Cities and Federal Agencies Are Split on Releasing Officials' ChatGPT Conversations

Government workers are using ChatGPT, Claude, and other AI chatbots to write letters, draft grant applications, and answer constituents. Some of those officials have had to hand over the transcripts. Others are refusing, and the fight over which is legally correct is heading toward the courts.

What's Already Come Out

A 2025 joint reporting project by Cascade Public Broadcasting and KNKX-TV, obtained under Washington's Public Records Act, found city workers and mayoral staff in two Washington cities using ChatGPT to write mayoral letters, generate grant applications, draft policy documents, and compose replies to citizens and reporters. The reporting alleged bid-rigging carried out through AI prompts and responses, according to media lawyer Steve Zansberg's article in the Fall 2026 issue of Communications Lawyer, published by the American Bar Association.

The same reporting found that roughly half of a letter sent by Bellingham's mayor to state authorities, seeking funding for a Native American tribe, was copied verbatim from ChatGPT output, Zansberg writes.

Separately, the Texas Department of Transportation produced numerous employee-ChatGPT conversations in 2025 after a records request under Texas's public information law, according to the same article.

Who's Stonewalling

Not every agency is playing along. Wired magazine reported, and Zansberg's article cites, that both the "Department" of Government Efficiency and the Department of Housing and Urban Development denied FOIA requests from the watchdog group Democracy Forward for AI prompts and responses used in crafting official agency policy. HUD invoked FOIA Exemption 5, the deliberative process privilege, to justify withholding the records.

At the local level, Colorado Springs refused to give a local television station copies of the mayor's AI chatbot interactions, citing two exemptions under Colorado's Open Records Act: material that counts as "work product assembled for the benefit of elected officials" and "deliberative process" documents, according to Zansberg's reporting, which was also republished by the Colorado Freedom of Information Coalition.

The deliberative process privilege exists for a real reason. It's meant to let officials hash out ideas candidly before a policy is finalized, without every draft thought becoming a headline. If an AI chat is functionally identical to an internal staff memo bouncing around ideas before a decision, agencies have a defensible case that Exemption 5 should apply the same way it applies to human deliberation. Whether a chatbot prompt is more like a finished work product or more like an internal brainstorm is exactly the kind of line no court has drawn yet. Zansberg notes that "no court has yet published a decision applying either FOIA or a state public records act to government officials' 'conversations' with AI chatbots."

The Federal Guidance Just Landed, and It's Thin

On August 21, 2026, the National Archives and Records Administration quietly published guidance on applying the Federal Records Act to AI materials, according to Just Security, an outlet that closely tracks executive-branch legal issues. Just Security's analysis, published September 21, 2026, by Asad Ramzanali and Margaret Mullins, argues the guidance only makes federal employees' AI prompts count as records "in narrow circumstances," leaving most day-to-day chatbot use in a gray zone.

Just Security's piece lays out why the stakes are high. The State Department is reportedly using AI to expedite visa revocations, which the outlet says raises First Amendment concerns. A VA inspector general found clinicians using Microsoft Copilot for patient care despite the tool lacking guardrails for high-impact uses like adverse-impact monitoring. Military personnel reportedly used Anthropic's Claude to support an operation targeting Nicolás Maduro. And DOGE staff reportedly submitted nonpublic federal data into Grok in ways that raised conflict-of-interest questions, used an OpenAI model to produce what Just Security describes as "error-ridden analysis" to flag VA contracts for cancellation, and ran employee responses to the "fork-in-the-road" buyout email through Meta's Llama models.

None of those specific federal uses have been confirmed through a public records release. They're described in Just Security's reporting as things that "reportedly" or "may have" happened, based on prior accounts, not documents the outlet has obtained and published in full.

What's Unresolved

The core legal question, whether a government employee's back-and-forth with a chatbot is a "public record" the way an email or memo is, remains untested in any published court ruling per Zansberg. Practical obstacles, like agencies not retaining chat logs at all, or claiming the cost of producing them is prohibitive, will likely shape the next wave of disputes as much as the legal exemptions themselves, according to the FOIA Advisor blog's summary of Zansberg's piece.

Democracy Forward has not indicated whether it plans to sue over the DOGE and HUD denials. Until a court rules, whether taxpayers get to see how their government is actually using AI to write policy will keep coming down to which state they live in and which agency they're asking.