GAO: AI coding tools may capture more diagnoses, and few studies have checked their accuracy

A federal watchdog now says in writing that AI coding tools may pull more diagnoses out of an encounter than a coder would, and that hardly anyone outside the vendors has measured how accurate they are. Every extra code still needs documentation behind it, and the ones that don't are the population an auditor samples.

That's the Government Accountability Office's Science & Tech Spotlight "AI for Medical Notes and Coding" (GAO-26-109116), published July 16, 2026. Two pages, no GAO-collected sample, and GAO's own footer says it "is not an audit product." It matters anyway, because there are few independent studies evaluating whether these tools are accurate, and the one accuracy figure GAO relays came from a vendor describing its own product.

GAO's summary line is the one to keep: "The accuracy of these tools may be difficult to verify, and the overall effects on health care spending are uncertain."

What GAO said about human coders

GAO doesn't hedge on where this is going. In its own voice:

"AI tools that analyze patient records and suggest codes for review by medical coders are already in widespread use. New AI tools may use generative and agentic AI technologies to review patient records and assign codes autonomously. This capability could make human coders faster or replace them entirely."

That's a federal agency putting the displacement question in a public document. Note the split GAO draws and most coverage won't: suggest-for-review tools are the ones already everywhere. Autonomous assignment is the new thing.

The rest of the Spotlight is why it isn't settled. GAO says there are "few independent studies evaluating the accuracy of these tools," and that some tools may not store recordings and transcripts, "which may limit the extent to which health care providers can conduct independent assessments."

The 95% number is a vendor's own, and GAO says so

The figure that'll get pulled out of this report and pasted into slide decks reads, in GAO's text:

"According to one AI medical coding software developer, when its software was deployed at a health system with five hospitals, it generated medical codes with more than 95 percent accuracy, and the system's emergency departments reduced annual coding costs by more than $1 million."

One unnamed developer. One health system. Reported by the developer. GAO relays it and, in the same document, concludes accuracy is difficult to verify because independent studies are few. Anyone quoting the 95% without the sentence that follows it is quoting a marketing claim with a government seal borrowed onto it.

Who actually said what

The challenge that actually lands on your desk

Of GAO's four challenges, one is a coding problem rather than a policy problem. GAO writes that these tools "could increase health care costs if they capture more diagnoses and services rendered during visits than non-AI approaches," and that inaccuracies "may result in patient harm or over- or under-reimbursement from insurers to providers."

Read that from a coder's chair and it stops being an abstraction. It is the same problem as an AI-drafted note that still needs coder QA, one step further down the pipe: now the tool isn't just writing the narrative, it's proposing the codes. If a tool systematically pulls more diagnoses out of the same encounter than a coder would, the extra codes aren't free. Each one either has documentation behind it or it doesn't, and the ones that don't are exactly what an auditor samples. "The tool suggested it" has never been a defense.

GAO's four challenges, and what you do about each

The capture-delta audit

GAO names the over-capture risk and stops. It's a watchdog Spotlight, not a procedure. Here's the procedure.

Pull a sample of encounters the tool has already coded. Have a credentialed coder assign codes from the same record, blind to the tool's output. Then diff the two sets in both directions.

Every code the tool assigned and the coder didn't goes into one of three piles:

• Supported: the documentation's there, the coder just missed it. That's the tool earning its keep.
• Query pile: the dx is in the note, but there's no monitoring, evaluation, assessment or treatment behind it at that DOS. Not a code yet. That's a query.
• Not supported in the documentation: nothing in the record backs it. That's the pile that turns into a repayment.

Then run the reverse column: every code the coder assigned and the tool didn't. GAO names under-reimbursement in the same breath as over-, and a tool that quietly drops a supported chronic condition costs you a recapture nobody flagged.

How those piles split is the only accuracy number for your shop that nobody has to take a vendor's word for. Run it again after any model or version change, because the answer isn't stable across updates.

This is the chart-level version of the claims-level check in coding intensity is climbing where AI documentation landed. That one compares documented severity against documented treatment across a population. This one compares two coders on one chart.

Where a human coder is still required

All four of GAO's challenges come back to a question a tool can't answer for you: is this diagnosis actually addressed at this encounter, or is it just in the note? That's a read of what the provider documented against what the code set demands, and if the answer is no, there's a repayment on the other side of it.

GAO says policymakers need more information about the performance of these tools to determine the appropriate level of oversight. That information doesn't exist yet. Until it does, the only accuracy number your shop has is the one you measured yourself.