“IF YOU SLING MUD, SOME OF IT WILL STICK.”
That saying is not about one lucky guess landing on a bullseye. It is about what happens after something damaging gets thrown. Even when the accusation is wrong, part of the stain may remain. Now put that problem inside a medical record, give the mud a polished AI voice, and let the same unsupported statement get copied into the next note. Bless its little digital heart, the machine may call it a summary. The Veteran may be stuck carrying the stain.
The FDA is asking for public feedback on how generative AI medical devices should be evaluated. The agency's discussion paper raises questions about risk, evaluation before market entry, monitoring after deployment, foundation models, and systems that can take more independent action. Comments are due October 19, but let us not turn that deadline into another breathless headline about FDA blessing or banning AI. FDA says this is a discussion paper. It is not draft guidance, final guidance, or a new set of regulatory expectations.
The useful question is harder and a whole lot more practical: what happens when a generated statement enters the care pathway looking cleaner and more certain than the evidence underneath it? For Wrist to Clinician™, competency cannot stop with whether the model produced a plausible answer. It has to include who checks that answer, what evidence supports it, where it lands, who can correct it, and how that correction follows every place the original statement traveled.
A Polished Note Can Leave a Dirty Trail
Generative AI can produce language that looks finished before the underlying work is finished. That may save a clinician from wrestling with a blank page, but it becomes dangerous when polished language starts passing for proven truth. A generated summary might pull together wearable trends, symptoms, history, and prior notes in a way that saves real time. It might also drop a source, flatten uncertainty, confuse two time periods, or state an inference like it personally stood beside the bed and watched the event happen.
Once that sentence enters a chart, the next person may not see the shaky foundation. They may see a clean sentence inside an official record and carry it forward. The statement can be repeated, summarized again, or used as context for the next decision until the inference starts looking like a fact with seniority. A responsible pathway has to preserve the difference between what a device measured, what the Veteran reported, what the model inferred, and what a clinician confirmed. If all four arrive wearing the same uniform, somebody is eventually going to salute the wrong one.
Correction Cannot Be a Note Nobody Finds
“A clinician should review this” is not a workflow. Which clinician? Where does the generated statement appear? Does that person receive a notification, or is the note sitting somewhere behind three clicks and a password reset, looking official and waiting for somebody to accidentally find it? When is review expected, what requires confirmation, and what happens when the model conflicts with a clinical measurement, the symptoms, or the Veteran sitting in front of the care team saying, “That is not what happened”?
Then comes the part too many technology conversations skip: if the statement is corrected, where does the correction go? Does it reach every downstream note, queue, summary, and person that received the original, or does the clean up crew mop one room while the mud keeps riding the elevator? A correction pathway needs an owner, a record of what changed, a way to notify affected users, and a clear distinction between the original source and the corrected conclusion.
VA Patient Generated Health Data already makes an important distinction because information can be shared without automatically notifying the care team. VA Remote Patient Monitoring Home Telehealth is a different lane with enrollment, equipment, training, a care coordinator, defined review, and monitoring during stated hours. Generated output does not get to erase that distinction simply because the note can speak in complete sentences.
Five Things Should Travel With the Statement
If an AI medical device creates a summary, recommendation, alert, or draft note, I want five things traveling with the statement:
Show the source. Identify the measurements, records, Veteran statements, and time periods that fed the output.
Label its status. Say whether it is measured, reported, inferred, generated, reviewed, or confirmed.
Name the reviewer. Identify the person and the actual workflow in which review happens.
Define confirmation. State what must be checked before anybody acts.
Build the correction path. Track where the statement traveled and how the people relying on it will learn that it changed.
That is not red tape. That is how we keep a smart tool from becoming a confident stranger wandering around the chart with a badge it printed at home. Industry should be able to show those five things without a three hour product demo, and a health system should be able to say where each one lives after the demo team packs up and leaves. If the answer to “How do we correct it?” is “Somebody can add another note,” that is not a correction system. That is putting a clean napkin on top of a mud puddle.
Competency Cannot Be a One Time Test
The FDA discussion paper also asks about monitoring performance after a device reaches the market, and that matters because competency is not a graduation ceremony. Generative systems can change through model updates, data changes, prompt changes, workflow changes, and the environment in which they operate. A system that performed well in one setting can still drift when the patient population, input quality, task, or surrounding process changes. When that happens, monitoring cannot stop at counting incorrect outputs. It must also ask where those outputs went and whether they were corrected before they became somebody else's source.
Industry needs a straight answer for what is monitored, what triggers review, how changes are evaluated, and who tells the care team when performance or intended use changes. Clinicians need to know what the tool can do today, not what the sales deck promised six months and three updates ago. Veterans need to know when AI contributed to something important about their care and how to challenge an output that does not match reality. Everybody does not need the same technical detail, but everybody needs an honest view of the part they are being asked to trust.
The Question for Industry, VA, Clinicians, and Veterans
Here is where I want the conversation to get specific. Industry, can your company show the source, status, review owner, confirmation rule, correction path, and downstream notice without hiding behind a three hour product demo? VA leaders, where should generated output enter the workflow, and which statements should never enter the record without a named owner? Clinicians, what would help you review and correct a generated summary without turning you into unpaid quality control for somebody else's product? Veterans, if AI put something in your record that did not match reality, what would you need to see before you believed the mud was actually gone?
The model may be able to write the note once. The chart may repeat it for years. If AI can sling the mud, the pathway better know who owns the bucket, the mop, and every muddy footprint headed down the hall.
Medical AI companies preparing for VA need more than a strong model and a polished demonstration. They need a defined Veteran use case, evidence boundaries, a named review workflow, a correction and notice plan, a purchasing path, implementation support, and measures that show whether the system helped without shifting an unlimited burden onto clinicians. JhetVet® helps companies assess that readiness, map the VA stakeholders and purchasing path, pressure test the operating model, and build a practical route from product promise to responsible execution.
Building an AI enabled medical device or clinical workflow for VA? Start a conversation with JhetVet® about the Veteran use case, evidence, review ownership, implementation plan, and path to scale.
Educational note: This Article is for general information and is not medical, legal, or regulatory advice.
