AI in Museums: Real Use Cases Beyond the Hype

A digital Van Gogh answering visitors' questions in Paris. A robot reassembling fresco fragments in Pompeii. A carbonised scroll from Herculaneum read without ever being unrolled. It sounds like a keynote slide. It all happened in the last five years.

AI in museums is no longer a conference topic. It is a set of real projects, with measurable results and a few very instructive failures. This article is for directors, curators and learning or visitor-experience leads who want to know what actually works, what it costs in terms of risk, and where to start.

You will find eight use cases, each tied to a project you can check. Then the risks: accuracy and "hallucinations", copyright, privacy and GDPR, the EU AI Act. At the end, a checklist to use before you sign anything.

Why AI in museums matters now

Two practical reasons.

First, the tools are now within reach of small museums. You no longer need an in-house lab or a global tech sponsor. You will see that in the example from a small valley museum near Bergamo, Italy.

Second, the rules have arrived. The EU's AI Act applies in phases. Some obligations already apply to organisations that simply use AI, museums included. More on that below.

One caveat. "AI" is a wide label. It covers chatbots, synthetic voices, computer vision and predictive models. Not all of it is generative. And the most useful part is often the least spectacular.

8 real use cases of AI in museums

1. Letting visitors talk to the artist

During the 2023–24 exhibition Van Gogh in Auvers-sur-Oise, the Musée d'Orsay installed "Bonjour Vincent": a 3D avatar of Van Gogh that answered visitors' spoken questions in French or English. The project was supervised by a Van Gogh scholar, who fed the system the artist's correspondence and historically checked information.

In Florida, The Dalí Museum opened "Ask Dalí" in April 2024. Visitors pick up a replica of Dalí's Lobster Telephone and ask a question. The answers come from models trained on the artist's writings and archival audio, in a voice designed to sound like his.

The lesson: a persona works when its "brain" is fenced in by curated sources. Without that fence, a conversation becomes improvisation.

2. Multilingual content adapted to each audience

In May 2025 the Museo della Valle in Zogno, a small museum in northern Italy, launched DIDA. It is a browser-based assistant, with no app to download. Visitors frame a label with their phone and DIDA returns it as text and audio in five languages. Content adapts to the chosen profile: adult, child, or person with a cognitive disability. The project was funded through a regional call backed by the European Regional Development Fund.

The lesson: here AI does not replace the curator. It multiplies the curator's text across languages and audiences. It is the use closest to audio guides, and the one with the fastest return.

3. Accessibility: describing art for blind visitors

The Rijksmuseum in Amsterdam is working with Microsoft on accessibility. That includes training AI to generate audio descriptions of artworks for blind and low-vision people. The stated goal is to make the whole online collection accessible.

Note the method. Inclusive design workshops with staff came first. For the wider picture, see our guide to museum accessibility.

The lesson: a generated description is a draft to be tested with the people who will use it. Not a finished product.

4. Describing and tagging collections

Harvard Art Museums have used AI since 2016 to generate descriptions and tags for their entire collection. They compare outputs from many commercial services and publish them. The aim is twofold: create new ways into the collection, and expose the limits and biases of the models. The museum is candid that machine descriptions sometimes read like a newcomer's, sometimes like an expert's.

The lesson: machine-generated metadata is useful as an extra layer. Label it as such and keep it separate from curatorial records.

5. Reconstructing what was lost

In 1715 strips were cut from all four sides of Rembrandt's Night Watch so it would fit a new wall. In 2021 the Rijksmuseum's research team used neural networks trained on Rembrandt's technique to recreate the missing parts, based on a 17th-century copy. The reconstructions were printed and hung temporarily around the original. The museum was explicit: they would not become part of the actual restoration.

The lesson: AI can show what might have been. Honest museums say so on the label.

6. Restoration with robots: Pompeii

RePAIR was an EU Horizon 2020 project coordinated by Ca' Foscari University of Venice, with the Italian Institute of Technology and the Archaeological Park of Pompeii. From 2021 it built a system combining AI and robotics. Software solves the "puzzle" of digitised fragments. Two robotic arms with soft hands place them. Tests focused on ceiling frescoes from the House of the Painters at Work, damaged by the eruption of 79 AD and later shattered by Second World War bombing.

Archaeologists flagged errors and made suggestions throughout. The park's director, Gabriel Zuchtriegel, put it plainly: no technology can replace people's work, but it can help when the amount of data is very large.

The lesson: AI proposes, the expert decides.

7. Reading the unreadable: the Herculaneum scrolls

The Herculaneum papyri were carbonised by Vesuvius. Hundreds are kept in Naples. Opening them physically risks destroying them. In 2023 the Vesuvius Challenge released CT scans of scrolls held at the Institut de France. Teams worldwide used machine learning to detect ink inside a scroll that was still rolled up. In February 2024 the grand prize went to a team that made more than 2,000 characters legible, about 5% of the first scroll: a never-before-seen philosophical text on pleasure.

The most useful detail for museums is how organisers guarded against hallucination. Independent reproduction of the code. Different models compared on the same area. Tiny analysis windows that detect specks of ink, not letters.

The lesson: trustworthy AI is verifiable. If a vendor cannot explain how errors are caught, take note.

8. Managing visitor flows

At the Uffizi in Florence, the University of L'Aquila built a system to end queues on free-admission days. One model predicts how many people will arrive and how long they will stay. A second decides how many entries to release every 15 minutes. Visitors collect a slip with an entry time and come back later.

According to the 2022 study, average waiting time fell from 64 minutes (2 September 2018, old system) to 14 minutes (7 October 2018, new system), while entries rose from 6,961 to 7,765.

This is not generative AI. It is statistics and optimisation. The authors name cameras and AI for counting people as their next step.

The lesson: the value often sits in data you already have, read well. The same applies to audio guide data: listening time, drop-off points, languages. More in museum KPIs.

The risks of AI in museums (and how to manage them)

Accuracy and hallucinations

In 2023 the Nasher Museum of Art at Duke University asked ChatGPT to curate a show from its own collection. The museum's account, published by the American Alliance of Museums, is worth reading in full. At first the chatbot proposed Van Gogh's Starry Night, which belongs to MoMA. Even after being connected to the collection data, it called a work on paper a sculpture, renamed objects to fit its theme and wrote vague or wrong labels. Human curators had to step in.

And when an error reaches the public, it is yours. In 2024 a Canadian tribunal held Air Canada liable for wrong fare information given by its chatbot. The airline argued the chatbot was responsible for its own actions. The tribunal replied that it makes no difference whether information comes from a static page or a chatbot.

How to manage it:

• restrict answers to sources approved by the museum
• design an explicit "I don't know" instead of an invented answer
• have someone who knows the collection review anything published
• sample real conversations every month

Copyright

Three points to keep in mind.

1. What goes into the model. EU rules on text and data mining, in the 2019 Copyright Directive, let rights holders reserve their works from mining. Before uploading third-party catalogues, essays or photographs into an AI tool, check your contracts.
2. Who owns the output. National rules are moving. Italy, for example, amended its copyright law in 2025 to protect works created with AI tools only when they result from the author's own intellectual work. Text generated with no real human contribution may not be yours to protect.
3. How the public reacts. In 2023 the Mauritshuis showed an AI-generated image among visitors' tributes to Girl with a Pearl Earring. It triggered a public backlash about artists' rights. Reputational risk matters as much as legal risk.

Privacy and GDPR

A conversational assistant collects questions, sometimes voices, sometimes data from children on school trips. That is personal data.

European regulators take this seriously. In March 2023 Italy's data protection authority temporarily restricted ChatGPT. Its later €15 million fine was annulled by a Rome court in March 2026, but the point for museums stands. The GDPR applies to chatbot conversations too.

In practice:

• give a clear privacy notice before the first question
• sign a data processing agreement with the vendor (Article 28 GDPR)
• know where data is stored and whether it is used to train models
• collect the minimum, especially with schools

EU AI Act: what already applies to a museum

The AI Act separates whoever builds a system (the provider) from whoever uses it (the deployer). A museum that adopts a vendor's chatbot or AI audio guide is usually a deployer. The dates that matter:

• since 2 February 2025, AI literacy applies: organisations using AI must take measures so their staff have a sufficient level of AI skills
• since 2 August 2026, the Article 50 transparency obligations apply. People must be told they are interacting with AI, unless it is obvious. Generated voices, images or video that resemble real people and could pass as authentic must be disclosed. For evidently artistic or creative works, a disclosure that does not spoil the experience is enough

For an avatar like the ones above, the practical rule is simple. Say it up front, clearly and accessibly.

Checklist: 10 questions before any AI project

Bring it to your next meeting, before you ask for quotes.

• Which visitor or staff problem does it solve? Write it in one sentence.
• Which sources does it draw on? Are they the museum's own, checked sources?
• Who reviews content before publication? How often?
• What happens when the AI does not know the answer?
• Does the visitor understand right away that they are talking to AI?
• What personal data does it collect? Where is it stored? Is it used for training?
• Do we hold the rights to the texts and images we upload?
• Can staff use the tool and explain its limits?
• How will we measure results? Usage, completion, recurring questions.
• If the vendor disappears tomorrow, can we export our content?

If you answer "I don't know" to three or more, you are not ready to sign. You are ready to ask better questions.

Where small and mid-sized museums should start

The headline projects (avatars, robots, particle accelerators) need research partners and budgets. But Zogno and the Rijksmuseum point to a more realistic entry point: the layer between the object and the visitor.

Three steps:

1. Start with the texts you already have. Labels, catalogue entries, guided-tour scripts. They are your safest raw material.
2. Use AI to multiply, not to invent. Translations, family versions, short versions. Then human review. For how to structure stops, see How to write an audio guide script.
3. Measure and fix. Find where people stop listening and rewrite there.

A museum audio tour app is where these three steps meet today. Platforms such as Museoo use AI for languages, audience versions and analytics, with an approval workflow that leaves the final word to curators. If you are starting from scratch, begin with How to create a museum audio guide.

Common mistakes

• Starting from the technology ("we want an avatar") instead of the problem.
• Letting the AI answer anything, with no source boundary.
• Publishing generated text without review by someone who knows the works.
• Hiding the AI from visitors.
• Uploading third-party material without checking rights.
• Forgetting staff training, which is already a legal duty.
• Measuring nothing, and so never knowing whether it worked.

Frequently asked questions

Will AI replace guides and curators?

The projects above say no. In Pompeii, at the Nasher and in the Vesuvius Challenge, the value came when experts checked and corrected the output. AI shifts human work towards selection, verification and storytelling.

Can a small museum afford AI?

Yes, if it picks a narrow use. Zogno shows a small museum adding a multilingual assistant thanks to regional funding. Translations and audience-specific versions are the most accessible starting point.

Do I have to disclose that a text was written with AI?

It depends. For a chatbot or avatar, since 2 August 2026 visitors must know they are interacting with AI, unless it is obvious. For text, the labelling duty covers text published to inform the public on matters of public interest, and does not apply where there is human review or editorial control. Disclosing it anyway is good practice for trust.

Can I use ChatGPT to write labels?

As a first draft, yes. As final copy, not without review. The Nasher documented errors in titles, media and descriptions even with collection data available.

Which project should we start with?

One that uses content you have already verified and gives a measurable result. For many museums, that is a multilingual audio guide with family and accessible versions. For narrative craft, see museum storytelling.

To wrap up

AI in museums works when it has a boundary: curated sources, human review, transparency with the public. Outside that boundary it produces confident errors, rights disputes and visitors who feel misled.

The eight projects you just read share one thing. None of them removed the experts from the room. They let them reach further.

Want to see how AI works inside an audio guide, with final approval in curators' hands? Book a demo, no commitment.