Public-sector unions gain AI guardrails amid widespread adoption

By Rebecca Stone

Nearly 70 percent of U.S. government agencies already use generative AI, according to an Ernst and Young survey cited by The Guardian.

That figure places AI policy inside public-sector operations at scale, even as unions negotiate limits on surveillance, automation and job displacement. The stakes include which systems agencies may use, which decisions remain under human control and what happens when AI changes staffing.

Kung Feng, a labor researcher at UC Berkeley, said employers are pushing automation to cut costs, according to The Guardian. His assessment supplies the economic pressure behind the labor response. The union gains recorded in California, New York, Pennsylvania, Ohio and Oakland show workers converting that pressure into specific protections.

What is driving the union push?

Nearly 70 percent is the strongest numerical measure in the report, but it measures agency use rather than the depth of deployment or the strength of safeguards. The Guardian does not provide the survey sample, survey date or a definition of "use," so the figure cannot establish how many workers face AI-assisted management or automated decisions.

The cited examples do not supply a national baseline for public-sector AI rules. They come from union bargaining, executive action, a product pilot and legislation. That mix matters because each route places control in a different place: a bargaining table, a governor's office, a deployment decision or state law.

The policy question has therefore moved beyond simple adoption. The Guardian's examples ask whether employees can challenge surveillance, retain human oversight and avoid dismissal or displacement when public employers introduce AI. They also show unions pursuing different remedies for the same cost-cutting pressure that Kung Feng described.

What protections have unions won?

• In Oakland, IFPTE Local 21 secured retraining programs instead of AI-related layoffs after a strike vote in October, according to The Guardian. The result substitutes a transition program for AI-related dismissal.
• In Pennsylvania, Gov. Josh Shapiro guaranteed SEIU Local 668 human oversight during a pilot of OpenAI's ChatGPT Enterprise, The Guardian reports. The guarantee attaches human oversight directly to the named product deployment.
• In Ohio, the Transport Workers Union won veto power over autonomous vehicles, according to The Guardian. The measure gives the union a formal role in deployment decisions involving a form of workplace automation.
• In California, The Guardian reports that Gov. Gavin Newsom banned AI surveillance software, often called "bossware," along with automated firings in September. The action bars two uses outright rather than making them subject to review.
• In New York, a law has protected state employees from AI displacement since late 2025, according to The Guardian. The measure creates a statutory response to job loss tied to AI use.

How do the protections differ?

The five examples cover different stages of AI deployment, from hiring and dismissal to software oversight and autonomous vehicles. California restricts surveillance and automated firings. New York addresses displacement. Oakland pairs automation with retraining. Pennsylvania embeds oversight in a ChatGPT Enterprise pilot. Ohio gives a union veto over autonomous-vehicle deployment.

That range matters because the reports do not describe one uniform model for public-sector AI. The measures vary from a training commitment to a statutory protection and a prohibition. Each reflects the specific workplace or policy in The Guardian's examples, rather than a single national bargaining standard.

The record also shows that unions are not rejecting AI in all forms. The leaders cited by The Guardian support safety systems when those systems help workers, while attaching human-control requirements to their use. Their approach centers on workplace deployment, job security and human control rather than a blanket ban on AI.

What do unions require from workplace AI?

Liz Shuler of the AFL-CIO, Lorena Gonzalez, and John Samuelsen of the Transport Workers Union support AI safety systems where they help workers, according to The Guardian. The report describes their shared condition as "mandatory human control over every AI application in the workplace."

That position accepts AI use while contesting the transfer of final authority to an automated system. The Guardian does not define the required control, identify who exercises it or specify what qualifies as an AI safety system. Those omissions leave room for disputes over whether a safeguard meets a union's test.

Pennsylvania shows oversight attached to a specific pilot, while Ohio places a veto before deployment. California goes further for surveillance and automated firings by banning both. New York sets a legal protection against displacement. Together, the examples cover complementary points in the employment relationship rather than one identical safeguard.

Why does this matter beyond the workplace?

Nearly 70 percent agency use makes the union campaign a public policy issue, not a series of isolated labor disputes. Public employers are adopting generative AI at a measured rate, and the rules written into labor agreements, executive actions and law will shape how that technology reaches employees.

OpenAI's ChatGPT Enterprise pilot in Pennsylvania shows how a commercial AI product can enter a public workplace under an explicit human-oversight guarantee. The Oakland retraining program and the New York statute address employment directly.

The Ohio veto gives organized labor a role in autonomous-vehicle decisions, while California's ban bars specified surveillance and firing tools. The Guardian's examples therefore cover deployment, job transition, legal protection and automated management.

The spread across five jurisdictions matters for other public employers facing the same employer pressure to automate. The reported measures do not create a universal standard, but they give unions and governments concrete models for other bargaining negotiations and public policy proposals. Whether those models endure will depend on implementation details that The Guardian's report does not set out.

What remains unresolved?

• The Guardian does not give the sample size, field dates or methodology for the Ernst and Young survey, so the nearly 70 percent figure cannot be compared with another adoption measure from the cited material.
• It does not define "AI safety systems" or "mandatory human control." No technical standard, review procedure or enforcement body appears in the examples.
• It provides no scope, funding or timetable for the Oakland retraining programs. It also does not identify which departments or job categories the Pennsylvania pilot covers.
• The Guardian gives September for Newsom's action and October for the Oakland strike vote without a year. It does not detail the Ohio veto, the terms of the New York law or the process for enforcing California's ban.

These gaps do not erase the reported protections. They do make the safeguards harder to compare and apply elsewhere. The survey figure frames the scale of the issue; the measures show the direction of the response.

With agency use already near 70 percent, the next test is whether the human-control conditions unions demand can govern real deployments as public-sector AI use expands.