The government misdirected $186 billion last year. Half the agencies fail the law that's supposed to count it.

The finding, in one paragraph: the federal government's official count of its own misdirected money reached $186 billion in fiscal 2025, and the count itself is the least reliable part of the story.▸ The law that requires agencies to measure improper payments is failed, in whole or part, by half of the agencies that matter; seven of them produce estimates their inspectors general judge unreliable; at least one large program is legally barred from producing a number at all. $186 billion is not the size of the problem. It is the size of the part the government is currently able to see.

The record here is one primary report and the oversight around it.▸ GAO-26-108694, published April 27, 2026, is the government's annual payment-integrity scorecard: 15 agencies, 64 programs, $186 billion estimated improper — up $24 billion from FY2024's $162 billion — and about $3 trillion cumulative since FY2003. Around it: the House Oversight subcommittee's request that GAO examine the cumulative $2.7 trillion highlighted at its hearing and produce anti-fraud recommendations (Sessions and Mfume, bipartisan), the Congressional Research Service's survey of legislative fixes (R48296), and — for the largest single-agency slice we have covered before — the SSA inspector general's $72 billion improper / $23 billion uncollected accounting.

About $153 billion — 82 percent — was overpayments: money out the door that shouldn't have gone, the recovery of which is its own costly system (we have measured one corner of it: Social Security spends $1.77 to claw back each dollar of its smallest overpayments).▸ Roughly $10 billion ran the other way — underpayments to people owed more. The Labor Department's unemployment insurance program alone accounted for about $5.6 billion at a 14.9 percent rate, which its administrators attribute to the antiquated state IT that processes claims.

The Payment Integrity Information Act exists to make this measurable.▸ Read GAO's compliance table against the totals: of the 24 agencies that account for 99 percent of reported improper payments, 12 fully complied with PIIA's criteria and 12 did not — five with inadequate risk assessments, seven with estimates their inspectors general deem unreliable. The instrument disagrees with itself: the same reporting regime that produced the $186 billion headline cannot certify the components it is built from. And some of the map is simply blank — HHS pays tens of billions through Temporary Assistance for Needy Families but, as the coverage of the report notes, does not calculate improper payments there "due to statutory limitations." A second disagreement is quieter: part of the $24 billion year-over-year increase reflects more programs being measured, not more money lost — which cuts both ways, since it means prior years' totals were undercounts.

The unemployment insurance case compresses the whole story into one program: states cite "outdated IT infrastructure" for a 14.9 percent error rate — while having accepted more than $204 million in UI IT modernization funding, a contradiction pointed enough that Labor's inspector general has opened an investigation into where the modernization money went.

Three open threads, each with a document attached.▸ GAO owes the Oversight subcommittee a response to the $2.7 trillion examination request. Labor's IG investigation into state UI modernization spending is newly opened. And the legislative fixes cataloged by CRS R48296 — including making more programs measurable — wait on Congress. There is genuine good news in the report: Medicare fee-for-service and FEMA's validate-as-you-go program both cut their improper totals materially, proof the number moves when a program is instrumented and worked.

Figures are GAO's FY2025 estimates as published April 27, 2026; the composition remainder is derived arithmetically from GAO's own components and labeled as such. Program-level UI figures as reported from the GAO data by Federal News Network.