Minutes are not a summary. They are the official, legal record of what a public body decided — the document a court reads to determine whether a vote was properly made, the thing a resident cites years later to prove what the council promised, the record a state archive keeps forever. In most states the minutes, once approved, are the meeting as far as the law is concerned. The recording is evidence; the minutes are the record.
That is exactly why AI-generated minutes are both the most useful and the most dangerous application of AI in a clerk's office. Useful, because turning hours of audio into a structured draft is genuinely hard, slow work that AI does well. Dangerous, because the same technology that writes the draft can also write things that never happened, in fluent, confident, official-sounding prose. And a plausible error in an official record is far worse than an obvious one.
The resolution is not to avoid AI. It's to build the workflow so that no AI output becomes the record until a human has checked it against reality. That principle has a name in the industry — "human in the loop" — but for clerks it's simpler than the jargon: the software drafts, a person verifies, the body approves. This post is about why that order matters, and what happens when it's skipped.
What "hallucination" actually means for a transcript
Large language models and the speech-to-text models built on them don't retrieve facts; they predict the most likely next piece of text. Most of the time the most likely text is also the true text, which is why they work at all. But when the audio is unclear, when a speaker trails off, when two people talk over each other, or when there's a stretch of silence, the model still produces its best guess — and its best guess can be a wholly fabricated sentence, delivered with the same confidence as an accurate one.
The National Institute of Standards and Technology gave this failure mode a formal name. In its Generative Artificial Intelligence Profile (NIST-AI-600-1, published July 2024), NIST lists "confabulation" — the phenomenon commonly called hallucination — among the risks unique to or exacerbated by generative AI, defining it as confidently stated content that is erroneous or false. When a federal standards body names a risk in its risk-management framework, that is not hype. It is a documented, expected property of the technology.
The most striking evidence for public bodies specifically comes from an Associated Press investigation into OpenAI's Whisper, one of the most widely used speech-to-text models. In reporting published October 26, 2024, AP journalists Garance Burke and Hilke Schellmann described researchers finding invented text — including fabricated racial commentary and imagined statements — across large samples of transcriptions. One University of Michigan researcher told AP he found hallucinations in eight of every ten audio transcriptions he inspected. The audio he was transcribing? Recordings of public meetings.
Eight in ten public-meeting transcriptions contained hallucinated text in one researcher's sample.
The figure comes from a University of Michigan researcher transcribing public-meeting audio, as reported by the Associated Press (Oct. 26, 2024). It is a sample, not a universal rate — but it shows the failure mode is common, not exotic, on exactly the kind of audio clerks work with.
Read that carefully, because it's easy to wave away. The number is not a universal error rate for all transcription tools, and modern systems have improved. The point is narrower and more important: fabrication on messy, multi-speaker, real-room audio is a common outcome, not a freak one — and a public meeting is close to the worst-case input. Gaveling, cross-talk, a resident at a podium three rows back, a member joining by phone from a car. Every one of those conditions is where a transcription model is most likely to guess.
Where the error hides
An obviously garbled transcript is not the real danger. If the minutes read "the council voted to purple the ordinance," a clerk catches it in one pass. The danger is the error that reads perfectly.
Consider the ways a confident fabrication can slip into a set of AI-drafted minutes and survive a casual read:
- A flipped vote. A model transcribes "the motion fails" as "the motion carries," or attributes an "aye" to the member who abstained. The sentence is grammatical and plausible. Only the recording disproves it.
- A misattributed statement. Diarization — the step that decides who said what — assigns a controversial remark to the wrong council member. Now the official record has a named official saying something they didn't.
- An invented condition. A summary model, asked to condense a rambling discussion into an action item, adds a specific dollar figure or a deadline that sounds like the kind of thing that was said, but wasn't.
- A dropped abstention or recusal. The one procedural detail that matters most for a conflict-of-interest challenge is the easiest for a summarizer to smooth away.
None of these look wrong on the page. That is the whole problem. Human-written minutes have errors too, of course — but a tired clerk's mistakes tend to be omissions and typos, not fluent inventions. AI's characteristic error is the opposite: confident, well-formed, and specific. It reads like the truth.
Once it's approved, it's the law
Here is the part that separates minutes from every other AI writing task. If an AI email draft is wrong, you fix it before you hit send. If AI-drafted minutes are wrong and the body approves them, the error doesn't just persist — it becomes authoritative.
Approved minutes carry legal weight. They are the evidence of what a quorum decided. They start the clock on protest and appeal periods. They're what a bond counsel relies on, what an auditor checks against, what a judge reads when a decision is challenged. An error in approved minutes is not a typo to be corrected in the next draft; correcting it usually requires a formal amendment at a subsequent public meeting, on the record, which is its own small embarrassment — and if anyone acted on the wrong version in the meantime, the amendment doesn't unwind what was done.
This is why "the AI wrote it" is not a defense a clerk ever wants to give. Responsibility for the record does not transfer to the tool. It stays with the clerk and the body, exactly as it did when the minutes were typed by hand.
The public-records trap
There's a second consequence clerks underestimate: the AI's mistakes don't necessarily disappear when you fix them. They may themselves be public records.
Public-records law generally turns on a record's content and function, not its format. A recording of a public meeting, a machine transcript of that recording, and an AI-generated draft of the minutes can each qualify as a record documenting public business. State retention schedules already treat the underlying material this way — South Carolina's general schedule for municipal records, for example, keeps voice recordings used to prepare minutes for two years before they can be destroyed, while the minutes themselves are permanent. The University of North Carolina School of Government's local-government law analysts have made the sharper point directly: an AI transcript a government creates in the course of public business can be a public record subject to disclosure even if it contains errors.
Sit with the implication. Turn on an AI note-taker for a public meeting and you may have created a discoverable record — one that a requester can ask for, one that might contain a hallucinated sentence, one you now have to account for. The convenience of an always-on transcription tool quietly enlarges your records footprint and your exposure. It's one more reason the raw AI output should be treated as a working draft inside a controlled workflow, not a casual byproduct scattered across a dozen staff laptops and cloud accounts.
A quiet governance question.
If council members or staff run a personal AI note-taker in the room, who controls that transcript, where is it stored, and is it now a record your government has to be able to produce? Deciding this before it comes up in a records request is a lot easier than after.
Governments are already grappling with this
This isn't a hypothetical debate for some distant future. Public bodies are working through AI-in-the-meeting questions right now. When two Santa Clara, California council members acknowledged using ChatGPT to look up information during public meetings, it set off a local debate about the accuracy of AI in official decision-making — reported by San José Spotlight in April 2025. That case was about AI research during a meeting rather than minutes, but it's the same underlying tension: an official process leaning on a tool that can be confidently wrong.
Meanwhile the tooling is spreading fast. Nieman Lab reported in March 2025 that local newsrooms are now using AI to transcribe and summarize public meetings across dozens of states, precisely because sitting through every school-board and zoning session by hand doesn't scale. If journalists are using AI to read your meetings, the pressure to use AI to write them is only going up. The question for clerks is not whether AI enters the minutes process. It's whether it enters it with a human checkpoint or without one.
Why "draft, then approve" is an architecture, not a slogan
The good news is that the fix is not exotic. It's a workflow design, and it maps cleanly onto how minutes have always been produced. A well-built AI minutes process has three separable stages, and a human sits at the seam:
- The machine drafts. Transcription and diarization turn the recording into text and a first-pass draft, with motions and votes structured. This is where AI earns its keep — it does in minutes what used to take a clerk hours.
- A human verifies against the source. The clerk reviews the draft with the recording available, not in isolation. Anything consequential — vote outcomes, who moved and seconded, dollar figures, conditions, abstentions — is confirmed against the audio before it goes anywhere. This is the step that catches the fluent fabrication.
- The body approves. Only after human review do the minutes go to the board for adoption, at which point they become the official record.
The design principles that make this safe are worth stating plainly, because they're the questions to ask of any AI minutes tool:
- The recording stays attached to the draft. Verification is only realistic if the clerk can jump from a line in the draft to the moment in the audio it came from. A tool that hands you a summary and throws away the timeline has made checking harder, not easier. (This is also why keeping the recording reliably available matters well beyond drafting.)
- The draft is clearly a draft until adopted. Nothing the AI produces should be able to masquerade as the official record before a human has signed off and the body has voted.
- The clerk edits freely. The human is the author of record, not a reviewer stuck rubber-stamping the machine's phrasing. The tool proposes; the clerk disposes.
- The trail is intact. You should be able to show what was drafted, what was changed, and what was approved — both for your own defensibility and because those artifacts may be records.
Put that way, the AI isn't replacing the clerk's judgment. It's removing the transcription drudgery so the clerk's judgment can be spent where it actually matters: on the handful of lines where being wrong has consequences.
What good AI-minutes practice looks like
If you're evaluating or already using AI to help produce minutes, a short discipline goes a long way:
- Never adopt an unread AI draft. Speed is the selling point, but the review is the value. Budget the time to check the consequential lines against the recording every single meeting.
- Verify the four things that carry legal weight first: the vote tallies, the mover and seconder, any numbers or dates, and any recusal or abstention. These are where a confident fabrication does the most damage.
- Know where the transcript lives. Decide, as a policy, what AI transcription tools are allowed in your meetings, who controls the output, and how long it's kept — before a records request forces the answer.
- Consider disclosing your process. There's a live debate, which we'll take up in a future post, about whether minutes should note that AI assisted in producing them. Whatever you decide, decide it deliberately rather than by default.
- Keep a human name on it. The clerk certifies the minutes. That signature means a person stands behind the record — which is exactly the accountability an official document needs and an AI can't provide.
The through-line
AI is going to make minutes faster, and for overworked clerks that is a real and worthy gift. The mistake would be to confuse "faster to draft" with "safe to publish." Those are different claims. The first is about the machine. The second is about the workflow around it.
The technology's defining weakness — confident, fluent, specific error — happens to be the single worst trait a tool could have for producing a legal record. That doesn't disqualify AI from the minutes process. It just means the human-approval step is not overhead to be optimized away. It's the part that turns a fast draft into a record you can stand behind. Draft, then approve. In that order, on purpose, every time.
Sources: Associated Press — "Researchers say an AI-powered transcription tool used in hospitals invents things no one ever said" (Burke & Schellmann, Oct. 26, 2024) · NIST — Artificial Intelligence Risk Management Framework: Generative AI Profile, NIST-AI-600-1 (July 2024) · San José Spotlight / Local News Matters — Santa Clara AI-in-meetings debate (Apr. 18, 2025) · Nieman Lab — "Local newsrooms are using AI to listen in on public meetings" (Mar. 2025) · UNC School of Government, Coates' Canons — analysis of AI and the Public Records Act · South Carolina Dept. of Archives & History — General Records Retention Schedule for Municipal Records. This article is general information, not legal advice; consult your attorney or state archives for your jurisdiction's rules.