UpDoc’s 510(k) cleared “concierge doctor” leaves a liability hole

TL;DR

UpDoc got what it says is the first FDA clearance for a patient-facing clinical LLM, via 510(k), for a concierge doctor supporting insulin and medication management. I argue this system does a clinician's work through a pathway built for fixed, testable devices. STAT asked last week whether the LLM is the interface or the decision-maker, and the clearance does not answer it. The filing is done, but who carries liability when the model is wrong has not been tested.

UpDoc says it has the first FDA clearance for a patient-facing clinical LLM. It got there through 510(k), the pathway that clears a product by finding it substantially equivalent to a device that already exists.

UpDoc's system supports insulin and medication management for adults with type 2 diabetes. Patients use it between visits. The company calls it a concierge doctor.

That is the regulatory crux. A system that supports medication management, and that the company markets as a concierge doctor, reached the market through a pathway built around substantial equivalence.

The direction I flagged in March

I flagged the direction in March, when the FDA granted Breakthrough Device Designation to a GenAI chatbot for post-surgical patients. That designation was the on-ramp. This is the clearance.

The device framework was built for instruments with fixed, testable behavior. A patient-facing LLM giving medication guidance is doing the work a clinician does. Same act, same consequence when it is wrong. Utah has already let AI renew prescriptions without a physician in the loop. The role keeps expanding while the regulatory category stays put.

STAT put the open question plainly last week: is the LLM the interface, or the decision-maker? The clearance does not answer it. When a cleared concierge doctor gets an insulin dose wrong, the accountability chain runs back through a device authorization that never contemplated the model making the call.

The filing is done. Who carries the liability when the model is wrong has not been tested. That is the gap pharma, providers, and platform builders are walking into in 2026.