How the FDA cleared a patient-facing LLM: the shape of UpDoc's 510(k)

The FDA has cleared a piece of software that talks to patients about their diabetes medication. It is a large language model, it speaks in natural language, and it is the first of its kind to make it through the agency's medical-device pathway. I want to walk through what actually got cleared, because the framing in most coverage ("FDA approves AI doctor") obscures the more interesting thing: the regulatory pathway itself shaped what UpDoc is allowed to be.

The device is called UpDoc V1.0. It received a 510(k) clearance dated 23 December 2025, and the company announced it in June 2026. The cleared indication is narrow: insulin and medication management for adults with type 2 diabetes, operating inside a treatment plan a clinician has already defined.¹ I read the McGuireWoods regulatory alert and the FDA database summaries, and the shape that emerges is not "an AI that practises medicine". It is something more constrained, and more interesting.

What 510(k) actually is

The pathway determines the product. Medical devices reach the US market through two main FDA routes. Premarket approval (PMA) is the heavy one: you submit independent evidence that your device is safe and effective, usually including clinical trials. 510(k) is lighter. You submit a premarket notification arguing that your device is "substantially equivalent" to a predicate device — something already cleared and on the market. If the FDA agrees the two are equivalent in intended use and technological characteristics, you inherit the predicate's regulatory standing.

That word substantially equivalent is doing an enormous amount of work. It means UpDoc had to point at an existing cleared device and say: we are like that, just with a language model on top. Analyses of the clearance describe the predicate as a drug-dose-calculator-type device — the kind of software that already exists to compute insulin doses inside clinician-set parameters.¹

So the LLM was not cleared as a novel category. It was cleared as a conversational interface layered on top of a bounded calculator. That framing is the whole story.

The system, concretely

Here is how I picture UpDoc from the public descriptions. A clinician sets up a treatment plan for a specific patient: which medications, which dose ranges, which titration rules, which red-flag symptoms should trigger escalation. This plan is the protocol envelope. Everything the LLM is allowed to say to the patient has to sit inside it.

The patient interacts through text or voice. They report a blood-glucose reading, ask what to do about a missed dose, describe how they are feeling. The LLM's job is to receive that input in natural language, map it to something the underlying protocol engine understands, and return an instruction the protocol has authorised — adjust the dose by this much, take it at this time, contact your clinician now.

The LLM is not deciding treatment. The protocol is. The LLM is the translation layer between messy human language and a constrained clinical logic that a clinician has already signed off on. This is what "patient-facing" means in the cleared device: the patient talks to the model; the model does not freelance.

Why this matters more than the indication suggests

The cleared use case is narrow — one chronic disease, one class of medication decisions, one patient population. It is not going to change most people's healthcare next week. Its significance is precedential.

The 510(k) pathway is recursive. Today's cleared device becomes tomorrow's predicate. When the next company wants to clear a patient-facing LLM for, say, hypertension medication management, they can argue substantial equivalence to UpDoc. The scope will widen one narrow indication at a time, each new clearance inheriting and slightly extending the last.

This is how the medical-device landscape has always evolved. What is new is that the predicate chain now includes a generative model talking directly to patients.

Human-in-the-loop is a spec, not a promise

The phrase "human-in-the-loop" appears throughout the coverage, usually as reassurance. I want to be precise about what it means here, because the phrase is often used loosely.

In UpDoc's case, the human-in-the-loop is the clinician who authored the treatment plan before the patient ever spoke to the model. They set the envelope. They are not sitting on the other end of the chat at 2am when the patient asks about a missed dose — the model handles that inside the protocol. If the patient's input pushes against the envelope (a reading outside expected range, a symptom that suggests escalation), the system routes to the clinician.

The loop is asynchronous and gated. The clinician sets the rules and receives exceptions. They do not review every message. That is not a criticism — it is how the system is designed to work, and it is how it satisfies the regulatory pathway. But "human-in-the-loop" here means something specific: a human defined the constraints, and a human handles the exceptions. It does not mean a human reviews every model output in real time. If you were imagining the second thing, adjust.

Where FDA clearance stops

One thing worth understanding: FDA clearance covers whether the device can be marketed. It does not resolve everything else.

State-level laws are moving in parallel and sometimes in tension. Maine's HB 2082 and California's AI companion law impose disclosure and consent requirements on AI systems that interact with people, and those apply regardless of federal clearance.² A nationally deployed patient-facing LLM has to satisfy the FDA plus a patchwork of state rules on disclosure, practice of medicine, and data handling.

There is also the consent question. A Nature Medicine editorial published the same week as the announcement argued that data-rights frameworks have not kept up with what LLM-mediated care actually does with patient data — every message a patient sends is training-adjacent material flowing through a model.³ Clearance does not settle that. It is a separate, unresolved layer.

And there is liability. If the model returns an instruction the protocol authorised but the outcome is harmful, the chain runs through the clinician who set the plan, the vendor who built the model, and the health system that deployed it. 510(k) clearance is not a shield for any of them. It is a permission to sell.

The one thing worth remembering

The FDA did not clear an AI doctor. It cleared a language interface bolted onto a bounded clinical calculator, for one narrow indication, with a clinician-authored protocol as the actual decision-maker. That is the object. Everything downstream, the precedent, the predicate chain, the state-law friction, the consent debate, flows from getting that object right in your head.

If you find yourself in a conversation where someone says "the FDA approved an AI doctor", you now have a more accurate sentence to offer them.

Footnotes and links

Further reading

• FDA 510(k) database entry K253281 (UpDoc V1.0) — for the primary regulatory record and cleared indications for use.
• FDA guidance on Software as a Medical Device (SaMD) — for the framework the clearance sits inside.
• FDA draft guidance on AI/ML-enabled device software functions — for how the agency thinks about lifecycle management of learning systems.

¶