The government is using ChatGPT to hunt $100 billion in Medicaid waste. The tool's known weakness is the catch.

The federal government is now using ChatGPT to look for fraud. In a program announced on May 21, the Department of Health and Human Services began running state and grantee audit reports through ChatGPT and other AI tools across all 50 states, chasing an estimated $100 billion to $200 billion in wasteful or fraudulent spending a year, by the estimate of the official leading it. As a use of AI, this is one of the more sensible ones in government. As a basis for decisions that can cut a state’s funding, it inherits the one weakness everyone already knows the tool has.

What the program actually does

The initiative is called AERO, for Audit Enforcement and Risk Oversight, and it is led by HHS assistant secretary for financial resources Gustav Chiarello. It uses AI tools to scan at least five years of audit history across all 50 states on a rolling basis, for any entity spending more than $1 million a year in federal funds. The program was first reported by The Wall Street Journal and uses ChatGPT and other AI tools to analyze the audit reports, and HHS sent letters to all 50 state governors and treasurers putting them on notice. The targets are chronic noncompliance, repeat deficiencies, and delinquent audits, and the penalty is loss of funding.

The framing from officials is that audits used to pile up unread. As Chiarello put it, the reports would land with a thud and no one did anything, and AI lets them dig into the backlog. It marks a shift from the old pay-and-chase model, paying first and clawing back later, toward flagging problems earlier.

Why the use case is reasonable, and where it isn’t

There is a version of this that is genuinely well-matched to what AI does. The system is reading large volumes of already-public audit documents to surface patterns a human backlog would never reach, which is close to the ideal task for a language model: a fast first pass over text that a person then verifies. Finding which of thousands of audits deserve a human’s attention is exactly the kind of triage these tools do well.

The catch is what happens after the flag. Critics note that the same tools frequently make mistakes and can carry unintended biases, and should be used with caution, and the stakes here are not a draft email. They are a state’s federal funding. The administration has already withheld hundreds of millions from Minnesota and more than $1 billion from California, and critics argue enforcement has disproportionately targeted Democrat-led states. Whatever one makes of that dispute, a confident AI error that helps justify a funding cut is a different order of problem than a confident AI error in a chatbot.

The precedent worth watching

AERO does not sit alone. The same government now forcing frontier models offline over safety concerns is also building AI into how it polices public money. The program builds on a February request for comment on broader fraud-detection rules, and the enforcement machine around it is already large: CMS cited a 2025 baseline of $5.7 billion in Medicare payments suspended and thousands of billing privileges revoked. What makes this one notable is not the dollar figure but the method. A federal agency has put a known-fallible language model into the workflow that decides which states get scrutinized and, downstream, which get paid.

The right guard is the same one emerging across every serious institutional use of these tools: treat the AI as a first-pass reviewer whose output a human verifies before it carries a consequence, never as the authority that delivers the consequence itself. Used that way, AERO is a reasonable answer to a real backlog. Used the other way, it turns a model’s most documented flaw into a lever over public money. Which version this becomes is the part to watch, because the program is rolling, not a one-time scan.