Pennsylvania's union representing 10,000 state workers negotiated job protections around generative AI in early 2025, making it one of the first labor agreements in the country to directly address the technology's impact on public employees. The deal between SEIU Local 668 and the Shapiro administration includes commitments to keep human workers in decision-making roles when AI tools are used.
Steve Catanese, president of SEIU Local 668, said he approached state officials with concerns about AI implementation and was surprised by their willingness to negotiate. The union's agreement includes a "human in the loop" requirement, meaning state workers oversee, review, and guide AI outputs for accuracy and fairness. It also affirms that state employees are defined as people, not AI substitutes.
"I think right now, whether you work for the government or big business, or whatever type of entity, you want to work for an employer that actually sees in the long term there is a role for humans," Catanese said.
As far as he knows, Pennsylvania is the only state government that has given a union collectively bargained protections around AI use, he said. But he acknowledged the technology is moving fast and the union must remain vigilant to keep those protections intact.
More than 3,000 state employees now use AI tools, up from 175 who piloted the technology in 2024. Another 6,000 are enrolled in training programs. The state's pilot with OpenAI - the company behind ChatGPT - was the first partnership of its kind between a state government and the company. Sam Altman said the pilot "showcases Pennsylvania's commitment to innovation."
State employees use AI for basic tasks: deciphering handwritten clemency applications, scanning documents, and triaging customer service requests. For the 721 employees with access to ChatGPT Enterprise, the cost is $25 per month each, totaling about $216,000 a year at current usage levels.
How the protections work
Governor Josh Shapiro published a public letter to state employees explaining the administration would implement AI to boost productivity "not with the agenda of substituting Commonwealth employees with AI." The state's Generative AI Governing Board reviews its AI policies annually.
Human involvement is required in any work produced by AI. Workers must disclose when customers interact directly with AI. State employees cannot submit personally identifying, confidential, or nonpublic information into AI tools.
The state worked with labor, Carnegie Mellon University, and the University of Pennsylvania on its rollout. Chief Information Officer Bry Pardoe said the aim is for "people to have confidence in the services they get from state government."
The first pilot
The January 2024 pilot involved 175 employees from 14 agencies. Participants included Office of Administration staff, lawyers, and construction project managers. Most used ChatGPT to assist with drafting emails or other writing. Employees reported saving an average 95 minutes per day, and 85% described their experience as somewhat or very positive.
Shapiro said the pilot reduced the time needed to onboard a new employee by 30 days and consolidated 60% of government technology policies. He said too many public officials "think they can stick their head in the sand" on AI. "I believe we need to embrace it and we need to use it for the betterment of society and the betterment of the people we serve."
What the experts say
Mark Fagan, a lecturer at Harvard's Kennedy School, said governments should first find friction points and inefficiencies, then decide if AI is the best solution. "AI is often a solution to pain, but that's not enough," he said. "If you're going to use AI, it should be the best solution to the problem."
Chinmayi Sharma, a Fordham Law professor, said a gap can exist between the appeal and the oversight needed for AI tools. She questioned whether enough workers are trained to "identify subtle or nonobvious errors" in AI outputs. She added that under-resourced governments may face pressure to shift funds away from human employees toward AI tools over time, and politicians rarely enjoy admitting a pilot failed.
Why this matters for government workers
Pennsylvania's experience offers a model for other public sector workers facing AI integration. The state's labor agreement is the only one of its kind that Catanese knows of, meaning most government employees across the country currently work without collectively bargained protections against AI displacement. The outcome shows that early, direct negotiations with officials produced defined safeguards - but those safeguards are not permanent and require ongoing enforcement as the technology evolves faster than policy.
Your membership also unlocks:
