Research · Field tool · v1.0.2
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.
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.
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
Nothing you type here leaves your browser.
02 The rules
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.
- The credential ledger
src/lib/secrets/inventory.mjs - The test that catches an unlisted credential
src/lib/secrets/__tests__/inventory.audit.test.js
- The credential ledger
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.
- The test that keeps contact details out of the code
src/lib/apply/sprint/__tests__/fixture.test.js - The test that walls client records off from public pages
src/app/__tests__/pii-boundary.test.js
- The test that keeps contact details out of the code
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.
- The check every human gate re-runs
src/lib/social/critic.mjs - Automatic sending, off unless switched on
src/lib/foia/send.mjs
- The check every human gate re-runs
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.
- The newsletter refusing to send a flagged issue
src/lib/newsletter/send-issue.js - The same check on outgoing records requests
src/lib/foia/send.mjs - The list itself (checked in code, not shown)
- The newsletter refusing to send a flagged issue
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.
- The nonpartisan rule (checked in code, not shown)
- The rule handed to the drafting critic
src/lib/social/critic.mjs
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.
- The read-aloud button, labeled as a device voice
src/components/cta/ListenButton.jsx - The test that blocks a faked voice
press/src/components/lightning/__tests__/no-device-voice.test.js
- The read-aloud button, labeled as a device voice
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.
- Sources attached to each answer
src/lib/site/agent/loop.mjs - Only approved sources may be cited
src/lib/rules/citation.js - Published figures re-checked on every change
src/lib/research/__tests__/claims.test.js
- Sources attached to each answer
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.
- The warning threshold
src/lib/ai-spend/brake.mjs - The stop at the ceiling
src/lib/ai-spend/brake.mjs
- The warning threshold
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.
- The per-visitor limits
src/lib/rate-limit.js - The pause for anonymous visitors
src/lib/site/agent/request.mjs
- The per-visitor limits
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.
- The written rule for every connection
.claude/rules/connection-wizards.md - The audit that every credential is accounted for
src/lib/secrets/__tests__/inventory.audit.test.js
- The written rule for every connection
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.
- The exact permissions requested
src/lib/social/networks/meta-oauth.mjs - The test that forbids broader access
src/lib/social/networks/__tests__/meta-oauth.test.js
- The exact permissions requested
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 has one switch that silences every notification channel at once, each channel's starting state is a recorded owner decision, and automatic sending of records requests stays off unless someone turns it on.
- The site-wide mute: one switch for every channel
src/lib/notifications/mute.mjs - Automatic sending that does nothing unless switched on
src/lib/foia/send.mjs
- The site-wide mute: one switch for every channel
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.
- The usage log
src/lib/ai-spend/events.mjs - Spend read from the log
src/lib/ai-spend/events.mjs
- The usage log
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.
- Fired, failing or silent
src/lib/burn/model.mjs - The list kept in step with the real jobs
src/lib/burn/__tests__/catalog-drift.test.js
- Fired, failing or silent
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 carries its own review date, kept in the same file as the rules, so the date and the rules change together.
- The review date
src/lib/ai-use-policy/policy.mjs - The check that reads it
src/lib/ai-use-policy/policy.mjs
- The review date
03 How this page stays honest
The evidence is tested
Every file named under “How this site enforces it” is checked on each change: it 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 repository is private, so the files are named rather than linked.
The review date is enforced
A test fails once 2027-05-06 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 shown
The site’s own list of things it never publishes is enforced and tested like everything else, but its file is not named here, because the list names subjects the site does not discuss.
04 Versions
- 1.0.2 2026-11-06 The off-switch rule's note says what is true now: Telegram and email notifications are on by the owner's decision, browser push is still off, and one switch still silences every channel.
- 1.0.1 2026-09-23 Evidence is named by file path instead of linked, because the repository is private.
- 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.