Two headlines landed in the same news cycle, and they should be read together. The Centers for Medicare & Medicaid Services established a new Office of Health Technology Products to oversee AI, interoperability and digital health tools — the first serious regulatory scaffolding around clinical AI. And a federal whistleblower lawsuit alleged a cover-up of AI error rates, cutting straight through the industry's preferred story of seamless clinical automation (MarketScale; The Health AI Brief).
Meanwhile the technology arrives regardless. A flexible on-body patch now detects life-threatening arrhythmias with 99.6% accuracy, processing data on the body rather than in the cloud. That is genuinely extraordinary. It is also the whole problem in miniature: clinical-grade AI is reaching the bedside faster than the machinery for questioning it is being assembled.
The thesis
Here it is plainly. If a tool participates in a clinical decision, its error rate is not a trade secret. It is a clinical fact — and it belongs to the people who will absorb the consequences. We have quietly accepted a standard for software that we would find grotesque in any other part of medicine. No hospital would stock a medication whose failure rate the manufacturer declined to disclose. No surgeon would use a device whose malfunction data was sealed as competitively sensitive. Yet we routinely deploy algorithms into triage, imaging and documentation while treating their performance characteristics as intellectual property.
You would never accept a medication whose failure rate was confidential. Why is an algorithm different?
And notice who pays for the asymmetry. Not the vendor. The nurse — who is told to "use clinical judgment to override the tool" while being denied the one number that would tell her when to. That is not decision support. That is liability transfer.
The honest counterpoint
Now the other side, argued fairly, because this deserves better than a slogan.
Error rates are genuinely slippery. A model's performance is not a single number — it shifts by population, by site, by scanner, by season, by how the data was labeled. A published figure from one health system can be actively misleading in another, and a naive "accuracy" percentage can create false confidence just as easily as secrecy creates false trust. There is also a real risk of over-correcting: if we bury AI vendors in disclosure obligations, the ones who survive will be the largest incumbents, not the best tools — and some of these systems already outperform an exhausted clinician at 3 a.m. Demanding perfect transparency from software while tolerating unmeasured human error is its own kind of double standard.
All of that is true. None of it justifies concealment. The answer to "error rates are context-dependent" is to publish them with their context — the population, the site, the failure modes, the known blind spots — not to publish nothing.
What we'd like to see
Error rates as a condition of deployment, not a courtesy. Disclosed to the clinicians using the tool, in plain language.
Failure modes named, not just accuracy claimed. Tell us who the model gets wrong, not only how often.
Post-deployment monitoring at each site. A model that performed at one hospital is a hypothesis at yours.
Whistleblower protection that actually works — because right now, a lawsuit is how we found out.
Where FHC stands
Our position hasn't moved: AI can carry the documentation; the nurse keeps the judgment. But judgment requires information. Ask your vendors for the error rate. Ask what happens when it's wrong. And if the answer is "that's proprietary," you have learned something important anyway.
This is an opinion piece reflecting FHC's perspective on a live policy debate. It is not legal, clinical, or procurement advice.
Sources
Digital health's July 2026 signal: AI wearables, a new CMS office, and the telehealth billing fight — MarketScale
The Health AI Brief — Week of July 13, 2026 — Yesil Science
