Nevada’s employment agency is rolling out a Google-run artificial intelligence tool to process appeals on unemployment benefit decisions, a development the agency said would rapidly speed up the process, but one that has some state lawmakers skeptical because of potential transparency and consent issues.

The Nevada Department of Employment, Training and Rehabilitation (DETR) has eyed using the tool since summer 2024. But less-than-desired accuracy rates prolonged the rollout, which has a total price tag of $2.6 million. About $1.1 million has been spent so far.

Other state agencies are already using AI, including the Nevada DMV’s AI-powered chatbot, but DETR’s planned use of this product in the coming weeks marks a dramatic expansion on the state government’s AI use. The agency had also planned to use AI for invoicing, but that project is not moving forward after determining it was not meeting the “promised benefits and expectation,” a spokesperson said.

State officials have touted AI as a way to simplify and accelerate existing processes, but they have also cautioned that it must not replace human oversight. Two state workers will be involved in the process, and officials have emphasized the AI tool is not making the final decisions because it requires human verification.

“AI is a great tool — but that’s what it is. It’s a tool,” DETR Director Christopher Sewell said in an interview with The Nevada Independent. “We have to have human review with everything that we do.” 

Still, there are skeptics in the Legislature about the project’s security and transparency, such as the fact that people are not required to give their consent for their appeal to go through the AI-driven process.

“When you start to contract with an AI company, where does the Nevada citizen fit inside of that relationship?” said state Sen. Dina Neal (D-North Las Vegas), who sponsored a failed bill last year to dramatically increase state oversight of AI. “Because you’re contracting a Nevada citizen’s rights without their consent, without their knowledge. And that is backwards.”

The release of the AI tool comes after the agency suffered a massive pandemic-induced backlog of unemployment appeals, which has since subsided. Officials said the tool could issue a ruling in five minutes, whereas the process without AI can take anywhere from 10 minutes to several hours depending on the complexity of the case.

Sen. Skip Daly (D-Sparks), who has been a vocal critic of the state’s use of AI, was skeptical of the purported efficiency improvements if humans still have to review the AI’s work, which can add between 10 and 30 more minutes to the process, an agency spokesperson said.

“I don’t think there should be a reliance on this, and this is where it starts,” Daly said. “You get used to it, and then you get comfortable with it.”

Under the current process, once a person appeals an unemployment benefits decision, a state employee serving as a kind of referee will conduct a hearing where all sides present evidence and testimony. The referee will then issue a written decision within 30 days.

Under the new process, the AI tool would review all information from an appeals hearing, including any relevant documents, and check any related Nevada laws or regulations — responsibilities that previously only were assigned to humans. It would then issue a ruling that must receive sign-off from the referee.

The agency has been testing the tool with old appeals to ensure it comes to the right decision. At times, the tool referred to an incorrect Nevada law and was not gathering information from all appeals documents, but Sewell said both of those issues have been fixed.

Sewell added that teaching the AI to make accurate rulings “is taking a lot longer than we thought.”

The contract with Google, which state officials approved in August 2024, required that the system have a 90 percent success rate, meaning state workers deemed the decision was correct nine out of 10 times. The state wants the true accuracy rate to be a bit higher.

Sewell said the need for accuracy is twofold.

“It could be dealing with a claimant — whether they’re eligible or not for benefits — and I want to make sure that if they’re eligible, they’re eligible, and they get their benefits,” he said. “Second, these determinations could end up in a court of law in Nevada.”

There are various security measures in place. The platform will ensure that the data does not leave the continental U.S., the state will have control over encryption keys and it is also in compliance with state laws on IT and personal information security.

After reviewing contract and security protocol documents, Junggab Son, a computer science professor at UNLV whose research areas include AI algorithm vulnerabilities and privacy protection, told The Indy that he believes the Google platform itself is secure, but he still has concerns on the actual deployment of the AI tool.

He said human reviews could place too much trust in the AI’s decision, and there is still a potential for model bias and hallucinations, where an AI tool produces information not grounded in reality but presented as fact.

“These kinds of things, maybe they need to address a little bit more,” he said.

Neal is also wary of giving Google access to Nevada government data.

“When you partner with certain companies who basically have control over almost all of our data, it is a problem,” Neal said. “It is almost becoming a monopoly on who controls data.”

Sewell said agency employees are “working on the security aspect of it daily.”

What AI oversight looks like

The most direct form of state oversight is a 2024 policy on agency AI use. It prohibits agency-level AI policies from being more lenient than the statewide one, as well as the use of AI to create discriminatory content or use personal data without anonymization.

The policy was published by the Governor’s Technology Office, which is largely responsible for setting statewide standards rather than directly involving itself in the daily operations of agency-level IT projects, said Michael Hanna-Butros Meyering, the office’s chief communication and policy officer.

He also chairs a group of agency IT officials that meets to discuss AI-related projects or problems they are experiencing. The group is planning to ramp up its efforts amid the rapid growth of AI.

There were talks in Congress last year of instituting a 10-year prohibition on state-level regulation of AI, but those failed to gather support after bipartisan opposition. 

President Donald Trump has since signed an executive order calling for a “a minimally burdensome national standard” on AI policy. The order also calls for a federal task force to challenge laws inconsistent with this expectation.

Daly said oversight is essential.

“If we’ve learned nothing from the little examples [of AI use] that we have now and everybody’s just racing forward … at some point it’s going to affect industries and people’s jobs,” he said.

In a statement, the American Federation of State, Council and Municipal Employees Local 4041, which is a union representing state workers, said “AI should never be used to replace public service workers’ judgment, expertise, experience and skills.”

“We know that AI tools often hallucinate and use biased data that can produce biased results,” the statement said. “And public programs — especially those that evaluate and make decisions about issues like benefits eligibility — must be held to the highest standards for responsible AI use.”

Meanwhile, Neal is planning to bring back an AI oversight bill during next year’s legislative session.

“It’s going to happen,” she said.