Artificial Intelligence (AI) in Museums: What's Changing?

AI in Museums: A Practical Guide for Visitors and Museum Teams

Where AI Shows Up Across the Museum

Zone	Typical AI Use	What It Needs	What It Can Improve
Topic Focus	AI in Museums used as practical tools for collections, interpretation, and daily operations	Clear goals, staff review, and permissions	Consistency, easier discovery, and time saved
Collection Records	Suggesting keywords, materials, and subject tags	Structured metadata and controlled vocabularies	Stronger search results and clearer catalogs
Digital Access	Supporting APIs, recommendation features, and related-works suggestions	Open datasets (where permitted) and stable identifiers	Deeper online exploration and research use
Archives	Matching images, grouping related photos, assisting with description tasks	Digitized images with basic captions	Quicker links between records and context
On-Site Experience	Interactive guides that respond to visitor interests	Carefully written prompts and content guidelines	Engaging discovery without overwhelming visitors
Education	Drafting lesson variations and accessible summaries	Approved interpretive texts and educator oversight	More options for different learning styles
Governance and Ethics	Reviewing disclosure, cultural bias, rights, factual risk, and environmental cost	Written rules, named reviewers, and public explanations	Clearer accountability and fewer avoidable errors
Back Office	Helping staff draft internal notes and turn routines into templates	Non-sensitive inputs and clear boundaries	More time for core museum work

AI now appears in museum search tools, collection databases, image-matching projects, translation systems, educational resources, and visitor guides. It can help staff connect a photograph to an object record or make a large digital collection easier to search. It can also influence which objects visitors see, how a disputed history is summarized, and whether generated text or imagery is mistaken for verified museum content. Every public use therefore needs a stated purpose, identified sources, staff review, and a clear way to correct errors.

What Museum Staff Mean by AI 🧠

In museum settings, AI is a broad label for methods that recognize patterns, suggest connections, classify material, or generate draft text and images. These systems do not verify museum facts on their own. Qualified staff remain responsible for the final public content, especially when a tool addresses provenance, attribution, cultural identity, human remains, contested history, or living communities.

Machine learning recognizes patterns in data (images, text, numbers).
Computer vision works with images (identifying visual similarities, grouping photographs).
Natural language tools assist with text (summaries, translations, keyword suggestions).
Recommendation systems support discovery (related works, thematic pathways).

Collections and Catalogs: AI Starts with Clean Data 🏛️

Before any model can help, museums need records that are consistent and well-structured. Think of AI as a high-speed assistant: it moves quickly, but it can only work with what you give it.

What “Data-Ready” Looks Like

Stable object IDs that don’t change unexpectedly
Clear fields for artist, date, medium, culture, geography
Controlled vocabularies for materials and subjects
Documented rules for how staff describe objects

Where AI Helps First

Suggesting tags for “materials” and “techniques”
Flagging near-duplicates across records
Helping staff spot missing fields
Improving internal search with smarter synonyms

One reason major U.S. museums are so useful for researchers is that they publish large, structured datasets. The Metropolitan Museum of Art, for example, makes select collection datasets available for unrestricted use covering more than 470,000 artworks, which encourages careful experimentation and responsible reuse [Source-1✅].

Open Access at Scale: Why It Matters for AI 🔍

Many AI systems depend on large sets of consistent records. When museums publish high-quality images and metadata with clear permissions, researchers, educators, developers, and visitors can reuse that material with fewer uncertainties about object identity and rights.

When Smithsonian Open Access launched in 2020, it removed Smithsonian copyright restrictions from about 2.8 million digital collection images and made nearly two centuries of collection data available for broad public reuse [Source-2✅].

How Open Access Changes the Visitor Experience

Even if you never download a dataset, you benefit when museums publish structured data. You see it in better search filters, richer “related works” suggestions, and clearer object pages that connect art to people, places, and techniques.

Computer Vision for Archives and Installation History 📚

Archives are full of photographs: gallery shots, installation views, event documentation, conservation imagery. They are powerful, but they can be hard to navigate at scale. This is where computer vision becomes practical: it can help connect an image to a record, or group visuals by similarity so staff can work faster.

At The Museum of Modern Art, a collaboration used machine learning to comb through over 30,000 exhibition photos and look for matches with the museum’s online collection of more than 65,000 works, recognizing over 20,000 artworks and creating new links between exhibition history and collection pages [Source-3✅].

What the MoMA Project Shows

Start with a clear target: “link installation photos to known objects.”
Keep the output assistive: staff confirm every proposed match before it is saved.
Store decisions back into the catalog so the work compounds over time.

AI Collection Explorers for Visitors 🎟️

Visitor-facing AI should be optional, easy to understand, and limited to a defined task. A collection explorer may suggest related objects, while an on-site guide may answer questions from approved museum texts. The interface should state when AI is being used and where visitors can find the museum’s verified object record.

The National Gallery of Art highlights this approach with interactive experiences that include an A.I.-powered explorer designed to help visitors uncover unexpected artworks across the collection [Source-4✅].

If You’re Visiting in Person

Use AI explorers to find a starting point, then slow down in the gallery.
Try “related works” after you’ve spent time with one object you love.
Choose one digital tool per visit; leave space for looking.

If You’re Exploring From Home

Follow a theme (portraits, textiles, landscapes) and let the system suggest paths.
Save object IDs or titles; return later with fresh eyes.
Compare two “similar” objects and note what the model noticed versus what you noticed.

Interactive Studios: When AI Becomes a Learning Space 🤖

Some museum-facing AI projects function as learning environments. They let visitors examine how a system groups visual structure, similarity, and style across a large collection. The museum can use the same interface to explain why machine similarity does not always match curatorial interpretation, historical context, or cultural meaning.

MIT’s Gen Studio project describes a collaboration with The Met and Microsoft that built an interactive “web studio” for exploring structure in artworks, using neural networks to help visitors experiment with visual relationships and search for similar works in the museum’s digital collection [Source-5✅].

Standards for an Interactive AI Studio

Transparency: visitors understand what the system is doing in plain language.
Boundaries: the tool is curated so outputs stay aligned with the museum’s mission.
Interpretation that stays close to the collection and avoids abstract technical language.

Curatorial Programs About AI and Collection Data 🧩

Museums also use AI as a subject of public conversation—especially when open access datasets invite artists, researchers, and designers to explore how images and metadata shape what we see. This is where curators can do what they do best: frame context, language, and meaning.

In a 2020 account of its early AI work, The Met described an Open Access collection of more than 406,000 images and noted that its images and structured data had become machine-accessible through an API in 2018 [Source-6✅].

Curatorial Questions Worth Asking

Which parts of the collection are most visible in a dataset, and which are quieter?
How do metadata choices influence what a model groups together?
Where can AI widen access—especially for visitors who prefer audio, plain language, or multilingual formats?

2026 Update: AI Governance and the Revised ICOM Code of Ethics

On 21 May 2026, UNESCO and the International Council of Museums launched a worldwide survey on AI use in museums. The survey asked institutions of different sizes and types to report how they use AI, what training they need, and what problems they encounter in collection work, research, access, education, visitor engagement, and daily operations. The participation period closed on 21 July 2026. UNESCO and ICOM state that the findings will feed into a joint report with baseline data, examples from museums, and areas where staff need further support [Source-7✅].

ICOM adopted its revised Code of Ethics for Museums on 25 June 2026 during the 41st Ordinary General Assembly. The text was approved by 85.90% of voting members after a multi-year consultation involving 114 National and International Committees, Regional Alliances, and Affiliated Organisations. ICOM states that the revised Code addresses digital technologies, the climate crisis, and the legacies of colonialism while retaining shared standards for collections, public trust, and professional conduct [Source-8✅].

What the 2026 Changes Mean for Museum AI

Museums need to document where AI is used across collections, interpretation, education, visitor services, and internal operations.
Pair technical approval with checks for public trust, cultural context, collection rights, and environmental cost.
AI rules should be connected to the museum’s existing duties for provenance, care, access, transparency, and accountability.
Institutions should expect shared international guidance to change as the UNESCO–ICOM survey findings are published.

Six Questions Before AI Reaches the Public

Issue	What Can Go Wrong	What the Museum Should Do
Visitor Disclosure	People may assume generated text, audio, translation, or imagery was written or verified directly by museum staff.	Label the AI-assisted feature, explain its task, identify the source material, and provide a route to the verified record.
Cultural Bias	Old catalog terms, uneven digitization, colonial classifications, and gaps in the collection can be repeated or amplified.	Review data with curators, archivists, affected communities, and language specialists before training or deployment.
Copyright and Training Data	A model may rely on images, texts, voices, or artist material that the museum is not permitted to reuse for that purpose.	Record licenses, consent, provider terms, and reuse limits separately from ordinary collection access.
Human Oversight	Generated output may contain wrong dates, invented quotations, false attribution, or unsupported interpretation.	Name the staff member or team responsible for checking each public output and approving corrections.
Synthetic Historical Material	A generated reconstruction can be mistaken for a photograph, document, or evidence from the period.	Mark synthetic material at the point of display and explain which elements are documented, inferred, or invented.
Energy and Environmental Cost	Large models, repeated generation, and unnecessary high-resolution processing can add avoidable computing demand.	Use the smallest suitable system, limit repeated generation, reuse approved outputs, and include resource use in project review.

Tell Visitors When AI Shaped the Content

A visitor should not have to guess whether a label, translation, spoken answer, reconstruction, or recommendation was generated with AI. The disclosure can be brief, but it should appear where the content is used. A separate technology page is useful for detail, yet it does not replace a clear label beside the tool or output.

The explanation should identify the tool’s limited role. A museum might state that an answer was generated from approved collection records and checked by an educator, or that a translation was produced by software and reviewed by a named language team. When live answers cannot be checked before display, visitors should be told that errors are possible and given a direct link to the official object page or staff contact.

Collection Data Can Carry Old Biases

Museum datasets reflect earlier cataloging choices. Object names may use outdated language. Geographic and cultural labels may follow colonial systems. Some departments may have detailed records and high-resolution images while other collections remain only partly digitized. A recommendation system trained on this material can make the best-documented objects more visible and leave other histories harder to find.

Cleaning spelling and filling empty fields does not resolve these issues. Teams need to review whose terminology appears in the record, which communities were excluded from the description process, and whether uncertain information is being presented as settled fact. Museums can keep historical catalog terms for research while adding current language, source notes, community-preferred names, and a record of why terminology changed.

Copyright and Training Data Need Separate Checks

Permission to display an image on a collection page does not automatically grant permission to use that image for model training, generated derivatives, commercial services, or voice imitation. The same distinction applies to exhibition texts, oral histories, artist interviews, donor records, conservation files, and licensed archival material.

Before a museum sends material to an outside AI service, staff should record who owns it, what the license permits, whether personal or culturally restricted information is present, how the provider stores inputs, and whether those inputs may be used to train another model. Open-access status should be checked object by object or dataset by dataset rather than assumed from the museum’s general access language.

Human Expertise Is Required Before Publication

AI can draft a label or suggest a match, but it cannot take professional responsibility for the result. Curators, registrars, conservators, archivists, educators, rights staff, and community advisers may notice different kinds of error. Review should therefore match the subject. A general editor may check readability, while a provenance claim needs a specialist and a translation concerning a living community may need review from speakers or representatives of that community.

Higher-risk uses need stricter review. These include human remains, sacred objects, restitution and repatriation, conflict history, genocide, racial classification, medical collections, disputed ownership, artist attribution, and content about living people. In these areas, generated answers should not be published without named expert approval and a stored record of the sources used.

Synthetic Images Can Be Mistaken for Evidence

Generated images and video can fill visual gaps, but realistic output may be read as a historical photograph or an accurate reconstruction. Small errors in clothing, architecture, tools, inscriptions, anatomy, or social customs can create a false impression even when the overall scene looks convincing.

Museums should identify synthetic material directly in the caption and keep it separate from digitized collection objects. The caption should state the evidence used, the uncertain elements, and the person who reviewed the result. A generated reconstruction should never receive an object number, archive-style date, or display treatment that makes it resemble primary evidence.

Energy Use Belongs in the Project Decision

AI projects use computing resources during model training, image processing, search, and repeated generation. The environmental effect varies by model, provider, hardware, electricity source, output size, and frequency of use, so a museum should avoid unsupported claims that a tool is either harmless or unusually damaging.

A practical review can still reduce waste. Teams can choose a smaller model for a narrow task, process images in batches, cache approved answers, avoid generating multiple high-resolution versions that will not be used, and retire tools that attract little visitor use. These choices connect digital work with the revised ICOM Code’s attention to the climate crisis without turning every project into a technical accounting exercise.

Museum Governance: Clear Rules Make Better Tools ✅

AI projects can move faster than collection review, rights clearance, interpretation, and public approval. Museums therefore need written boundaries before staff begin testing a tool with collection data. The rules should define which materials may be entered, which subjects require specialist review, what visitors will be told, and who can suspend the system when errors appear.

One practical model comes from the Amon Carter Museum of American Art, described through the American Alliance of Museums, where staff developed guidelines and basic literacy resources to steer how AI may be used across the organization [Source-9✅].

Steps for an Internal Museum AI Rule Set

Write short staff rules covering acceptable inputs, private material, rights checks, and approval levels.
Keep a shared “approved sources” folder with interpretive texts, object records, vetted terminology, and current program language.
Assign review roles for factual accuracy, cultural context, rights, accessibility, and visitor disclosure.
Keep prompts, model versions, source lists, staff edits, and reasons for approval or rejection.
Create a public correction route and name the team responsible for responding.
Set a stop rule for repeated factual errors, rights concerns, harmful output, or unexplained changes after a model update.

Review Checks Before and After Release 🧭

Effective museum AI tends to feel steady because it has gone through careful review. Clear routines built around purpose, quality, and trust help teams test and improve systems over time. The National Institute of Standards and Technology offers a practical risk management model that many organizations use when designing, deploying, or working with AI systems [Source-10✅].

Before Launch

State one goal in plain language.
Decide what acceptable output means for accuracy, clarity, accessibility, and cultural context.
Choose the smallest dataset and model that can perform the task.
Confirm permissions for every input category, including training and generated reuse.
Name the reviewers and list subjects that require specialist approval.
Write the visitor disclosure before the interface is released.
Test false dates, invented quotations, biased recommendations, and misleading synthetic images.

After Launch

Sample outputs weekly and record factual, cultural, and rights-related errors.
Invite feedback from gallery staff, educators, researchers, and visitors.
Refresh approved source texts and retest after any model or dataset change.
Keep a human override path visible and easy.
Correct public errors in the same place where the original output appeared.
Review output volume, repeated generation, and computing use against the value visitors receive.

Visitors should be able to tell when AI shaped a museum answer, image, translation, or recommendation. They should also be able to reach the verified object record, report an error, and understand who reviewed the content. Museums that keep these controls visible can use AI for search, access, and routine assistance without weakening the authority of collection records or the responsibility of museum professionals.