Are AI Chats Public Records? What City Clerks Need to Know

Somewhere in your city, a staff member has already asked a chatbot to tighten a memo, summarize a long packet, or draft a reply to a resident. Maybe it was on a city laptop with an approved tool. Maybe it was on a phone, in a personal account, late in the evening. Either way, it raises a question most cities have not answered yet: is that conversation a public record?

The honest answer is that it depends, and most of what it depends on is already familiar to you. The answer turns on your state's public records law, how that law defines a record, your retention schedule, and how the tool stores what was typed into it. AI does not create a new body of records law. It creates a new kind of material that has to be sorted under the rules you already have. The sorting is harder than it looks, because a single chat thread can contain a draft, a reference lookup, a pasted confidential document and a personal question, one after another.

This guide is for clerks, records managers and the staff who support them. It covers:

• Why prompts and outputs may count as records
• What an AI interaction actually contains
• The cases that are genuinely hard
• What to ask your city attorney
• Practical steps on tool selection, logging, retention and staff guidance

It does not cite specific statutes. They differ by state and they change. Treat everything here as a set of questions to bring to your city attorney and your state clerk association, not as a legal conclusion.

Why the content matters more than the format

Most public records laws were written before chat tools existed. Many of them were written broadly on purpose. In a large number of states, the definition of a public record focuses on what the material is and why it exists, not on the form it takes. Definitions commonly ask whether something was created or received in the course of public business, and many say outright that physical form or format does not matter.

Cities have been through this before. When email arrived, the question was whether a message was a record. Then came text messages, then social media posts, then collaboration tools. In each round, the deciding issue was usually less about the technology and more about whether public business was being conducted in it.

Apply that logic to AI tools and the outline of an answer appears:

• A staff member pastes facts about a rezoning application into a tool and asks for a first draft of the staff report. The prompt and the output may be working material for an official document, much like a draft in a word processor.
• A staff member uses a tool to draft a response to a resident complaint and sends it with light edits. The exchange may document how the city handled the complaint.
• A staff member asks a tool to fix the grammar in one sentence of a birthday card for a coworker. Most attorneys would likely see little or no public business there.

The middle example is where many cities will spend their time. It looks a lot like ordinary drafting, and ordinary drafting is already something your laws and schedules address.

Three separate questions

When people ask "are AI chats public records," they are often asking three different things at once. Keep them apart, because the answers can differ:

1. Is it a record? Does the material meet your state's definition of a public record?
2. Is it disclosable? If it is a record, does an exemption apply? Some states exempt certain drafts, preliminary notes, attorney communications or personnel matters from disclosure. Being a record and being disclosable are not the same thing.
3. How long must it be kept? If it is a record, which retention series does it fall under, and can it be destroyed after a set period or once it is superseded?

A chat can be a record, be exempt from disclosure, and still need to be kept for a defined period. Another chat might be a record that is disclosable and eligible for destruction after a short time. Mixing up the three questions leads to bad decisions in both directions: deleting things that should have been kept, or treating every quick query as a permanent file.

What an AI interaction actually contains

Before anyone can classify AI material, they need a clear picture of what exists. A single interaction can leave several distinct pieces behind, and they may not all have the same status.

• The prompt. What the staff member typed. It may include facts, names, questions or instructions that reveal how the city was approaching an issue.
• Pasted or uploaded material. Documents, spreadsheets, emails or excerpts the staff member gave the tool. These are often already records somewhere else in your system. The copy inside the tool may be a duplicate, or it may be the only place a particular combination of material exists.
• The output. What the tool produced. It may be discarded, edited heavily, or used nearly word for word.
• Follow-up turns. Many interactions involve several rounds of revision. Each round can change the substance.
• The final work product. The memo, report, letter or post that came out of the process. In most cases this is clearly a record, in the same way it would be without AI involved.
• Metadata. Who used the tool, when, under which account, and sometimes which settings or model were involved.
• Saved instructions and templates. Some tools let users or administrators store standing instructions or reusable prompts. If those shape official documents, they may matter too.
• Vendor-side logs. Separate from what the user sees, the vendor may keep its own records of activity for security, troubleshooting or other purposes, under its own terms.

The final work product is rarely the hard part. The real questions are about the trail behind it: the prompts, the intermediate outputs, and the copies of source material sitting inside a third party's system.

The hard cases

Some situations come up often enough that it helps to name them. For each one, the right answer depends on your state's law and your attorney's reading of it.

Personal accounts and personal devices

This is the case that worries attorneys most. Staff may sign up for a free tool with a personal email address and use it for work because it is convenient. Many states have already dealt with public business conducted through personal email or text messages, and the same reasoning may reach AI tools.

The practical problem is as serious as the legal one. If the material sits in a personal account, the city may not know it exists, cannot search it when a request arrives, cannot apply a retention schedule to it, and cannot place it on hold. The staff member may end up responsible for searching their own account. Most cities will be on firmer ground with a clear rule that public business happens only in city-approved tools under city-managed accounts.

Drafts and working notes

Many cities already treat drafts in one of a few ways: as records kept until the final version is adopted, as transitory material with a short retention period, or as records that are kept but may be exempt from disclosure. AI drafts can often follow whatever approach you already use for drafts, if your attorney agrees. The complication is volume. A tool can produce many versions quickly, and staff may not think of each regenerated output as a draft.

Transitory and reference use

Many retention schedules include a category for transitory records, meaning material with short-term value that can be destroyed once its purpose is served. Asking a tool to define a term, suggest a synonym, or explain a general concept may fit there. Check your schedule's actual wording, though. Transitory categories are usually defined narrowly, and staff tend to stretch them.

AI features inside software you already use

Word processors, email clients, meeting platforms and agenda systems increasingly include built-in AI features, sometimes turned on by default. These are easy to overlook because nobody "adopted an AI tool." Ask IT for an inventory of AI features in licensed software and find out where their prompts and outputs are stored. Some may save material inside the document itself. Others may store it somewhere separate or not at all.

Meeting transcription and summaries

Automated transcripts and summaries of meetings deserve their own conversation with your attorney. They can interact with open meetings rules, with your official minutes, and with recordings you may already be required to keep. A machine summary that circulates internally and differs from the adopted minutes can cause confusion about which version reflects what happened. Decide in advance which document is official and how the others are treated.

Deleted chats and vendor retention

A user deleting a chat does not necessarily mean the material is gone. Depending on the vendor's terms, it may remain in backups or logs for some period. The reverse is also possible: a vendor may delete material on its own schedule, regardless of what your retention schedule requires. Neither situation is under your control unless your contract addresses it.

Questions to take to your city attorney

Your city attorney is the right person to interpret your state's law. You can make that conversation far more productive by coming with specific questions and real examples. Bring two or three anonymized chats from actual use, if you can, so the discussion is grounded in what staff are really doing.

Questions worth asking:

1. Under our state's definition, are prompts and outputs from AI tools records when they relate to public business? Does it matter whether the output was used?
2. How should we treat AI drafts? Do they follow our existing approach to drafts, or is something different needed?
3. Which of our current retention series apply to AI material, and is any of it eligible for the transitory category?
4. Which disclosure exemptions might apply to AI material, such as those for drafts, deliberative material or attorney communications?
5. How should we handle public business done in personal AI accounts or on personal devices, both going forward and for anything that already happened?
6. When a records request comes in, how far do we need to search? Does that include AI tools, built-in AI features, and vendor-held logs?
7. How should legal holds apply to AI tools, and how will we enforce them technically?
8. Is there any guidance from our state archives, records agency, attorney general or state clerk association on AI material?
9. Are there confidentiality concerns with pasting certain records into third-party tools at all, such as personnel, health, law enforcement or attorney-client material?
10. What contract terms should we require from vendors on data ownership, retention, deletion and access?

Ask for the answers in writing, even a short memo. Policy decisions about records are much easier to defend when they rest on documented legal advice. Include your records manager and IT lead in the meeting, because many of the answers will turn on technical facts they can supply.

Also check with your state clerk association. Some associations and state archives have started discussing AI material, and others may not have addressed it yet. Either way, you will learn where your state stands and whether neighboring cities are taking a consistent approach.

Choosing tools with records in mind

Records obligations are much easier to meet when you choose tools that support them. Too often, a tool is selected for what it can draft, and records questions come up only after the first request arrives. Put records questions into the evaluation from the start.

The table below lists questions to ask any vendor, whether the tool is a general chatbot, a specialized government product, or an AI feature added to software you already license.

Some practical points on using this table:

• Get answers in writing. Sales conversations are not commitments. Ask for written responses and, where it matters, put terms in the contract.
• Confirm in a demo. Ask the vendor to show you the export, the administrator search, and the hold function, not just describe them.
• Recheck on renewal. Features, settings and terms change. A capability that existed at signing may work differently a year later.
• Treat consumer and enterprise versions as different products. The free version of a tool and the organizational version of the same tool can have very different storage and administration options. Confirm which one staff are actually using.

None of this requires choosing one type of tool over another. It requires knowing what each tool does with your material before staff start putting public business into it.

Logging and capture: deciding what you keep

Once a tool is approved, you still need to decide how AI material will be captured. There is no single right approach. Each option has tradeoffs, and the choice should follow your attorney's guidance on what counts as a record in your state.

Option 1: Rely on the tool's own history

The simplest approach is to let the tool keep history under city-managed accounts and treat that as the record. This works only if the tool lets administrators search, export and hold material, and if its retention settings match your schedule. If any of those pieces is missing, this option leaves gaps.

Option 2: Centralized organizational logging

Some organizational versions of tools offer central logging or export to an archive. This gives the city one place to search and apply retention. It also means you are keeping a large volume of material, much of which may be transitory, and some of which may contain sensitive information staff should not have pasted in. Central capture makes searching easier but can make requests larger.

Option 3: Capture at the point of use

Staff save relevant AI material into your existing records system as part of their work. For example, a prompt and output that shaped a staff report get saved with the report's working file. This puts AI material alongside the records it relates to. It also depends on consistent staff behavior, which is hard to sustain.

Option 4: Keep only what becomes a work product

Some cities may decide, with legal advice, that only AI outputs incorporated into a final document need to be kept, and treat the rest as transitory. This keeps volume down. It also carries the most legal risk if your state takes a broader view, so do not choose it without a clear written opinion from your attorney.

Principles that apply whichever option you choose

• Decide before you configure. Settings such as automatic deletion or history being off should reflect a records decision, not a default someone accepted during setup.
• Do not destroy outside your schedule. In many states, destroying records outside an approved schedule, or after a request or hold, creates serious problems. Confirm how this applies in yours.
• Write down what you chose and why. If someone later asks how the city handled AI material, a short document explaining your approach and the legal advice behind it is far better than a reconstruction from memory.
• Keep logs of activity, not just content. Knowing who used a tool, when, and on what is useful even if you do not keep every prompt.

Applying your retention schedule to AI material

Many cities will find that most AI material fits into series they already have. The work is mapping it clearly enough that staff apply the mapping consistently. Your schedule's wording, and any guidance from your state archives or records agency, will decide the specifics.

A typical mapping exercise looks something like this. These are categories to discuss with your records manager and attorney, not conclusions:

A few cautions:

• Holds override schedules. When litigation is anticipated or a request is pending, the normal destruction timeline generally stops. Make sure AI tools are part of your hold process.
• Do not create a new series without checking. Some states require approval for new retention categories. Your state archives or records agency can tell you whether existing series cover AI material or whether something new is needed.
• Revisit the mapping. As tools change and your state issues guidance, the mapping may need updates. Put a review date on it.

Guidance staff can actually follow

Even the best legal advice fails if staff cannot apply it at their desks. Staff guidance on AI and records should be short, concrete and tied to what they actually do. A one-page reference is often more effective than a longer policy section, provided it points back to the full policy.

Rules that cover most situations

• Use only approved tools under city accounts for city work. No personal accounts, no unapproved apps, no personal devices unless the city has specifically approved them.
• Assume what you type may be read later. Write prompts the way you would write an email that might be produced in a records request.
• Do not paste confidential material unless the tool is approved for that type of information. When in doubt, ask first.
• Save AI material that shaped official work using whatever capture method the city has chosen.
• Do not delete history yourself unless the city's guidance says you may for that category of material.
• Tell the clerk or records manager if a records request or hold might involve AI material you have used.
• Verify before you use. Any output that goes into an official document should be checked against source records. Your name is on the result, not the tool's.

A quick self-check before using a tool for work

Encourage staff to run through these questions:

• Is this an approved tool, and am I signed in with my city account?
• Does this task involve public business?
• Am I about to paste anything confidential or restricted?
• If this output ends up in an official document, do I know how to save the trail?
• If someone requested this conversation next month, would I be comfortable with how it reads?

Make it easy to ask

Staff will run into situations the guidance does not cover. Name a contact, usually the clerk or records manager, for AI records questions, and make it clear that asking is expected. Questions that come in are also useful signals of where the guidance needs work.

Next steps

You do not need to solve every AI records question at once. A sensible sequence for the next few months:

1. Find out what is in use. Ask departments which AI tools and built-in features staff use for work, including personal accounts. Make it safe for people to answer honestly.
2. Meet with your city attorney. Bring the questions above and a few real examples, and ask for written guidance.
3. Check with your state clerk association and state archives to learn whether any guidance on AI material exists in your state.
4. Run the vendor questions for every approved tool, get answers in writing, and confirm the key functions in a demo.
5. Choose a capture approach and map AI material to your retention schedule, with your records manager and attorney.
6. Publish a one-page staff guide and name a contact for questions.
7. Set a review date for the guidance and the mapping, since tools and state guidance will keep changing.

If you are evaluating tools as part of this work, including Govera, which keeps each city's data in its own database schema and audit logs every action, put the same records questions to every vendor and compare the written answers side by side.