Many state government employees’ conversations and prompts with artificial intelligence chatbots will likely remain shielded from public record, after a ruling from Pennsylvania’s open records agency this month.
WITF sought employees’ ChatGPT logs from more than two dozen state agencies last year under Pennsylvania’s public records law. Gov. Josh Shapiro’s office rejected those attempts, arguing the chats were exempt because they were “notes and working papers used solely for that official’s or employee’s own personal use” and “internal, predecisional deliberations.”
Pennsylvania’s independent Office of Open Records, which oversees appeals to denied public records requests, largely sided with the administration.
The decision reveals a potential gap in Pennsylvania’s public records law. And transparency advocates say it could block government employees from facing accountability in how they use the technology.
But the ruling does not mean that AI chat logs are always exempt, according to Melissa Melewsky, media law counsel for the Pennsylvania NewsMedia Association, of which WITF is a member. Every AI chat is “presumptively public,” she said, unless proven exempt by an agency.
“If AI is used to create a policy, the policy would be a public record,” Melewsky said. “If AI is used to formulate and create a press release, the press release is public.”
In Washington, which has a more open public records law than Pennsylvania, a reporter at public media station KNKX obtained thousands of pages of ChatGPT conversations from local government officials. Those documents revealed that government staff used the software to respond to constituent emails, generate social media posts, draft letters seeking state and federal funding, write speeches and talking points, and perform many other day-to-day tasks.
KNKX also found that many of the AI-generated documents lacked a clear label noting the technology’s use — in direct violation of the state’s guidelines.
The Shapiro administration maintains a similar disclosure policy in Pennsylvania.
“Generative AI use must be disclosed even if it was only used to generate a portion of the content,” the policy reads. “The disclosure shall be prominently displayed and include an indication that the content was generated either entirely or in part by generative AI and identify the Generative AI system and version that was used.”
But the Shapiro administration’s policy allows generative AI only for drafting materials that are then reviewed by a human employee before being published.
That clause — if followed by every employee as written — could lead to most AI use being exempt from public record, since the governor’s office argued the documents were solely “research and initial drafting” in denying WITF’s request.
Philip Hensely-Robin, executive director of good-government advocacy group Common Cause Pennsylvania, said knowing whether Shapiro’s administration is following its own disclosure policy is unclear.
“For the public to be fully on board and to have confidence in these initiatives, that would really benefit from increased transparency about how (AI is) being used and if the safeguards that the administration has put in place are actually being followed,” he said.
Melewsky agreed.
“There cannot be accountability without access,” she said. “There’s going to be a struggle to have any bright line test apply here (for AI).”
EMBRACING AI
Throughout 2024, the Shapiro administration conducted a pilot program with OpenAI, the creator of ChatGPT, to implement a more secure tool than their commercial product into state government operations. The end goal, officials said, was to enhance staff productivity.
That test run saved participating employees — 175 total across 14 state agencies — about 95 minutes per day in writing, researching, summarizing and technology support, according to the administration.
Shapiro’s office said the licenses needed for roughly 300 employees using OpenAI cost more than $7,500 per month total — or $90,000 annually. Another 2,200 or so use other AI tools like Microsoft Copilot, which has its expenses included as part of a general Microsoft license agreement.
There was a single chat log generated during the pilot program shared with WITF through its record request. And the information from that conversation matches some of what was included in Shapiro’s newly released Housing Action Plan.
The anonymous employee asked ChatGPT 5.2 to summarize a Center for American Progress article focused on “Expanding Housing in America by Reforming Local Land Use.”
The software interpreted the article as recommending that policymakers reform zoning laws by ”allowing more multi-family housing, reducing parking requirements, and permitting accessory dwelling units.” It also suggested a program to incentivize local governments to reform their policies in “high-opportunity areas.”
“Can you provide more details about the proposal?” the employee asked. “Are other states successful in doing this?”
ChatGPT pointed to states like California, Oregon and Washington and wrote a glowing review about their policies to permit higher-density housing and accessory dwelling units, which it said have expanded housing supply and affordability.
“A ‘high-opportunity area’ refers to neighborhoods that offer strong access to resources like quality schools, job opportunities, public transit, and safe environments,” ChatGPT wrote.
Shapiro’s plan uses similar language, though the overlapping proposals are commonly pitched as reforms by housing advocates.
“Local land use regulations – such as restrictive zoning, excessive parking mandates, and lengthy permit approval processes – further limit housing construction, particularly in areas where demand is highest,” Shapiro’s plan states.
The Shapiro administration, in response to this story, said the ChatGPT summary “did not meaningfully influence decision-making or determine the content of the Housing Action Plan” and therefore no disclosure of AI use in the housing plan was necessary.