The FDA just made you the check on AI in your charts.

What the FDA actually did

On January 6, 2026, the FDA issued revised guidance that expanded its enforcement discretion for AI and generative-AI clinical decision support software. In plain terms, the agency said it will not regulate many low-risk AI tools as medical devices.

The reasoning, as legal analysts at Faegre Drinker and Orrick read it, is that these tools are meant to support a clinician's judgment rather than replace it. If a doctor remains in the loop, the FDA is comfortable treating the software as an assistant rather than a regulated device, and it leans instead on professional judgment and post-market accountability to keep things safe.

For a busy physician, the temptation is to read this as the FDA getting out of the way, which means fewer hoops. That reading is half right and dangerous. The hoops did not disappear. They moved onto your desk.

The condition that lands on you

The exemption is not unconditional. It holds only when the clinician can independently review the basis for the recommendation. That single clause is the entire deal.

Read it the way a regulator would. The FDA is willing to step back precisely because it assumes a licensed human is standing between the AI and the patient, able to see why the tool made its recommendation and able to override it. The light-touch regulation and your independent review are not two separate things. The review is the reason the light touch exists.

So when the agency steps back, it does not create a gap. It hands the gap to you. If the AI gets something wrong and you signed it without being able to explain the basis, the regulatory framework that was supposed to catch the error was you, and it was not there. That is the accountability the new guidance leans on.

This is a familiar trade in medicine, even if the technology is new. Whenever a regulator decides a tool is low-risk enough to deregulate, it is making a bet that trained professionals will catch what the lighter oversight no longer does. The bet is reasonable, but only if the professionals actually perform the review the bet assumes. The danger is a quiet drift where the tool feels reliable, the reviews get faster, and one day the independent judgment that justified the whole arrangement has thinned out to a signature. The FDA did not say that is acceptable. It said the opposite, by making your review the condition of the exemption.

What "independently review the basis" means on a Tuesday

This phrase can sound abstract, so make it concrete. To independently review the basis for an AI recommendation means you can answer, in your own clinical reasoning, why the recommendation makes sense for this patient, without simply trusting that the model must be right.

That has a few practical edges:

• You can see enough of the AI's reasoning or inputs to evaluate it, not just its conclusion. A black-box output you cannot interrogate is harder to defend.
• You apply your own clinical knowledge to confirm the recommendation fits the patient in front of you, not a generic case.
• You are willing and able to reject it. A review you would never act on is not a review.

If you can do that, you are exactly the safety net the FDA built its lighter framework around. If you are rubber-stamping outputs you cannot explain, the framework has a hole in it, and your signature is on the hole.

A verify-before-you-sign workflow for any AI output

The good news is that the discipline this requires is the same discipline that already protects you everywhere else AI touches the chart. The companion briefing on how doctors use AI for clinical notes safely lays out the core review habit, and it maps directly onto what the FDA now expects.

A workable routine for any AI recommendation that reaches the record:

• Pause before the signature. The moment of attestation is the moment your accountability attaches, so make it deliberate.
• Ask whether you can explain the basis. If you cannot say why the output is right for this patient, you are not done reviewing.
• Check it against the actual patient, not the typical patient the model assumes.
• Document your call where it matters, so the independent review you performed leaves a trace.

None of this is slow once it is a habit. It is the difference between using AI as a tool and being used by it.

The post-market accountability piece deserves one more line, because it is where the lighter framework has teeth. Enforcement discretion is not a promise the FDA will never act. It is a decision not to gate these tools up front, while reserving the right to respond when something goes wrong in the field. If a pattern of harm shows up, the question asked afterward will be whether the clinician exercised the independent review the framework assumed. A documented habit of real review is the answer you want to have ready. An absence of it is the answer you do not.

Where this sits next to consent and state rules

One boundary, so you do not conflate three different things. This briefing is about FDA oversight and where accountability for AI recommendations now sits. Whether your patient must be told or must agree to AI involvement is a separate question covered in the briefing on when AI scribes need patient consent. The FDA framework is about your review duty. Consent is about the patient's awareness. Keep them in separate boxes and both get easier.

The skill under the oversight

Regulators will keep adjusting how tightly they hold AI in medicine. This year they loosened the grip on low-risk tools. The constant underneath every version of the rules is the same: a licensed human who can look at what the machine produced and independently judge whether it is right.

That skill is not optional anymore. It is the literal condition the FDA attached to the freedom physicians just gained. AI for Physician Notes builds the verify-before-you-sign workflow that makes you a real safety net rather than a rubber stamp, and the two minute course quiz will point you to the right starting place for your practice.