AI for Medical Use: How FDA-Cleared Devices Are Authorized and What That Actually Means

When someone searches for "AI for medical" use, they are usually asking one of three very different questions: Is this tool actually cleared to be used on patients? What does the FDA authorization actually permit it to do? And how does the clearance pathway affect what the manufacturer can change about the algorithm after it ships?

These questions have specific, traceable answers — but they get muddled because the phrase "FDA-cleared AI" circulates in vendor materials, press releases, and procurement conversations without much precision. A device can be cleared for a narrow screening task and marketed as a broad diagnostic tool. A 510(k) clearance and a De Novo authorization carry different regulatory histories. And a Predetermined Change Control Plan (PCCP) changes what a manufacturer is allowed to do to an algorithm post-market.

This record explains the authorization landscape for AI in medicine — how devices get cleared, what the major pathways mean in practice, which clinical specialties hold the most cleared devices, and what limits apply regardless of clearance status.

The Three Pathways to FDA Authorization for AI Medical Devices

FDA regulates AI-enabled medical devices as Software as a Medical Device (SaMD) under three primary authorization pathways. Each pathway has a different evidentiary bar, a different regulatory history, and different implications for how the device can be updated.

The 510(k) pathway is the workhorse of the registry. It requires a manufacturer to demonstrate that a new device is substantially equivalent to a legally marketed predicate — meaning similar intended use and similar technological characteristics, or different characteristics that don't raise new safety questions. For AI tools, this often means pointing to an earlier cleared CAD (computer-aided detection) system or an established decision-support tool in the same clinical domain.

De Novo is the pathway that gets used when a device is genuinely novel — no predicate exists — but the risk level is low to moderate. Once a De Novo is granted, it establishes a new device type and can itself serve as a predicate for future 510(k) submissions. Several early AI diagnostic tools in ophthalmology and radiology came through De Novo, and subsequent competitors then used those decisions as predicates.

PMA applies to the highest-risk devices. For AI, this is rare — most AI tools are positioned as decision support rather than autonomous diagnostic systems, which keeps them out of the highest risk tier. But as AI systems move toward more autonomous functions with fewer human-in-the-loop safeguards, PMA becomes more relevant.

What "Cleared" Actually Permits — and What It Doesn't

FDA clearance authorizes a device for a specific intended use. That intended use statement is the operative document — it defines the clinical task, the patient population, the care setting, and the level of clinician oversight required. A device cleared to "assist radiologists in identifying potential nodules on chest CT" is not cleared for autonomous nodule diagnosis, and it is not cleared for use without a radiologist in the loop.

This distinction matters practically. A hospital procurement team verifying whether a tool is "FDA-cleared" is asking a necessary but not sufficient question. The follow-up questions — cleared for what, in what population, with what human oversight requirement — are where the operational decisions live.

Intended Use vs. Indications for Use

Two terms in FDA submissions that are sometimes conflated: intended use describes the general purpose of the device (what it's meant to do), while indications for use is more specific — it defines the disease or condition the device is intended to diagnose, treat, or monitor, and the target patient population. For AI devices, the indications for use statement is the more operationally relevant document for clinical deployment decisions.

The Specialty Distribution of Cleared AI Devices

AI device authorizations are not evenly distributed across medicine. Radiology and medical imaging dominate the registry by a significant margin — estimates based on FDA's public database consistently place imaging-related AI at roughly 75% of all cleared AI/ML-enabled devices. This reflects both the data availability (imaging produces structured, annotatable datasets at scale) and the regulatory history (CAD for mammography was one of the earliest cleared AI applications, establishing precedent and predicates for subsequent submissions).

Predetermined Change Control Plans: What They Allow Post-Market

One of the most practically significant developments in AI device regulation is the Predetermined Change Control Plan (PCCP). Traditional device regulation assumed a relatively static product — you clear a device, you ship it, significant changes require a new submission. AI algorithms can improve (or drift) over time with new training data, which creates a mismatch with that model.

A PCCP allows a manufacturer to pre-specify the types of algorithm modifications they anticipate making post-market — and to get FDA agreement in advance that those specific modifications, if executed within defined parameters, won't require a new premarket submission. The PCCP must describe the modification protocol, the performance goals the modified algorithm must meet, and the methodology for verifying those goals.

PCCPs became formally available through FDA guidance finalized in 2023 and have been incorporated into a growing number of AI device submissions since. Devices with an active PCCP are identifiable in the FDA's submission database, though the PCCP document itself may require a separate request to access in full.

Generative AI and the Current Authorization Gap

As of Q2 2026, no large language model or generative AI system has received FDA authorization as a medical device for clinical diagnostic or treatment decision tasks. This is not a minor detail — it is the defining regulatory boundary for a large category of tools that are actively being deployed in clinical settings.

Several generative AI tools operate in healthcare settings under regulatory frameworks that don't require FDA clearance — primarily ambient documentation tools that are positioned as clinical workflow software rather than medical devices making clinical determinations. Whether that positioning is accurate for all use cases is an active area of FDA scrutiny.

How to Verify an AI Device's FDA Status

FDA maintains two primary public databases relevant to AI device verification. Neither is perfectly organized for the task, but both are authoritative.

• The FDA 510(k) database allows lookup by submission number, device name, or applicant. The full decision summary PDF for each cleared device contains the intended use statement, the predicate device(s) cited, and the performance data reviewed.
• The FDA AI/ML-enabled device list is a curated spreadsheet of devices the FDA has identified as AI/ML-enabled. It is updated periodically but may lag behind the main 510(k) database by weeks or months.
• For De Novo authorizations, the De Novo database is the authoritative source. De Novo decisions include an order letter that specifies the device type created and the special controls established.
• Submission numbers follow a predictable format: 510(k) numbers begin with K, De Novo numbers begin with DEN, and PMA numbers begin with P. A vendor claiming clearance should be able to provide the exact submission number — if they can't, that is a red flag.

What Clearance Status Doesn't Tell You

FDA clearance answers a narrow question: did this device meet the regulatory standard for its stated intended use at the time of submission? It does not answer questions that matter equally for deployment decisions.

• Performance on your patient population: Training data demographics and the patient population in your institution may differ substantially. A device cleared on predominantly one demographic group may perform differently on others — and the clearance record may not surface this clearly.
• Real-world workflow integration: Clearance says nothing about PACS compatibility, EHR integration, latency under clinical load, or how the tool behaves when input data quality is lower than the training set.
• Post-market performance drift: Algorithms can degrade over time as clinical practice patterns, imaging hardware, or patient populations change. Post-market surveillance requirements exist but are not uniformly rigorous across device types.
• Clinical outcome evidence: A device can be cleared based on technical performance metrics (sensitivity, specificity on a test set) without prospective evidence that using the device improves patient outcomes. Many cleared devices lack RCT-level outcome data.

Reading the FDA Submission Record: A Practical Guide

The 510(k) decision summary is a public document that most procurement teams never read. It typically runs 10–40 pages and contains more operationally useful information than any vendor data sheet. Key sections to locate:

1. Indications for Use — the operative statement of what the device is cleared to do, in what patient population, in what clinical setting.
2. Device Description — what the algorithm does technically, what inputs it requires, and what outputs it produces.
3. Substantial Equivalence Discussion — which predicate device was cited and why the new device is considered equivalent. This reveals the regulatory lineage of the tool.
4. Performance Testing — the dataset used for validation, sample sizes, demographic breakdown (if disclosed), and the performance metrics reported. Sensitivity and specificity figures here are test-set results, not real-world performance guarantees.
5. Limitations — often buried but sometimes explicit about population restrictions, hardware requirements, or conditions under which the device should not be used.

The Relationship Between Clearance Pathway and Risk Level

FDA device classification runs from Class I (lowest risk) through Class III (highest risk). Most AI/ML-enabled medical devices cleared to date are Class II — moderate risk — which is why 510(k) and De Novo dominate the registry. Class III devices require PMA, and very few AI tools have pursued that pathway.

The risk classification reflects the intended use and the consequences of device failure, not the technical complexity of the algorithm. An AI tool that flags potential findings for radiologist review is lower risk than one that makes autonomous treatment recommendations with no clinician review step. That distinction drives both the pathway chosen and the post-market obligations that follow.

One consequence: the 510(k) pathway's substantial equivalence standard means FDA is not independently testing device performance — it is reviewing whether the manufacturer's own testing demonstrates equivalence to a predicate. The rigor of that review depends on the quality of the submission and the FDA reviewer's scrutiny. It is not equivalent to an independent clinical trial.

Scope and Limitations of This Registry

The FDA's own AI/ML device list and the 510(k) database are the authoritative sources. This registry organizes and contextualizes that information for readers whose primary task is verification and comparison — not browsing vendor materials or navigating FDA's database interface directly.