Kaël E. Khoram
← Writing

Article

When Museum AI Adoption Outpaces Governance

The first UNESCO–ICOM global survey on AI in museums finds AI spreading through everyday museum work while formal governance, staff training and evaluation remain much less developed. For museum leaders, the immediate task may be to understand the uses already in place before deciding what kind of policy they need.

Published 24 September 2026 · 8 min read

By Kaël E. Khoram

In this article

I take a look at the first UNESCO–ICOM survey on AI in museums to examine what governance looks like once AI is already part of ordinary museum work. The starting point is to identify the uses already in place, understand where professional judgement is happening and decide which uses are worth continuing. From there, museums can match governance to the consequences of each use, give staff enough guidance to recognise when an institutional decision is needed, and turn those decisions into a policy they can actually apply.

How AI Becomes Part of Ordinary Museum Work

In September 2026, UNESCO and the International Council of Museums published the results of their first global survey on the use of AI in museums.

Of the 403 museums included in the analysis, 57% reported using AI. At the same time, 55% reported having no internal AI policy, strategy or guidelines. The way museums describe that use is just as interesting as the headline figures: among AI-using respondents, 77% said they were exploring possible uses, 58% reported individual staff use, 40% limited operational use in specific departments and 33% pilot projects. Only 11% described AI use as widespread across the institution. Participation was voluntary, and UNESCO and ICOM caution that the findings should not be treated as representative estimates for the global museum sector.1

According to these figures, AI adoption in a museum does not necessarily arrive as an “AI programme”. Respondents describe uses ranging from summarisation and transcription to semantic collection search, object identification and automated metadata generation. The way museums access AI is also consistent with use becoming distributed through ordinary work rather than taking the form of a single institution-wide programme: among AI-using respondents, 84% access AI through external tools or platforms, while only 13% report developing and using systems in-house. For many museums, then, governing AI will involve tools they did not build and decisions that arise inside ordinary work.

A generic drafting task raises different questions from putting unpublished collection material into an external service. Translation for an internal document is different from generated interpretation presented to visitors. The same technology might be used in all these cases, but what is being processed and what follows from its use can be drastically different.

Where Professional Judgement Happens

There is a finding in the report that makes the picture more complicated, and more encouraging. Among AI-using museums, 81% say AI-generated outputs are always reviewed by a human before they are used or published. Another 12% say they are usually reviewed. This figure is supported by the qualitative responses: museums repeatedly describe AI as supporting professional work rather than replacing professional judgement.1

This is important since the governance gap should not be mistaken for an absence of care. A curator may already know that an AI-assisted description needs checking against collection records. Someone in communications may have developed sensible habits about what they will and will not enter into an external service. Those judgements are part of the institution’s existing response to AI. But the harder question is whether the museum knows where they are happening and whether staff know what to do when professional judgement alone does not settle the issue. Different teams need enough clarity about consequential uses and the review they require, including who can make a decision when a case becomes difficult.

Seven Questions Before Writing an AI Policy

Museum-specific thinking about AI governance predates the current generative-AI wave. The Museums + AI Network’s 2020 planning toolkit was developed with museum professionals and academics and asked institutions to think about AI projects through their purpose, ethical implications and affected stakeholders. Six years later, parts of its technological context have inevitably dated. What remains useful is the way it treats AI as an institutional decision rather than simply a technical one.2

The UNESCO–ICOM 2026 survey suggests that the problem has also moved on. AI is now entering ordinary museum work through staff use and external services, without necessarily taking the form of a formally planned institutional project. The governance question therefore starts earlier: does the museum know where AI is already being used, who can decide whether a use is acceptable, what kind of review it needs and when it should stop?1

For that situation, I would start with seven questions:

  1. Which AI tools or AI-enabled services are already being used?
  2. What tasks are they being used for?
  3. What information or material enters them?
  4. Where do the outputs go, and who is affected by them?
  5. What human review already takes place?
  6. Who can approve, restrict or stop a particular use?
  7. What evidence would tell the museum that the use is worth continuing?

These questions aim to make the uses that matter visible enough for the organisation to make decisions about them.

Before AI Becomes Routine, Ask What It Improves

Governance discussions often centre on the risks involved in implementing AI, with adoption treated as the starting assumption. But this survey gives museums another angle from which to look at actual uses: many still do not know whether those uses are producing enough benefit to justify becoming normal practice. Among AI-using respondents, 21% reported improvements they considered measurable. Fourteen per cent said AI had not produced measurable improvements, while 65% said it was too early to assess.1

These figures describe a gap between experimentation and evaluation: museums are increasingly identifying perceived benefits, but relatively few appear to have formal ways to measure them systematically.

There are clear examples of benefits in the report as well, including accessibility, visitor services, education, administration, collections documentation and research. The question for an institution is whether a particular use deserves to move from experiment to routine. What problem is it solving? What would count as improvement? How much checking does it add? Would another approach serve the task as well or better?

This question predates the UNESCO–ICOM survey. A 2023 National Lottery Heritage Fund briefing advised heritage organisations to consider their values and strategic aims when deciding whether, how and when to use AI.3 A museum can therefore stop an AI use because it is unnecessary, disappointing or poorly suited to the work. It does not need to wait for something to go wrong.

Governance Should Match the Use

Among AI-using museums, the most frequently reported challenge was accuracy, followed by copyright and intellectual-property issues and data protection. Respondents also reported lack of technical expertise, limited capacity and funding, trust and several forms of bias. But these concerns do not have the same significance in every use. What matters depends on the task, the material involved, what happens to the output and the safeguards already in place.1

Not every museum needs the same governance structure. A small museum using a few general-purpose services may find that a short record of approved uses and a named person to resolve difficult questions is sufficient for many of those uses. A larger institution running collection-scale or visitor-facing systems may need formal review procedures, procurement requirements, training and responsibilities shared across several departments.

What matters is whether each institution has enough structure for the uses actually taking place and the consequences they can create.

I use Governance through Coherence here as a way of checking whether the governance around a particular AI use fits the institution’s actual practice and the responsibilities that use creates. Different uses can therefore lead to different decisions even when the underlying technology is the same.

Some Decisions Belong to the Institution

Only 17% of responding museums provide professional training to staff on AI use, while the demand for support is much higher. Eighty per cent identify technical training on AI for core museum functions as a need, 67% identify ethical AI use and 62% ask for guidance on data governance and intellectual-property rights.1

Training can help someone understand how a system works and recognise where its outputs may fail. It can also explain why the information entered into an external service matters. Some questions, however, belong to the museum itself. Can this type of collection material be entered into an external AI service? Can generated text be used in public interpretation? When should another person review a use? Who decides when staff disagree?

Training can help staff recognise these questions and understand when they need to be raised. But the museum still has to decide what its answer is.

The 55% figure will understandably lead to calls for more museum AI policies, and for some institutions, that is exactly what should follow. But the UNESCO–ICOM findings show that much of the underlying activity is already there: staff are experimenting, reviewing outputs, encountering problems and asking for more useful guidance.1 What many museums now need is a clearer connection between those everyday practices and institutional decisions.

Once those decisions are clearer, a policy can record them in a form that staff can use consistently. If the policy is written before the museum understands the situations it actually needs to govern, it risks remaining too general to help with the decisions people face in practice.

For an individual museum, the starting questions are much more concrete: where is AI already being used? Which uses matter enough to record? Who reviews the consequential ones? Who can decide that a use should change or stop? Those answers will give a policy something real to govern and support staff in the decisions they already face.

Sources

Footnotes

  1. UNESCO & International Council of Museums (ICOM), Final Report: Findings of the UNESCO-ICOM 2026 Global Survey on the Use of Artificial Intelligence in Museums, 17 September 2026.
    https://unesdoc.unesco.org/ark:/48223/pf0000399452_eng ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7

  2. Oonagh Murphy & Elena Villaespesa, AI: A Museum Planning Toolkit, Museums + AI Network, 2020.
    https://themuseumsai.network/toolkit/ ↩

  3. Mathilde Pavis, Digital Heritage Leadership Briefing: Artificial Intelligence, commissioned by The National Lottery Heritage Fund, 2023.
    https://www.heritagefund.org.uk/sites/default/files/media/attachments/Digital%20Heritage%20Leadership%20Briefing%20-%20Artificial%20Intelligence_English.pdf ↩