AI prompts and chat logs as public records: retention and disclosure

Reference article · 5 min read

When an employee uploads a staff report to an AI assistant and pastes the answer into a memo, four things have been created: a prompt, a copy of an agency document sitting in a vendor environment, a generated response, and a conversation history in somebody's account. Records officers and counsel eventually ask whether any of that is a public record. In most states some of it almost certainly is, and the harder problem is that the agency cannot find it, cannot export it, and has not decided how long to keep it.

Public records statutes differ by state and the case law on AI is still developing. What follows frames the question for your records officer, your counsel, and your vendors; confirm specifics against your state statute and your state archives or records management program.

Why prompts and outputs usually meet the definition

Most public records laws define a record by content and use rather than by format. The common formulation covers a writing, recording, or electronically stored information prepared or received by an agency in connection with public business. Nothing there excludes text typed into a chat window.

Prompts. A request for a policy summary, an analysis of complaints, or a rewrite of a staff report is agency business expressed in writing, and it often quotes the underlying record.
Uploaded files. The document is usually already a record elsewhere. The upload creates a copy in the vendor environment, which is a custody question more than a records question.
Outputs. Once generated text is pasted into an email, a document, or a case note, it is a record in that system regardless of what happens to the chat history.
Conversation history. Retained threads are the closest analogue to instant messages, which several states have treated as records when the content concerns public business.

Some of it is transitory, and schedules account for that. The distinction turns on content and use, not on the tool.

Find out where the logs actually live

Settle this before writing policy, because most agencies have more than one deployment in use.

Assistants inside the productivity suite. Prompts and responses are typically stored in the agency's own tenant, tied to a user account, and reachable by the same eDiscovery, hold, and retention tooling used for email.
Standalone enterprise accounts. Conversations sit in the vendor's system under an agency-administered workspace. Export, audit logging, and retention controls vary by plan tier.
Consumer accounts. Staff who sign up individually produce conversations the agency cannot search, hold, or export. This is the hardest case, and common wherever no guidance exists.
Features inside line-of-business systems. A drafting feature in a permitting, case management, or dispatch system may log prompts in that application's database, in the provider's platform, in both, or in neither. Where software calls a model through an API, logging is whatever the developers wrote, so specify it rather than assume it.

Decide retention deliberately

Default settings are a policy decision made by someone outside the agency. Vendors variously discard conversations after a fixed window, keep them until a user deletes them, or keep them indefinitely.

Two failure modes matter. Keeping AI logs longer than the schedule allows creates discoverable material the agency was entitled to dispose of. Losing them sooner, including through an automatic vendor purge, can amount to unauthorized destruction of records.

Classify AI conversations against the existing schedule before inventing a new series. They often fall under a category the agency already has for correspondence, drafts and working papers, or transitory records. Where the fit is unclear, state archives programs have a process for amending a schedule, and several have published guidance on generative AI.

Handling a request that reaches AI conversations

Treat it as an ordinary request against an unfamiliar repository.

Scope the search. Identify which custodians used which tools during the period covered, including unapproved use.
Search what you can reach. Tenant-hosted assistants respond to existing eDiscovery tooling. Standalone workspaces usually require an administrator export.
Review for exemptions. The usual exemptions apply on their own terms: personal information, security information, criminal justice information, attorney-client material, and, where recognized, deliberative material. Uploaded files carry the exemptions they already had.
Expect inaccurate content. A generated passage may contain errors the agency never adopted. It is still a record of what the tool produced, so release it as it exists rather than correcting it in place.
Document the search. Record the systems searched, the terms used, and what was withheld and why.

What to ask vendors and put in the contract

Storage and access. Where prompts, uploads, and outputs are held, which subprocessors handle them, and which vendor personnel can read them.
Retention controls and holds. Whether retention is configurable by the agency, whether vendor-side purging can be prevented, and whether a legal hold can suspend deletion.
Export. Whether an administrator can export full conversation history, with attachments and timestamps, in a usable format, and whether that depends on the plan tier.
Training use. Written confirmation of whether agency prompts and uploads are used to train models.
Deletion and termination. What deletion means operationally, including backups, and how conversation data is exported at the end of the contract.

Where to start

Inventory actual use first, including tools nobody approved. Policy written against an imagined deployment will not match what staff are doing. Issue interim guidance on which tools are permitted and what may be entered into them. Turn on administrative logging and test an export while nothing is pending. Then classify the records with your records officer and set retention to match the schedule.

HQT works with state and local agencies on AI deployments and the systems around them, including logging, export, and retention configuration.

Frequently asked questions

Are prompts typed into an AI chatbot public records?

In most states a record is defined by content and use rather than by format, covering writings and electronically stored information prepared or received in the conduct of public business. A prompt about agency work, an uploaded agency document, and the generated response all fit that description, and generated text pasted into an email or case note is plainly a record in that system. Some material will be transitory under your schedule, but the tool does not decide the question. Confirm the specifics with your records officer and counsel, since definitions and case law vary by state.

How long do we have to keep AI chat logs?

As long as your retention schedule requires for that content, which usually means classifying the conversations against series the agency already has, such as correspondence, drafts and working papers, or transitory records. The risk runs both ways: keeping logs longer than the schedule allows creates discoverable material you were entitled to dispose of, and an automatic vendor purge can destroy records early. Check the vendor's default retention setting and align it with the approved schedule rather than accepting the default. Where no existing series fits, state archives programs have a process for amending a schedule.

An employee used a personal AI account for agency work. Is that a problem?

It creates two problems at once. The agency generally cannot search, hold, or export those conversations, and in many states records concerning public business can be subject to disclosure even when they sit in a personal account or on a personal device. The practical remedy is to provide an approved tool with administrative logging and to state in guidance that agency business stays inside it. Ask counsel how your state treats records held in personal accounts before responding to a request that reaches them.

How do we respond to a records request that asks for AI conversations?

Treat it as an ordinary request against an unfamiliar repository. Identify which staff used which tools during the period covered, search tenant-hosted assistants with your existing eDiscovery tooling, and use administrator export for standalone vendor workspaces. Apply the usual exemptions on their own terms, remembering that uploaded source files carry whatever exemptions they already had. Document the systems searched, the terms used, and what was withheld and why.

What should we require from an AI vendor about conversation logs?

Require that the agency can configure retention, apply a legal hold that suspends deletion, and export full conversation history with attachments and timestamps in a usable format. Ask whether those capabilities depend on the plan tier, since export and audit logging are sometimes missing from lower tiers. Get written answers on where the data is stored, which subprocessors handle it, whether prompts and uploads are used to train models, what deletion means including backups, and how data is returned at termination.