Research · Field tool · v1.0.0

AI Acceptable-Use Policy

A plain-language template for small governments and nonprofits.

Fifteen rules, each written two ways: how a small office follows it with a spreadsheet and a sign-off, and where this site enforces the same rule in code. Answer six questions and the template trims itself to what you need.

Version
1.0.0
Published
2026-09-23
Review by
2027-03-23
License
CC BY 4.0

This is a starting template, not legal advice. Laws on public records, privacy and procurement differ by state and by funder. Have counsel review it before you adopt it.

01 Which rules do you need?

Every question starts on, so you see the full template first. Switch off what does not describe your organization.

15 of 15 rules apply. 4 of them apply to every organization.

Nothing you type here leaves your browser.

02 The rules

  1. Use only approved AI tools

    Keep one list of the AI tools your organization has approved and what each one may be used for. A new tool goes on the list before anyone uses it for work, not after.

    Why

    Tools nobody approved are tools nobody checked. You cannot protect information that flows into a service you do not know exists.

    Without engineers

    A shared spreadsheet with four columns: tool, approved uses, who approved it, date. Review it every quarter and remove what nobody uses.

    How this site enforces it

    This site keeps a ledger of every outside service it holds a credential for. A test fails when code starts using a credential the ledger has never heard of.

  2. Keep sensitive information out of AI tools

    Do not paste personal information, case details, health, financial or law-enforcement records into an AI tool unless that tool is approved for that kind of information.

    Why

    What goes into a chat window may be stored, reviewed by the vendor or used to train a model. Once it is out, it is out.

    Without engineers

    Add a column to your approved-tools list: which kinds of information each tool may see. When in doubt, remove names and identifying details first.

    How this site enforces it

    This site keeps personal details out of its public code by test: one test fails if contact details appear in the repository, and another fails if a public page imports the module that handles a client's records.

  3. A person approves before AI-drafted work goes out

    Anything an AI drafted that will be published, sent to the public or sent to another agency is read and approved by a named person first.

    Why

    AI output can be wrong, off-tone or confidently invented. The approval step is where a person takes responsibility for it.

    Without engineers

    Pick one approver per channel. Nothing AI-drafted is sent until the approver replies "approved" in writing, and you keep that reply.

    How this site enforces it

    On this site a drafted social post can only reach "needs review", never "approved", without a person. Public-records requests are drafted by software but sent only after approval; automatic sending is off by default.

  4. Check drafts against a written list of things you never publish

    Keep a short written list of topics and claims your organization never publishes. Check every AI-drafted piece against it before it goes out.

    Why

    An AI does not know your organization's lines unless you write them down. A list turns judgment calls into a check anyone can run.

    Without engineers

    Five to ten lines on one page. The approver in the previous rule checks each draft against it and says which lines were checked.

    How this site enforces it

    This site runs a pre-send check against its own list in its outbound paths; a newsletter issue that trips it never sends. The list itself is kept in the code and tested for consistency.

  5. Keep public-facing AI output nonpartisan

    AI-assisted content that speaks for a public body stays neutral on candidates, parties and ballot questions, and critiques systems rather than people.

    Why

    Public bodies owe residents neutrality, and many are bound to it by law. An AI draft can slip a political frame in without anyone noticing.

    Without engineers

    Add a nonpartisan line to your never-publish list, and have your approver check each draft for it.

    How this site enforces it

    This site's list includes a nonpartisan rule, and the same rule is written into the instructions for the AI that critiques drafts before a person sees them.

  6. Say when AI is involved, and never fake a person

    Tell people when AI wrote, summarized or answered something they rely on. Never present an AI voice, image or message as a real person.

    Why

    Trust in a public body or a nonprofit is the whole asset. Being caught passing AI off as a person spends it all at once.

    Without engineers

    One standard sentence on AI-assisted material, for example "Drafted with AI assistance and reviewed by staff." No AI voices or avatars that imply a real employee.

    How this site enforces it

    On this site a computer reading a page aloud is labeled as a device voice, and in its language course a test keeps a synthetic voice from standing in for a native speaker.

  7. Check facts and sources before you rely on AI output

    Before AI output is used in a decision, a publication or a reply to the public, a person checks its facts against the original source.

    Why

    AI tools produce fluent text that can be wrong. A citation you did not open is not a citation.

    Without engineers

    Require a source link for every factual claim in AI-assisted work, and have the approver open at least the ones that matter.

    How this site enforces it

    This site's assistant attaches the sources it drew on to every answer, its civic assistant only cites approved official sites, and research figures are recomputed by test on every change.

  8. Put a spending cap on paid AI

    Every paid AI service has a monthly budget, a warning before it is reached, and a stop when it is.

    Why

    Usage-based billing has no natural ceiling. A runaway automation or a busy public tool can spend a month's budget in a day.

    Without engineers

    Set the vendor's own billing limit and alert, name who receives the alert, and write down what gets switched off at the limit.

    How this site enforces it

    This site tracks model spend month to date against a ceiling, warns at 80 percent, and pauses non-essential AI work when the ceiling is reached.

  9. Limit how much the public can use your AI tool

    A public AI tool has a per-person limit and a total budget, so one visitor cannot use it all or run up the bill.

    Why

    Anything free and public will be used at volume, sometimes by scripts. Limits keep the tool available and the cost predictable.

    Without engineers

    Use the limit settings your vendor offers, and set a separate, lower budget for the public tool than for staff use.

    How this site enforces it

    This site's public assistant allows a set number of questions per minute, hour and day for each visitor, and pauses for anonymous visitors when the monthly budget is spent.

  10. Keep keys and passwords out of documents, chats and browsers

    API keys and passwords for AI services live in one secure store. They are never pasted into documents, chat messages, AI prompts or shared screens.

    Why

    A leaked key is a bill and a breach. Keys pasted into chats and prompts end up in logs and histories you do not control.

    Without engineers

    Use a password manager with a shared vault for service keys. Rotate a key whenever someone who had it leaves.

    How this site enforces it

    This site's setup flows read a credential on the server and write it straight to the store, so it never passes through a browser, and an audit checks that every credential in use is accounted for.

  11. Give AI tools the least access they need

    When an AI tool connects to email, files or other systems, grant only the permissions its approved use requires, and review them when the use changes.

    Why

    An AI tool with broad access can read, send or delete far more than anyone intended, and a compromised account inherits all of it.

    Without engineers

    Before approving a connection, write down which permissions it asks for and why each one is needed. Refuse any you cannot justify.

    How this site enforces it

    This site requests only the permissions a feature uses, and a test fails if a connection ever asks for the forbidden ones.

  12. Keep an off switch anyone on duty can use

    Every automated AI process that sends, posts or acts on your behalf can be stopped quickly by a named person, without waiting for a developer.

    Why

    When something goes wrong at night or on a weekend, the question is how fast it stops, not who built it.

    Without engineers

    For each automation, write one line: how to pause it and who is allowed to. Test the off switch once before relying on it.

    How this site enforces it

    This site's notifications ship switched off and are turned on deliberately, and automatic sending of records requests stays off unless someone turns it on.

  13. Keep a record of what AI did

    Keep a simple record of which AI tool was used for what, when, and by whom, for anything that leaves the organization or informs a decision.

    Why

    Public records requests, audits and complaints ask what happened. "Someone used a chatbot" is not an answer.

    Without engineers

    Note the tool and the reviewer in the file or ticket where the work lives. For recurring automations, keep the vendor's usage log.

    How this site enforces it

    This site logs every paid AI call with what it was for and what it cost, and reads its monthly spend from that log.

  14. Watch for automations that quietly stop working

    Check on a schedule that each AI automation is still running and still doing its job, not only that it has not reported an error.

    Why

    The most expensive failure is the one nobody notices: a reminder that stops sending, a report that stops arriving.

    Without engineers

    A monthly checklist: each automation, when it last ran, and whether its output arrived. Anything that did not gets a person assigned.

    How this site enforces it

    This site has a dashboard that classifies every scheduled job as fired, failing or silent, and a test keeps that list in step with the jobs that actually exist.

  15. Name an owner and review this policy on a schedule

    One named person owns this policy. It is reviewed at least every six months and whenever you add a tool, a use or a kind of data.

    Why

    AI tools change monthly. A policy nobody owns describes last year's tools.

    Without engineers

    Put the owner's name and the next review date at the top of the policy, and put the review on a calendar.

    How this site enforces it

    This page's own review date is enforced by a test that fails once the date passes, so it cannot go stale quietly.

03 How this page stays honest

The evidence is tested

Every link under “How this site enforces it” is checked on each change: the file must exist and still contain the rule it is cited for. If the enforcement disappears, the build fails before the page can claim it.

The review date is enforced

A test fails once 2027-03-23 passes. The policy cannot quietly describe last year’s tools; someone has to review it and publish a new version.

Some sources are checked, not linked

The site’s own list of things it never publishes is enforced and tested like everything else, but its source is not linked here, because the list names subjects the site does not discuss.

04 Versions

  • 1.0.0 2026-09-23 First release: fifteen rules, each with a no-code way to follow it and a link to where this site enforces it in code.

Reuse it under CC BY 4.0: adapt it, publish it, credit this page. Cite it from its research record.

Need something like this built?

I design and ship AI tools, full-stack apps, and data pipelines — end to end, to production. Tell me the problem in a sentence; I'll give you an honest read on fit within a day.

Work with me →