The same way an agency answers a request for email or texts. Treat staff AI chats about agency work as records, find every account where that work happened, have employees search the accounts only they can reach, export what is responsive, review it for exemptions and produce it. Washington cities did exactly this in 2025, for thousands of pages of ChatGPT logs. The hard part is practical: the record usually sits in many separate accounts.
Is a ChatGPT log covered by the request?
When the chat is about agency work, usually yes. State records laws define a record by what it is about, not by the tool that made it; we walk through California, Washington and Texas separately. The state archivists agree. Texas: “AI-generated records should be classified the same as human-generated records.” Washington: “The retention period for public records depends on the content and function of the record, not its format.”
Who has to search, and where?
The agency can search the accounts it controls. For the rest, the courts have put the search on the employee:
- Washington. In Nissen v. Pierce County (2015), employees search “their files, devices, and accounts” and produce “any public records (e-mails, text messages, and any other type of data).” An employee who holds something back as personal “must submit an affidavit with facts sufficient to show the information is not a ‘public record.’”
- California. In City of San Jose v. Superior Court (2017), the agency’s search “need only be reasonably calculated to locate responsive” records, and employees may search their own accounts.
- Texas. Under Government Code section 552.233 (2019), the employee is a temporary custodian and “shall surrender or return the information to the governmental body not later than the 10th day after” the agency’s officer for public information asks.
So for every employee who used a personal AI account for work, the request becomes a separate search, with a sworn statement or a deadline attached.
What does producing the logs look like?
The Washington requests show it. In 2025 Cascade PBS and KNKX asked nearly a dozen cities for two years of ChatGPT logs. Bellingham and Everett “were the fastest and most comprehensive in their responses”; other cities, like Seattle, were “slowly responding to the request in installments.” Most logs came out in date order. Some ended with the employee asking ChatGPT “How do I export my ChatGPT history?”, likely after hearing about the request. Before release, some chats were redacted, for example where they held sensitive city code or details of an active police investigation.
What if the chats were never saved?
Then there is nothing to produce, and the question becomes whether they should have been kept. The Washington requests returned no chat histories from elected officials, and one official said the chatbot had been used while logged out, so nothing was saved. Washington State Archives warns against the opposite mistake too: “Retaining everything is a strategy for chaos.” The archivists’ answer in Washington and Texas is the same: keep AI records under the retention series that fits their content and function, and settle that before the next request arrives.
A checklist for the next request
- Know the tools. List the AI tools staff use for agency work, and which are agency accounts.
- Know who can search what. The agency searches its accounts; for personal accounts, set the employee search, the affidavit or the surrender deadline your state uses.
- Know how to export records from each tool before a request arrives.
- Name who reviews for exemptions. AI chats hold the same confidential material as email: code, investigations, personnel matters.
- Map AI records to your retention schedule by content and function.
Where Verillian fits
Verillian keeps the record of every captured AI interaction on your agency’s own servers, so for the AI use it captures, a records request is a search your staff run. Verillian never receives your prompts, responses or records.
Sources read on 27 October 2026.