On October 22, the Department of the Treasury published a one-page notice describing Fraud.gov as a government-wide intake and referral portal for reports of suspected waste, fraud, abuse, improper payments, misuse of federal funds, and related misconduct affecting federal programs. The form can collect a person's narrative, supporting documents, contact information, and other details needed to evaluate and route a submission.
The notice also puts an AI step into the public record. An artificial intelligence tool will process the initial entry, populate standardized fields, and suggest where the report should be routed. The person submitting the tip can review and edit those fields before sending it. Treasury personnel will then apply review and quality-control procedures before a referral goes to a federal agency, inspector general, law-enforcement organization, or another authorized entity.
This creates a visible chain between what a person reports and how the government first organizes it. The initial narrative remains the source material described by Treasury. The AI turns that material into structured fields and a routing suggestion. The respondent gets a correction point before submission, and Treasury adds another review before referral. The notice does not give the AI authority to decide whether an allegation is true or whether an investigation should begin.
What changed
Treasury is seeking an extension without change of a currently approved information collection under the Paperwork Reduction Act. That procedural label matters because the notice does not establish a new enforcement power or prove that every part of Fraud.gov is newly built. At the same time, Treasury says it is launching the portal and describes the AI-assisted intake sequence in concrete terms.
Treasury expects 80,000 respondents each year, with one response per person. It estimates 30 minutes per response and 40,000 total annual burden hours. The portal is intended to help Treasury receive, review, triage, categorize, de-duplicate, and refer allegations. Those functions turn a public report into material that can move across institutional boundaries.
The AI tool operates near the start of that process. It structures the initial entry and proposes routing before submission. A person may edit the populated fields, which creates a chance to catch a wrong program name, an omitted detail, or a category that changes the apparent destination. Treasury's staff review comes later, before referral.
Why it matters
A government-wide portal has to translate ordinary descriptions into the categories agencies use. People reporting suspected misconduct may know what happened without knowing which office owns the program, which inspector general has jurisdiction, or how Treasury labels a particular allegation. Standardized fields and routing suggestions can reduce that translation burden.
The same step can also shape what the institution sees first. If the AI compresses a detailed account into the wrong category, leaves out a qualification, or suggests the wrong destination, the structured version may frame the tip before a receiving office reads the underlying narrative. Respondent editing and Treasury quality control are meaningful safeguards. Their value depends on whether people can recognize a bad transformation and whether reviewers can compare the structured record with the original submission.
Treasury's notice gives no error rate, model identity, or description of how often users change the populated fields. It does not say whether the original narrative, AI output, respondent edits, and final routing decision remain linked in an audit trail. Those omissions do not show that the controls are absent. They mark the questions this particular notice cannot answer.
Verification bottleneck
The critical record is the path from allegation to referral. A useful receipt would preserve the original narrative, the fields the AI proposed, the respondent's edits, Treasury's review, and the final destination. Without that chain, a receiving office may see a clean set of fields without being able to tell where the system summarized, inferred, or misclassified the report.
The notice also leaves the downstream process outside its scope. It does not say that every tip will be referred, explain how a routing suggestion affects investigative priority, or describe what a receiving agency must do. A referral is a handoff for review and potential action, not a finding against a person or organization.
Watch next
Comments on the information collection are due December 21, 2026. The most useful submissions would press for measurable details: which fields the AI may populate, how the interface shows uncertainty, whether the system retains the original text beside the structured record, and how Treasury evaluates routing errors before and after referral.
Separate primary records may be needed to answer privacy, retention, security, access, and correction questions. Any later system-of-records notice, privacy impact assessment, OMB package, inspector-general procedure, or technical documentation could show whether the correction points described here survive into the operational record.
This Signal Card draws on a Paperwork Reduction Act notice. The source supports the collection purpose, AI-assisted field population and routing suggestions, respondent review and editing, Treasury quality control, estimated burden, and comment deadline. It does not establish model performance, investigative outcomes, or the complete data-governance design for Fraud.gov.
Source
- Department of the Treasury, Federal Register, FR Doc. 2026-21578, published October 22, 2026.
