equivalence-based clearance is a weak bar for a technology whose behaviour depends on training data the predicate device never used, and that stacking new approvals on old predicates compounds the problem over time.

January 2026: the FDA loosened one set of rules

On January 6, 2026, the FDA issued revised guidance expanding enforcement discretion for clinical decision support software. Where a tool provides a single clinically appropriate recommendation and a health care provider can independently review the basis for it, including the underlying logic and data inputs, it may fall outside device regulation. This applies to AI, and to certain generative AI features.

The agency also broadened its general wellness policy so that more non-invasive consumer wearables reporting physiologic metrics, including blood pressure, oxygen saturation or glucose-related signals, can fall under enforcement discretion if paired with non-diagnostic notifications. That is a material expansion from the 2019 version of the policy.

Then it refused to loosen another

The same year, the FDA rejected a proposal to ease oversight of several categories of radiology and diagnostic AI software, including computer-aided detection systems that identify abnormalities in medical imaging. The proposal was published in the Federal Register and drew 47 comments before the February 27, 2026 deadline.

Read together, the two decisions describe a line rather than a direction. Software that advises a clinician who retains judgment gets lighter treatment. Software that substitutes for clinical judgment, or influences time-critical care, stays inside the device framework. Whether clinicians actually retain independent judgment under time pressure is the part nobody can regulate directly.

Performance drift is the unresolved problem

In 2026 the FDA asked for public comment on how to measure AI device performance in the real world, with specific attention to performance drift. An AI system's accuracy can shift after deployment because clinical practice changes, patient populations change, or the data being fed to it changes. The device approved is not necessarily the device running two years later.

This is the structural mismatch both sides acknowledge. Traditional review evaluates a fixed product at a fixed moment. Supporters of stronger oversight argue this means continuous post-market monitoring is essential. Industry argues it means the pre-market gate is the wrong place to concentrate effort. They agree on the problem and disagree entirely about the remedy.

Where the debate actually splits

The left generally argues that software influencing diagnosis can cause harm at scale, that performance drifts after deployment, that unrepresentative training data produces tools that work worse for some groups, and that liability remains unsettled enough that a harmed patient may have no clear defendant.

The right generally argues that frameworks built for static devices fit continuously updating software poorly, that approval delay has a patient cost too, that over a thousand authorized devices without systemic failure suggests the process works, and that heavy compliance costs favour large incumbents over smaller developers.