Nevada DETR generative-AI unemployment appeals

Nevada's Department of Employment, Training and Rehabilitation (DETR) contracted Google to build a generative AI, one that writes text, for unemployment appeals. It reads a hearing's transcript and evidence, and searches Nevada law and past appeals decisions. It then drafts a recommended ruling, to approve, deny, or modify the claim, with the written decision, for a referee's review and signature.

How it is meant to be used

Claimants can appeal decisions on their unemployment benefits. Each appeal has a hearing, recorded in a transcript, with evidence. An appeals referee signs the ruling, which is then issued to the claimant. The case file does not define the referee's role further.

Officials presented the AI as a way to clear a backlog of appeals from the pandemic. More than 10,000 appeals were outstanding as of June 2024, with roughly 1,500 pandemic-era claims still pending. The Markup separately reported a drop from more than 40,000 at the pandemic peak to under 5,000.

A referee's determination could take as much as several hours, elsewhere put at about three hours. Officials projected about five minutes per case with the AI. DETR said the required human review adds an estimated 10 to 30 minutes. The five-minute figure leaves that review out.

Who decides

A referee must review and sign each drafted ruling, and DETR says two state workers are involved. The sources do not say whether the referee is one of the two. Director Christopher Sewell said no AI-drafted written decision goes out without human review. He called AI "a great tool" and added, "It's a tool." The signed ruling is then issued to the claimant.

The design's safety rests on that human sign-off. The AI exists to clear a backlog, and the case file argues that a referee who keeps rejecting drafts is then cast as the bottleneck. The drafts carry both the ruling and its reasoning. From outside, a real review and a rubber stamp can look exactly the same.

How accuracy is judged

The contract set a 90 percent success requirement. State workers had to deem the AI's decision correct nine times out of ten on test decisions. The sources do not say what the workers compared each decision against. No independent outside audit checks that grading. DETR said it wanted accuracy above 90 percent before going live.

Critics cite outside studies of legal-research software built by the same method. Those found 17 to 33 percent of answers incorrect and 18 to 63 percent incomplete. The studies did not measure Nevada's AI.

DETR's internal governance committee watches for bias and for hallucinations, content an AI makes up. It meets weekly during fine-tuning, when the AI is adjusted before launch, and quarterly after launch.

Testing, delays, and cost

Rollout was delayed again and again because accuracy in testing fell short. The AI cited incorrect Nevada statutes and failed to pull information from all hearing documents. Officials said those problems were fixed.

In 2024, officials projected a launch within months. As of the latest reporting, in March 2026, the AI was still being tested on past appeals. It was described as launching in the coming weeks. No source confirms it is deciding live claimant appeals.

The reported cost grew over time. It was about 1 million dollars when the Board of Examiners approved the contract in 2024. A later legal analysis put it at roughly 1.38 million. By early 2026 the total price was 2.6 million dollars, with about 1.1 million spent.

Data and consent

The AI runs on Google's Vertex AI Studio cloud platform. DETR says the data stays in the continental United States, and the state holds the encryption keys. It says Google gave assurances it would not access personal information for other purposes.

Claimants are not required to consent to AI processing of their appeal. Reporting says it is unclear whether they are told the AI is involved. No opt-out is described.

What each side says

Legal scholars, attorneys who represent claimants, and a former U.S. Department of Labor official warned that backlog and speed pressure could hollow out the review. They said it could create incentives to rubber-stamp the AI's drafts. One attorney said the time savings "only happens if the review is very cursory." A legal analysis warned that staff "might feel pressured to authorize AI decisions on appeal with haste."

State Sen. Dina Neal said Nevada would be "contracting a Nevada citizen's rights without their consent, without their knowledge." Sen. Skip Daly said, "I don't think there should be a reliance on this, and this is where it starts."

A UNLV computer scientist warned of possible model bias and hallucinations. A public-employee union argued AI should not replace worker judgment. Sen. Neal said she plans to reintroduce AI-oversight legislation after an earlier bill failed.

This deference risk is projected by experts, not measured. No referee override or rejection rate has been published.

What this network is drawn from

This network follows the pattern the case file describes. It does not reconstruct the actual AI. It shows the drafting AI, its search of Nevada law and past appeals, the appeals records, the referees, DETR's governance committee, and Google's cloud. Claimants are outside the network.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network where a mistake made by one part can be passed on to the other. A tool closes a pathway when mistakes stop passing along it. The work along it may go on.

This case has a budget of 12 units. Explore (No Targets) sets no targets. The other three levels all ask for the network's mistakes to be contained, meaning corrected rather than building on each other.

Under Service Targets Only, the targets can be met with every tool at its standard setting. The cheapest combinations cost 4 units, and each includes Escalate checks. One is Escalate checks with Gate vendor updates. Escalate checks at its stronger setting also meets them alone, for 4 units.

Under Service and Safety Targets, the targets can be met. This level also asks you to close every failure pathway, among other targets. Six are open before any tool is used. The first three are Drafted ruling handed to the referee, Law and past decisions searched for each draft, and Draft saved to the appeals records. The other three are Signed decision filed in the appeals records, One AI drafts every appeal, and Appeal files processed on Google's cloud.

The cheapest combination costs 10 units: Escalate checks, Gate vendor updates, Gate record entries, and Vet connections. Every combination that meets these targets includes all four. Escalate checks is the only tool on offer that closes the pathway from drafts to the referee.

Under All Governance Targets, one combination within the budget meets the targets, and it spends the whole budget of 12 units. It is the same four tools plus Understand the system. At this level Understand the system costs 4 units. While it is on, four tools cost 1 unit less: Keep skills sharp, Escalate checks, Upgrade model, and Vet connections.

At this level, Vet connections and Check copied records work at full strength only while Understand the system is on. Without it, the four tools leave two pathways open: Law and past decisions searched for each draft, and One AI drafts every appeal.

More is not better here. Using every tool at its strongest setting costs 38 units, over three times the budget. It contains the mistakes and closes every failure pathway. It still meets the targets at none of the three levels that set them, because the drafting AI then adds too little to the work.