Catching AI Hallucinations in Government Documents Before Council

The draft looks finished. The background section walks through the project history in order, the fiscal impact paragraph names a budget account, and the resolution's recitals refer to an earlier council action that set the policy in motion. Then someone pulls the minutes for that earlier meeting and finds no such action. No one on staff invented it. The drafting tool did, and it wrote the false sentence in exactly the same confident tone it used for the true ones.

That is what people mean by an AI hallucination: output that reads as factual but is not supported by any real source. For most people using AI tools, a hallucination is an annoyance. For a city, it can end up in a council packet, in a resolution the council adopts, or in minutes that become the permanent record. Council members, reporters and residents read those documents closely, and some of them will check.

This guide is for the staff who prepare and review those documents. It covers:

• What a hallucination is, in plain terms, and why it happens
• The forms hallucinations take in staff reports, resolutions, minutes and briefing memos
• Why grounding and citations reduce the risk, and where risk remains
• A review checklist for any AI-assisted draft
• A verification routine for each common document type
• What to do if an error reaches council anyway

Human-written drafts contain errors too: a paragraph copied from last year's report, a transposed number, a stale contract amount. Most of the routine below catches those as well. The difference with AI output is that its errors lack the usual warning signs. A rushed staff member's mistakes often look rushed. A model's mistakes look polished.

What a hallucination is, and why it happens

A large language model, the technology behind most AI writing tools, produces text by predicting what words are likely to come next given everything before them. It learned those patterns from an enormous amount of text. It does not contain a database of facts it can look up, and it has no built-in sense of whether a sentence is true.

Most of the time, likely text and accurate text overlap. Ask for a general paragraph explaining what a consent agenda is, and the likely answer is also a reasonable one. The trouble starts when the model is asked for something specific it does not actually have: the date your council approved a lease, the section of your municipal code that governs sidewalk vending, the amount of last year's grant award. In those cases the model may still produce something with the shape of a correct answer. A resolution number in your city's format. A dollar figure of about the right size. A code section that resembles your real numbering.

It helps to separate hallucinations from a few related problems, because each needs a slightly different check:

• Fabrication: the claim has no source at all. An invented prior action, a nonexistent code section, a quote no one said.
• Misreading: the source is real, but the draft states it wrong. The contract says three years with two optional renewals, and the draft says five years.
• Stale information: the source is real and correctly read, but it has been superseded. The draft cites an old fee schedule after the council adopted a new one.
• Blending: details from two real sources are merged. The vendor from one contract, the amount from another.

All four produce the same result for your council: a document that says something untrue. Your review needs to catch all of them, not just the obvious inventions.

What hallucinations look like in city documents

Hallucinations in government documents cluster around a few predictable spots. Knowing where to look is half the work.

Invented legal authority

Citations to state statutes, municipal code sections, attorney general opinions or court cases are among the highest-risk content in any draft. A model knows what a statute citation looks like and can generate one that is formatted correctly and refers to nothing, or refers to a real section about a different subject. Be especially alert to "pursuant to" or "as authorized by" followed by a section number. Law also differs from state to state, so a model drawing on general patterns may describe a rule that applies somewhere else.

Phantom council history

Recitals and background sections often summarize what the council did before: "On March 12, the City Council directed staff to..." These sentences are easy for a model to produce and hard for a reviewer to doubt, because they sound like institutional memory. Every reference to a prior action needs to match an actual agenda item, motion, resolution or ordinance.

Numbers that look right

Dollar amounts, vote tallies, dates, acreage and unit counts are where blending and invention show up most. A fiscal impact paragraph might carry a figure that is close to a real budget line but not the one that applies. A vote count in draft minutes might total the wrong number of members. Plausible numbers are more dangerous than wildly wrong ones because no one stops to question them.

Confident statements of requirement

Watch for "state law requires," "the city is obligated to," "best practice is," and "typically." These phrases turn a general pattern into a claim about your city's obligations. Sometimes the claim is right. Sometimes it describes a requirement your state does not have, or a deadline that is different where you are.

Names, titles and roles

Drafts may attach the wrong title to a person, list a council member who has left office, name the wrong lead department, or use a vendor's former name. In minutes, a model working from a transcript or notes may attribute a comment to the wrong speaker, a serious error in a document that becomes the official record.

Boilerplate that does not fit

Resolutions follow conventions, and a model knows them well. That fluency can produce "whereas" clauses stating findings the council never made, or a standard sentence about public hearing notice that does not apply to this item. Each recital should be supported by something in the record for this item.

Omissions that look like completeness

A tidy draft can leave out the inconvenient facts: the prior recommendation that went the other way, the pending litigation, the condition attached to the grant. Nothing in the draft is false, but the council gets a misleading picture. This is the hardest category to catch, because you are looking for what is not on the page.

Why the stakes are higher in a public document

Most organizations can fix a bad internal memo with a follow-up email. Cities have fewer quiet options.

Staff reports in a published packet are generally public records once distributed. Resolutions, once adopted, carry legal effect and can be cited in future actions, contracts and disputes. Minutes, once approved, are the official account of what a public body did. Correcting any of these usually requires a public step: an amended report, a corrective action at a later meeting, or a formal correction under your rules of procedure. The details depend on your state law, your charter and your council's rules, so confirm the process with your city attorney or clerk before you need it.

There is also a trust cost. A council member who finds one invented citation in a packet will reasonably wonder what else is wrong, and may start reading every staff report with suspicion. A resident or reporter who finds one may write about it. Neither outcome depends on how the error got there. The council will not care whether a person or a tool wrote the sentence. Staff put their names on the document.

That is the right frame for review. AI tools change how a first draft gets produced. They do not change who is accountable for the final document.

Why grounding and citations reduce risk, and where risk remains

Not all AI drafting works the same way, and the difference affects how much review you need.

General tools versus grounded tools

A general-purpose chatbot answers from patterns in its training data. Ask it to draft a staff report on your downtown parking program and it knows nothing about your program unless you paste information in. It may fill the gaps with plausible inventions.

A grounded tool works differently. Before drafting, it retrieves relevant documents from a defined collection, such as your adopted resolutions, minutes, prior staff reports and contracts, and drafts from that material. Well-designed grounded tools also show where each claim came from: the document, the page and the meeting. That changes the reviewer's job. Instead of asking "is this true?" with no starting point, you ask "does this source say this?" and can check quickly.

Grounding reduces hallucination risk because the model has real text to work from. Citations reduce the cost of verification because they point you to the evidence. Together they make review faster and more reliable. They do not make review optional.

What can still go wrong with a grounded draft

• The citation is real, but the claim overstates it. The cited page discusses the topic but does not say what the draft says. Read the cited passage, not just the citation.
• The tool found an older document and missed the newer one. If a resolution was amended or repealed, a draft may cite the original. Confirm each source is current.
• The source itself was wrong. A prior staff report may contain an error, and a grounded draft will repeat it faithfully. Grounding inherits the quality of your records.
• Uncited sentences slip in. Treat any factual claim without a citation as unverified until you check it.
• The records collection is incomplete. If older minutes were never digitized or contracts live in a separate system, the tool cannot draw on them and may not tell you they are missing.

When you evaluate any tool, ask what it does when it cannot find a source. A good answer is that it says so plainly or leaves a visible gap rather than producing text anyway. Ask to see this in a demo using your own records, and get the answer in writing. Capabilities change between releases, so confirm current behavior rather than relying on what you heard earlier.

A review checklist for any AI-assisted draft

Use this checklist on every AI-assisted document before it leaves staff hands. It works for human drafts too.

Then ask:

• Does every factual sentence have a source you personally checked?
• Did you read each cited passage, or only confirm the citation exists?
• Is every cited document the current version?
• Does the draft make a recommendation, finding or judgment the department did not make?
• Would the department head sign this as written?

Keep a short list of phrases that should always slow you down: "previously approved," "pursuant to," "state law requires," "consistent with past practice," and any sentence that begins "Historically." They are not always wrong. They are always worth checking.

A verification routine for each document type

Different documents fail in different ways. A routine tuned to the document type catches more with less effort.

Staff reports

Staff reports carry the most varied content, so they need a structured pass.

1. Check the recommendation first. Confirm it matches what the department actually recommends. A model may soften, strengthen or reword it in ways that change its meaning.
2. Trace the background section. Every reference to a prior council action, study session or directive should map to a specific meeting and item. Pull the minutes.
3. Verify the fiscal impact with finance. Confirm the amount, fund, fiscal year and whether the money is budgeted. Council members ask about this section most.
4. Check every legal reference. Send statute and code citations to the city attorney's office when the report relies on them.
5. Compare against attachments. Read any summary of a contract, agreement or study against the attachment itself. Summaries drift.
6. Read for omissions. Ask the department lead what a skeptical council member would want to know that the report does not say.

Resolutions

Resolutions are short, formal and legally operative, so precision matters more than readability.

1. Confirm the operative language. The "now, therefore, be it resolved" sections must state exactly what the council is being asked to do. Compare them word by word with the staff recommendation.
2. Check each recital against the record. Every "whereas" clause should be supported by the staff report, attachments or prior actions. Delete any recital you cannot support.
3. Verify references to other resolutions and ordinances. Confirm numbers, titles and adoption dates against your official records, and confirm none has been amended or repealed.
4. Check names, amounts and terms. Contracting parties, dollar figures, terms and effective dates must match the underlying agreement exactly.
5. Route to the city attorney under your normal approval process. AI drafting is never a reason to skip legal review.

Draft minutes

When a tool produces draft minutes from a recording, transcript or notes, the risk shifts to attribution and actions.

1. Verify every motion, second and vote against the recording: who moved, who seconded, how each member voted, and the outcome.
2. Check attendance. Confirm who was present, who arrived late and who left early, and that vote tallies are consistent with attendance at that point in the meeting.
3. Check attribution of any summarized comments.
4. Look for invented detail. Models sometimes add a plausible summary of discussion that did not occur, or tidy up an action in a way that changes it. Minutes reflect what happened, not what usually happens.
5. Apply your minutes style, whether action, summary or verbatim.

Agenda item titles and summaries

Short text gets less scrutiny, yet the agenda title is what the public sees.

1. Confirm the title accurately describes the action the council may take. In many states an inadequate agenda description can create open meetings problems. The standard varies by state, so ask your city attorney what yours requires.
2. Match the summary to the staff report. A summary should not introduce facts the report does not contain.
3. Check consistency. Title, summary, staff report and resolution should describe the same action in compatible terms.

Council briefing memos

Briefings prepared for council members ahead of a meeting are often less formal, which makes them easy to under-review.

1. Check every historical claim. Briefings lean on "the council last considered this in..." statements.
2. Separate facts from framing. The memo should present information neutrally and not imply positions the council has not taken.
3. Assume it may be a public record. Depending on your state law, a memo sent to council members may be subject to disclosure.

Building verification into your workflow

A checklist only works if it fits how your office actually operates.

Assign the review, do not assume it

Decide who verifies what. The author verifies their department's content. Finance verifies fiscal impact. The city attorney's office verifies legal citations and resolution language. The clerk checks consistency across the packet and references to prior actions. Write these assignments down so they survive staff turnover.

Separate drafting from review where you can

The person who generated the draft is primed to see what they expected. For resolutions and anything with legal or fiscal weight, have a second person verify. In a small office that might be the department head, the manager or the clerk.

Budget time for review

AI drafting can make a first draft faster. It does not make verification instant. If your agenda calendar assumes drafting time savings carry straight through to the packet deadline, review gets squeezed. Build verification time into your submission timeline.

Keep a simple error log

When a reviewer catches an error, note the document type, what was wrong and how it was caught. Over time you will see your own patterns, such as fiscal figures drifting or original resolutions cited instead of amendments. Use that to sharpen the checklist and to raise specific issues with your vendor. Keep the log factual and blame-free so staff report errors instead of hiding them.

Watch for review fatigue

The better a tool performs, the more tempting it becomes to skim. Long runs of correct drafts train reviewers to stop checking. Rotate reviewers, keep the checklist visible, and treat the high-risk rows in the table as always-check items no matter how reliable the tool has been.

If an error reaches council anyway

Even good routines miss things occasionally. Have a plan before it happens.

• Before the meeting: Issue a corrected document or supplemental memo through your normal process, and make clear to council and the public what changed.
• At the meeting: Staff can note the correction on the record when the item is introduced. If the error affects the action itself, the council may need to amend the motion or continue the item. Your city attorney can advise.
• After adoption: Correcting an adopted resolution or approved minutes usually requires formal action under your rules and state law. Never edit an adopted document quietly. Confirm the proper process with your city attorney and clerk.
• Afterward: Log the error, identify which check missed it, and adjust the routine.

Be straightforward with council. Explaining what was wrong, how it was caught and what you have changed preserves more trust than a vague correction.

Next steps

You do not need to overhaul your process to start catching hallucinations. This week:

1. Post the checklist table where staff who prepare reports and resolutions will see it.
2. Assign verification owners for fiscal figures, legal citations and prior council actions, in writing.
3. Add review time to your agenda submission calendar for AI-assisted items.
4. Start an error log, even a simple spreadsheet.
5. Test any tool you use or are considering with your own records. Ask about a topic whose history you know well, check whether the citations hold up, and see what the tool does when it has nothing to cite.

If you are comparing grounded drafting tools, Govera is one option to consider: it drafts staff reports, resolutions and agenda items from your city's own records with citations to the source document, page and meeting, and staff review every output. Whatever tool you choose, a consistent verification routine is what keeps errors out of the packet.