AI Policy for Museums: Checklist for Labels, Audio Scripts and Visitors

A small museum's AI policy should cover four things: a named curator who checks every AI-assisted label and audio script for accuracy, a rule for telling visitors when AI helped, a rule on image and text rights, and a ban on pasting visitor or donor data into unapproved tools. One page is enough.

Most museums that write a policy get stuck by trying to cover every tool. Do the reverse: write down the five or six jobs people already do with AI, and make a rule for each. A policy that matches real behaviour gets followed. A policy that describes an ideal museum gets filed and ignored.

Follow me on Instagram@sagnikteaches

Why labels and audio scripts need their own rules

A label is a claim made in the museum's voice. Visitors read it as fact, and it often sits next to an object that cannot be checked on the spot. AI drafting tools produce fluent text that sounds equally confident whether a date is right or invented. That is the whole risk in one sentence: the prose is smooth, so errors do not announce themselves.

Connect on LinkedInSagnik Bhattacharya

Audio scripts add a second problem. They are read aloud, often by a synthetic or hired voice, and they get recorded and reused for years. A wrong place name in a printed label can be corrected with a new sticker. A wrong name baked into a hundred-minute audio tour means re-recording, re-approving and re-uploading to whatever hosting service the guide runs on.

Subscribe on YouTube@sagnikteaches

Broad ethics papers exist, including UNESCO's Recommendation on the Ethics of Artificial Intelligence, and professional bodies publish their own guidance. They are worth reading once, because they give you vocabulary for transparency, accountability and human oversight. They do not tell a three-person museum who checks a label on a Tuesday afternoon. That is what your page does.

The one-page policy, section by section

Here is a filled-in version for an illustrative museum with two paid staff, a curator on three days a week and about thirty volunteers. Adapt the wording; keep the structure. Each section is short enough to read aloud at a trustees' meeting.

AI USE POLICY - [Museum name]
Owner: [name and role]. Reviewed: [date]. Next review: [date].

1. Approved uses
   Staff and volunteers may use [named tool] to: draft first versions of
   labels and audio scripts, simplify text for younger readers, suggest
   alt text, and summarise visitor comments.

2. Not approved
   Pasting collection records marked "restricted", donor names, loan
   agreements, visitor lists, or anything from the membership database.
   Using AI to write or alter provenance, dates, measurements or
   attributions. Using personal AI accounts for museum work.

3. Accuracy check
   Every AI-assisted label or script is checked against the catalogue
   entry by the curator (or named deputy) before printing or recording.
   The checker initials the draft. The draft is kept with the file.

4. Disclosure
   Printed interpretation carries no AI note if a curator wrote the
   final text. Audio guides and website text drafted with AI carry:
   "Drafted with AI assistance and checked by our curator."

5. Images and rights
   AI images are never used as if they were photographs of objects.
   Reproductions follow the licence on the source image. Any AI-made
   illustration is captioned as such.

6. Visitor data
   No visitor personal data goes into an AI tool unless the tool is
   on the approved list and the privacy notice covers it.

7. Problems
   Mistakes found after publication go to [name] within a week and are
   corrected on the web or on the next print run.

Notice what is missing: no technical vocabulary, no long list of banned products, and no promise to "monitor developments". Vendor products change monthly, so the policy names a tool only in section 1, where a single edit updates it.

Labels: the curator check in practice

The accuracy rule in section 3 only works if the check is concrete. "Review for accuracy" invites a skim. Give the curator a four-line test and a place to tick it.

1. Every proper noun. People, places, makers and organisations are matched against the catalogue entry, character by character.
2. Every number. Dates, dimensions, quantities and prices. If the draft contains a figure that is not in the catalogue entry, delete it.
3. Every cause and consequence. Sentences that explain why something was made, or what it led to, are where AI fills gaps with plausible guesses. Each one needs a source the curator can name.
4. Every adjective of rank. "Earliest", "finest", "only surviving", "rare". These are claims. Remove them unless the curator can defend them.

Here is an illustrative before and after. The catalogue entry says: stoneware jug, salt-glazed, made in a small local pottery, about 1870 to 1890, handle repaired, donated 1974. A volunteer asked an AI tool for a 60-word label and got this back.

DRAFT FROM AI TOOL (illustrative)
This salt-glazed stoneware jug was crafted around 1875 by a master
potter at one of the region's most celebrated workshops. Used daily in
farmhouse kitchens, it is among the finest surviving examples of its
type. The carefully mended handle tells the story of a much-loved
household object.

The curator's pass catches four problems. "Around 1875" narrows the catalogue's 1870 to 1890 range. "Master potter" and "celebrated workshops" are not in the record. "Among the finest surviving examples" is a rank claim with no evidence. "Used daily in farmhouse kitchens" is a guess about use. The corrected label reads: "Salt-glazed stoneware jug, made between about 1870 and 1890. The handle has been repaired. Donated in 1974." It is plainer, and it is true.

That edit took about three minutes. Across a gallery of forty labels, budget two hours for the curator and say so in the policy, so nobody is surprised when the check slows the print deadline.

A prompt that makes the draft easier to check

You can reduce the damage at the source by telling the tool to stay inside the facts you supply. This does not remove the need for the check; it makes the check faster because there is less invented material to strip out.

Write a 60-word museum label for general visitors, reading age 12.
Use ONLY the facts below. Do not add dates, names, places, uses or
comparisons that are not listed. If a useful fact is missing, write
"[MISSING: what]" instead of guessing.

Facts:
- Object: stoneware jug, salt-glazed
- Made: about 1870 to 1890, small local pottery
- Condition: handle repaired
- Donated: 1974

Run the same facts twice. If the two drafts disagree about anything, that disagreement points to a place where the tool is filling a gap, and the curator should look there first.

Audio scripts: spoken text needs a second read

Audio scripts get the same four-line test, plus three checks that only matter when words are spoken.

• Pronunciation list. Keep a one-page list of names and place words as they should sound. Synthetic voices mangle unusual names, and a hired narrator needs the list anyway.
• Listen once at full speed. A script that reads well can sound stilted aloud. A volunteer who was not involved in the drafting listens once and flags anything unclear.
• Version control. Each recording carries the script version and the date the curator signed it off. When a fact changes, you know which recordings to redo.

If you use a synthetic voice, the disclosure rule applies to the voice as well as to the text. Some visitors do not mind; others feel misled when they discover a human-sounding guide is not human. A single line at the start of the guide ("This audio tour uses a computer-generated voice") settles the question cheaply.

A short illustrative case: a small maritime museum wants a twelve-stop audio tour. A volunteer drafts the stops with AI and the curator checks them against the catalogue, striking two invented anecdotes about a ship's captain. The tour is recorded with a synthetic voice, and the pronunciation list fixes three vessel names. Total curator time: about four hours, against perhaps twelve if she had written the stops from scratch. The saving is real, but only because the check was planned, not squeezed in.

A realistic mistake and how it showed up

Picture a museum that used AI to turn a long catalogue entry into a 90-second audio stop. The script said the object "was exhibited at the great exhibition of that year". The catalogue never mentioned an exhibition. The model had merged the entry with a well-known story about similar objects. Nobody caught it in drafting because the sentence read naturally; a visiting teacher caught it a month later and emailed the museum. The policy fix was small: the checker now highlights every sentence in the script and writes the catalogue field it came from in the margin. A sentence with no field next to it is deleted or sent back for sourcing.

The margin-note habit is worth copying whatever your size. It turns "does this sound right?" into "where did this come from?", which is a question a volunteer can answer without being an expert.

Handling updates after publication

Labels and scripts are not finished when they go up. New research, a corrected provenance or a family's feedback can change the facts. The policy's section 7 should name who receives corrections, and the log should record the revised version. For audio, keep the source script in a shared folder with the version number in the file name, so a correction means editing one script and re-recording one stop, not hunting through exports.

Disclosure: deciding what to say, once

Disclosure fights are usually about inconsistency, not principle. A museum that discloses on the audio guide but not the website invites the question of why. Pick a rule by channel and write it down.

The distinction in the columns matters. If a person rewrote the text and the AI only supplied a starting point, most museums reasonably treat it as the museum's own writing. If the AI wording went through with light edits, say so. Whatever you choose, apply it identically across galleries, and keep the wording of the disclosure line in the policy so volunteers copy it rather than inventing their own.

Object photographs, licensed images and AI illustrations

Museums handle three kinds of image, and the policy should treat them differently.

1. Photographs of your objects. These are records. Never replace one with an AI-generated picture, and never use AI to "restore", extend or tidy a photograph of an object without labelling it as an interpretation. A visitor must be able to tell what is the real object.
2. Third-party images. Archive photographs, loaned images and press pictures come with licences. Uploading them to an AI tool to describe or edit them may breach the licence or the lender's terms. Check the licence before you upload, and write "check the licence first" into the policy.
3. AI-made illustrations. Useful for a children's activity or a reconstruction scene. Caption them as AI-made. Ownership of AI output is unsettled and OpenAI's and Anthropic's terms assign you whatever rights they have "if any", with no promise that those rights exist. A museum should not rely on exclusive rights in a generated picture.

Alt text is a separate, useful job. AI can draft alt text for accessibility quickly, and a person checks it against the object, since a model can misname a material or invent a pattern. The policy line is simple: the curator or a trained volunteer checks alt text for objects, as for labels. For the mechanics, see AI alt text: generate and check image descriptions.

Visitor data: what must stay out of the tool

Visitor data is the section most likely to cause real harm, and the easiest to write plainly. What counts as visitor data here:

• Names, emails and phone numbers from bookings, memberships, donation forms or newsletter sign-ups.
• Free-text comments that name an individual, mention a health condition or describe a child.
• Photographs of identifiable visitors, especially children, and CCTV stills.
• Donor and legacy records, which are both personal and commercially sensitive.

The rule: none of it goes into an AI tool unless the museum has approved that tool for that data, the vendor's business or education terms say what happens to inputs, and your privacy notice covers the use. Consumer chatbot accounts are not the place for any of it. Business plans generally let administrators control training on inputs; consumer accounts rely on a setting each user must switch off. How to read those terms is covered in what to check in an AI tool's privacy policy and terms.

A common and legitimate job is summarising visitor feedback. The safe way is to strip names and contact details first, then paste only the comments. Three steps: export the comments column alone, scan it for names or "my daughter Anna" style identifiers, and delete or replace them with "[child]" before anything is pasted. The follow-on tutorial on turning visitor feedback into next season's plan goes through the analysis once the data is clean.

If you have visitors or members in the EU, data-protection law such as the GDPR applies to this data, and a visitor-facing AI chatbot is more than a drafting aid. Ask a qualified adviser before you launch one.

Introducing the policy to volunteers

A policy is read by the people who wrote it. For everyone else, you need a ten-minute conversation. A museum volunteer team typically includes people who have never used AI and people who use it daily, and the second group is where the risk sits, because they will already have habits.

A short induction works well. Show one real label going through the process: draft, curator markup, final. Show what not to paste. Hand over a card with the three rules that matter most: the facts-only prompt, the curator initials, and no visitor data. Ask volunteers to tell the policy owner what they already use AI for; you will almost certainly learn about a use you had not planned for and can add it to section 1 or section 2.

Keep a simple log of AI-assisted items: title, date, tool, who checked. A spreadsheet with four columns is enough. When a visitor questions a label, or a trustee asks how much you use AI, the log answers in a minute.

Checking the policy works

After three months, pick five items from the log at random and audit them.

1. Was the curator's initial on the draft?
2. Does the final text match the catalogue on every name and number?
3. Does the disclosure line appear where the table says it should?
4. Did any visitor data go into a tool? Ask the team directly and without blame.

If two of five items fail on the same step, the step is too hard or too vague. Simplify it rather than telling people to be more careful. If everything passes, check whether anyone is skipping the log, because passing audits and an empty log usually means the work is happening off the record.

What the policy deliberately leaves out

Some topics belong in other documents. Procurement of a visitor-facing chatbot belongs in the vendor checklist, covered in questions to ask an AI vendor. Collections-management data sharing belongs in your data-sharing agreements. Staff employment questions belong with HR. Keeping those out stops the one-page policy from growing into a twelve-page document nobody opens.

If you are collecting policies for accreditation or a funding bid, this page slots in as one item of evidence; the museum accreditation checklist shows where it sits among the others.

Three short cases to test your wording

Run these past your trustees. If the policy gives a clear answer to each, it is specific enough.

1. The helpful volunteer. A volunteer pastes the full membership export into a free chatbot to "tidy the spelling". Section 2 and section 6 both say no. The follow-up is to ask the volunteer what they were trying to achieve and offer an approved route, such as a spreadsheet clean-up done inside your own tools.
2. The schools pack. The learning officer drafts a worksheet with AI and prints forty copies. Section 3 requires the curator check; section 4 puts the disclosure line on the back. If the pack was printed before the check, the fix is a reprint, not a footnote, and the policy should say so.
3. The donor letter. A trustee wants AI to polish a thank-you to a legacy donor. The letter names the donor and refers to the bequest, both sensitive. Without the donor's name or amounts the tool can polish the tone, and the trustee adds the specifics afterwards. Write that substitution into your guidance: draft with placeholders, fill in by hand.

If a case produces an argument rather than an answer, the argument tells you which sentence in the policy to rewrite. Do that before the next trustees' meeting, and put the new review date at the top of the page.

If you would rather map this with someone, that is what my AI implementation consultation covers; you can get in touch here.

Questions trustees and volunteers ask about the policy

Do we have to tell visitors a label was drafted with AI?

There is usually no single rule forcing it for a curator-checked label, but a short disclosure is the safest default for interpretation text, and some funders and professional bodies expect it. Decide once, write it into the policy, and apply it to every label and audio script alike. Check any funding agreement for its own wording.

Can volunteers use ChatGPT on their own phones for label drafts?

Only if the policy says so. The safer rule is that volunteers draft in the museum's approved tool and never paste collection records, donor details or visitor lists into personal accounts. Put the rule in the volunteer handbook and the induction, not only in a policy file.

How often should the policy be reviewed?

Twice a year is enough for a small museum, plus a review whenever you add a new tool or a vendor changes its terms. Put the review date on the page and give one named person the job of checking vendor terms.

Who signs off an AI-assisted label?

The curator or collections lead who would sign off any label, not the person who ran the prompt. If the museum has no curator, name a trustee or an outside subject expert for each gallery and write that into the policy.

Further reads

• Can an AI Guide for Museums Stick to Your Own Collection Notes? — Keep a visitor-facing AI guide inside your own collection notes.
• How to Write an AI Usage Policy for Your Small Business — The general AI usage policy this museum page builds on.
• How to Roll Out an AI Policy So Staff Actually Follow It — Get staff and volunteers to follow the policy, not file it.
• How to Write an AI Policy for a Small Charity — Many small museums are charities; compare the charity version.
• AI Policy for a Small Library: One Page, Patron-Data Rules — A one-page policy for a library's patron data.
• AI Chatbot for a Small Library: Hours, Catalogue and Events — A neighbouring cultural-sector example with visitor-facing AI.
• Can You Make an AI Audio Tour Guide? Scripts, Voices and Licensing — A claim-log method for scripts, what four speech services' own terms say, consent wording for cloned voices and the royalty-free music trap.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: UNESCO Recommendation on the Ethics of Artificial Intelligence (2021); vendor terms and privacy pages of the AI tools you adopt (check each yourself); your professional body's current guidance.