Is there an adopted ICOM guideline for AI in museums?

No. On 25 June 2026 ICOM adopted a revised Code of Ethics that addresses digital technologies. A working group on digital ethics and artificial intelligence has been recommended, but according to AVICOM in August 2026 it has not yet been formally established. UNESCO and ICOM have carried out a global survey. The report has been announced, but there is still no finished ICOM standard to rely on.

When should the visitor know that the voice is synthetic?

There is no Danish law that specifically requires this of museum audio guides. As professional practice you should say so as soon as the voice could be taken for a human, and always if it gives voice to a named person, a curator or a historical figure. The shortest places are the first seconds of the audio, the landing page after QR scanning, and an icon on the sign.

What should the sign next to an AI audio guide say?

One sentence is enough on the wall: that the audio uses a synthetic voice, and that the text is the museum's. The icon should show that there is audio. The rest (supplier, model, training data, who has quality-assured it) belongs one click away on a short page you can point to from the sign, the website and the audio guide.

Should the museum disclose what data the AI model was trained on?

Jonas Heide Smith from SMK and Lisbet Tarp from Aarhus University say publicly that museums should both declare AI use and disclose training data. You do not need to publish the entire dataset. You should be able to say whether the model was trained on your own texts and collections, on open datasets, or on a supplier's general model, and who has checked the result.

How do you quality-assure an AI translation of interpretation texts?

Write the original in plain language, lock proper names and years, let the machine translate, and have a person with language and subject knowledge proofread before the text goes into the exhibition. Unpublished automatic translations of gallery texts and audio guides are the most common source of error. Put a date on the version, and make it easy to correct mistakes.

May we use AI in interpretation while ICOM is still working on guidelines?

Yes. ICOM's revised code already covers digital technologies, and Danish museums already use AI in interpretation. What is missing is a shared recipe for declaration. Until it arrives, the safest practice is to be specific with the visitor: what AI has done, what a human has approved, and where to read more.

How museums should declare their use of AI

On 25 June 2026 ICOM adopted a revised Code of Ethics with 85.9 per cent support. The new text addresses digital technologies, the climate crisis and the after-effects of colonialism, and is aligned with ICOM's museum definition from 2022.

Practical guidance on implementation is on the way. An adopted ICOM guideline on artificial intelligence, however, still does not exist. That is the gap most Danish museums stand in right now: you already use AI, or you are considering it, and the visitor cannot tell a recorded human voice from a synthetic one.

You can use the recommendations below regardless of which supplier you choose. Ask them for the same things you would ask of yourselves.

ICOM has adopted a new ethical foundation. An AI guideline is still waiting

ICOM's 41st Ordinary General Assembly adopted the revised code after a multi-year process involving 114 national and international committees. A PDF dated May 2026, which is still on icom.museum, was a draft. The current text is the one adopted on 25 June.

The code says that digital technologies form part of the museum's ethical responsibility. It does not tell you how to label a synthetic voice, how to disclose training data, or when a chatbot in the exhibition should say that it is a model.

The next step was already formulated in November 2025. At ICOM's 40th General Assembly in Dubai, members adopted Resolution no. 3, Positioning Museums for an Equitable, Ethical, and Sustainable Digital Future through Technologies. The resolution recommends an ICOM Working Group on Digital Ethics and Artificial Intelligence.

According to AVICOM's newsletter from August 2026, that working group is still not formally established.

UNESCO and ICOM sent out a global questionnaire survey on AI in museums in 2026. The deadline for replies was 21 July 2026. A joint UNESCO-ICOM report has been announced. It has not been published yet.

AVICOM points to two things that are more useful for a Danish museum than waiting for the finished standard. AI literacy should be a standing part of continuing professional development, and that applies to directors, curators, registrars, conservators, interpretation and communications staff, and suppliers. Those who present museum content must be able to understand when AI is being used.

The last of these is declaration in practice. If the person presenting the content does not know where the machine has been in the text, the visitor cannot get an honest answer either.

What Danish professionals are already saying out loud

In Denmark you do not need to wait for Paris. Jonas Heide Smith from the National Gallery of Denmark (SMK) and associate professor Lisbet Tarp from Aarhus University say publicly that museums should declare their use of AI and disclose which dataset it was trained on. That is the clearest Danish professional position available right now.

SMK is internationally known for working with AI in interpretation. The museum's open dataset of 45,000 works has been used by AI researchers globally. SMK itself writes about artificial intelligence at the museum and is building SMK Link, a browser-based, QR-code-based guide on the visitor's own device without an app. AI is used for machine translation and text-to-speech for accessibility reasons. Launch is expected in 2027.

At the QR code, SMK uses icons so the visitor can see whether audio, sign language, video or fuller descriptions await. That is declaration before the visitor has scanned: you say what the format is, so nobody is surprised.

You do not need SMK's dataset or development capacity to do the same on the sign. You should be able to say three things without hesitation:

Where has AI been in this piece of interpretation?
What has a human approved?
Where can the visitor read the rest?

If you cannot answer that internally, you cannot answer it at the ticket desk either.

Museum Helsingør already has AI in its interpretation

Helsingør Municipal Museums, together with Digital Impact ApS, have developed To Prompt, described as an AI-based history guide to Helsingør's historic town centre. The model is trained on information from around 18,000 articles, historical articles and archival records from the museum's written collection.

Project manager Sebastian Gyrst-Longsig Christensen has explained the rationale in concrete terms: the museum has 1,200 linear metres of archives and nearly 40,000 objects in store. Only a small part can be shown in the exhibitions. AI makes it possible to put the collections to work.

At the same time, Museum Helsingør received 1,575,000 kroner from the Agency for Culture and Palaces' development fund for the project Archaeology for Beginners. The fund is 10 million kroner a year and can be applied for by museums in basic grant categories 1-3. The project includes an AI archaeologist, where QR codes give digital access to finds on the museum's website.

Two different projects, the same museum organisation, the same core: AI is used to open something the gallery cannot hold. The visitor meets it as interpretation. The question is whether the visitor also meets a sentence about what kind of interpretation it is.

If you apply to the development fund yourselves, put declaration into the project from the start. It is cheaper than writing it on afterwards, and it is easier to explain to both the agency and the visitor.

When should the visitor be told about a synthetic voice?

There is no Danish special rule that says a museum must label a synthetic voice in an audio guide. The accessibility act and the web accessibility act deal with other things. ICOM, as mentioned, has not adopted an AI guide. What you have is an ethical code on digital technologies, AVICOM's requirement to be able to see when AI is used, and a Danish professional consensus on declaration.

Use that as a threshold, not as a loophole.

Always say it when the voice could be taken for a human. A fluent, natural AI voice in headphones sounds to most people like a recording. If you do not say so, you have implicitly claimed that a human read the text aloud.

Say it clearly when the voice is put in the mouth of a named person. A curator, an artist, an archaeologist at the museum, a historical figure. Otherwise the visitor will believe that that particular person said those particular sentences.

Say it before the conversation starts when the visitor can put questions to an AI archaeologist, a chatbot or a guide trained on your archive. Write that it is a model, what it is trained on, and that it can be wrong. A QR code that opens a chat without that sentence is the worst variant, because the format resembles a member of staff.

A short label is enough when you use text-to-speech on a text the museum's specialists have written, and the voice is clearly presented as an audio guide. Here the content is yours. The machine has only read it aloud. The visitor should still know. You do not need an eight-line legal disclaimer.

Do not hide the information in the terms and conditions, in a PDF under About the museum, or in a preface nobody hears because playback starts in the middle of the track.

A good rule of thumb: if a visitor asks a member of staff in the room, is it a human speaking?, the answer should already be on the sign.

What should the sign say?

The sign has three jobs. It should say that there is something to listen to. It should say that the voice is synthetic. It should point onwards if anyone wants to know more.

Keep the wall text short. The visitor is standing up, the phone is in one hand, and the room has other things to look at.

Short version for the wall:

Scan and listen. The voice is synthetic. The text is written by the museum.

Version when you also translate:

Scan and listen in your language. The voice is synthetic. The Danish text is the museum's. Translations are AI-generated and reviewed by the museum.

Version when the visitor can ask a model:

Here you can put questions to an AI trained on the museum's material. It is not a member of staff. Answers may contain errors.

Put an icon by the QR code, as SMK does: audio, text, video, sign language. The icon is declaration of format. The sentence is declaration of method. You need both.

Avoid formulations such as powered by artificial intelligence without substance, and avoid a genuine museum voice if it is not. Synthetic voice is understandable. AI narration is also fine if you use it consistently.

On the sign there should be one place to go for the rest: a short URL or the same QR that opens the guide. The landing page should have the declaration at the top, before the play button. If you only say it after three tracks, you have spoken to the visitor for several minutes under a false premise.

Do not print eight languages on the wall. One sentence in Danish and English is enough in most rooms. The rest sits in the guide, where the visitor has already chosen a language.

How to declare AI on the website

The website is the place you can be specific without filling the room. Make one short page, and link to it from the footer, from the audio guide and from press releases when you launch a new guide.

The page does not need to be called AI policy. It can be called How we use artificial intelligence in our interpretation. Keep it to one screen.

Write four blocks, in the order the visitor actually asks:

What do we use AI for? Translation, narration, archive search, a chatbot in the exhibition. Be specific. We use AI is too thin.

What has a human done? Who wrote the original text. Who approved the translation. Which languages have been checked by a person with both language and subject knowledge. Which languages are machine-translated only, if you have any of those.

What kind of model and data? You do not need to publish the model weights. You should be able to say whether you use a general language model from a supplier, a voice model, and whether it has been fine-tuned on your own texts, your collection or an open dataset. SMK and Tarp specifically ask for that information. Trained on the museum's own interpretation texts is an answer. We do not know, it is the supplier's model is also an answer, and then you should ask the supplier until you know more.

How do we correct errors? An email or a form. A date for the latest review. Without that block, the declaration becomes a sign you never maintain.

If you already have an accessibility statement, cross-refer between the two pages. Do not mix them together. Accessibility is about who can use the solution. Declaration is about what the solution is.

Also put one sentence on the ordinary visit page that links to the audio guide: The guide uses a synthetic voice. The visitor planning the visit from home should be able to see that before they stand in the room.

The audio guide: three places the visitor should be able to see it

An audio guide has an advantage over a wall sign. You have text, audio and a screen. Use all three, and repeat yourselves. The visitor scans in the middle of a conversation and rarely hears the track from second one.

1. The landing page after scanning. The first line under the title. Not below the fold, not in a Read more. Example: The voice you hear is synthetic. The text is written by [museum name].

2. The first seconds of the audio. One sentence before the interpretation begins. This track is read aloud with a synthetic voice. Then the specialist text. Yes, it costs three seconds. That is the price of the visitor knowing what they are listening to.

3. The player. A small, fixed label by the play button or by the language selector. Synthetic voice. When the visitor switches from Danish to German, the label should still be there. The translation is a new place AI has typically been.

If you offer both a human recording and a synthetic version, say which one is playing. The choice itself is declaration.

Keep the spoken disclaimer to one sentence. A long ethical statement in the ear, before the painting is explained, makes the visitor skip ahead. The details sit on the landing page and on the website page.

Test it with a colleague who has not been involved in building the guide. Give them the phone, ask them to scan, and after twenty seconds ask: did you know the voice was synthetic? If the answer is no, the text is in the wrong place.

How to quality-assure an AI translation

Declaration without quality assurance is a sign on a product you have not read yourselves. Tarp and Smith talk about training data. For most Danish museums the more pressing question is whether the German gallery text actually means what you think it does.

Start with the original. An unclear Danish sentence becomes an uncertain sentence in twelve languages. Write to be listened to: short sentences, specialist terms explained, proper names spelled as you want them spelled.

Lock what the machine must not guess at. Personal names, place names, work titles, years, materials, designations of origin. Give the translation a glossary. The Jelling stones must not become three different things in German.

Then let AI translate. That is the part that is cheap and fast. The part that decides whether you can stand behind it is the next one.

Have a human proofread before the text goes into the exhibition or into the ear. Ideally someone with both language and subject knowledge. If you only have one of those, prioritise the languages you have most visitors in, and be honest on the website page about which languages have been human-checked.

Check in particular:

negations (not, never, only), which models often reverse
numbers and centuries
designations for people, religion, colonial history and gender
quotations that suddenly are no longer quotations
instructions (do not touch, go to the right), which must be unambiguous in the room

Publish with a version date. When you correct the Danish original, the other languages must follow. An audio guide stuck on an old machine translation is worse than no translation, because it sounds finished.

Give the visitor a way to say that something is wrong. One sentence on the landing page is enough: If you find an error in the translation, please write to us. You get both better texts and an argument that the declaration is meant seriously.

If you do not have time to check a language, do not switch it on in the exhibition. A language you will not put your name to is not one you should ask the visitor to listen to.

A declaration model you can copy

You do not need three policies. You need one form that interpretation, communications and the supplier can fill in before anything goes live.

Level	What the visitor sees	What you have under control internally
1. Format	Icon on the sign: audio, text, video, chat	You know which format the QR code opens
2. Method	Synthetic voice / AI translation / AI chat	You know where in the chain AI has been
3. Responsibility	The text is written by the museum or answers are generated and may contain errors	Name of the person who approved the original and the language
4. Data	Link to a short page on model and training data	A supplier contract that actually answers the question
5. Correction	Email or form, version date	A routine for correcting all languages when the original changes

You can have levels 1-3 on the wall and in the audio this week. Levels 4-5 require a conversation with the supplier and a page on the website. If the supplier cannot say whether the voice is synthetic, or what the model is trained on, that is a procurement problem before it is a signage problem.

Write it into the requirements specification. The solution must be able to show a visible declaration on the landing page, in the player and in the first second of the audio. The museum must be able to edit the text. Then the responsibility sits in the right place.

Use the same words everywhere. If the sign says synthetic voice, the audio and the website should say the same. Artificial voice, AI voice and generated narration at three different touchpoints make the visitor think three different things are going on.

What you can do this month

You do not need to wait for ICOM's working group or for the UNESCO report. The code has been adopted. The Danish professional position has been stated out loud. Several museums already have AI in the room.

Do three things that can be seen:

Write the short wall sentence and put it by the QR code.
Put the same sentence at the top of the landing page and at the start of the audio.
Publish a short page on what you use AI for, what a human has approved, and how to report an error.

Then the conversation most people skip: ask your supplier for training data, model type and an explanation you dare put your logo next to. If the answer is unclear, the declaration is unclear too.

The visitor does not come to read an AI policy. The visitor comes to understand what you have on the wall and in the collection. The declaration is there so that understanding is not built on a voice you pretended was a human.

At museum-guide.com we declare it when the voice is synthetic. Ask your supplier for the same.

Sources

About the author: Thøger Elung-Jensen is co-founder of Museum Guide and works with AI-generated audio guides for museums. We declare it ourselves when the voice is synthetic.