Insights from a global AI survey of 400 museums
Declaration: AI wasn’t used in the writing or editing of this article’s text (or outlining of content), nor summarising or analysing any documents discussed (everything mentioned was read in full, minus some appendices). I use spell-check features (via Google Docs) for support with my dyslexia (with grammar-check turned off for personal preference). I do use AI for website coding and templating, including modifying how articles appear visually once fully written.
Hanna Barakat & Archival Images of AI + AIxDESIGN / https://betterimagesofai.org / https://creativecommons.org/licenses/by/4.0/
On 17 September 2026, the United Nations Educational, Scientific and Cultural Organization (UNESCO) released its ‘Global Survey on the use of Artificial Intelligence in Museums’ (read it here). Representing 403 museums from 90 different countries, it’s one of the most interesting and valuable resources I’ve seen published on this topic this year.
For my own work, I tend to use reports like this to adjust my core training and policy support offering, and generally shift around the emphasis of the content in my sessions based on what concerns or areas of interest I’m seeing being discussed at different stages in time or in different work areas. I combine this with the other ways I keep up to date with emerging concerns and conversations about AI within the sector (you can learn more about that in this previous article about AI use), and then try to present that back to my clients (to the best of my ability) so their own discussions, experiments and decisions can be better informed.
Rather than just summarise the full report, I thought it could be more interesting to instead highlight five items which I found particularly insightful, and combine these with some of my own reflections based on my own perception of what I’m seeing within the museum and culture sector.
1) Evaluation is lagging behind experimentation
From the report: “Museums are increasingly able to identify perceived benefits of AI, particularly the reported improved efficiency, but comparatively few appear to have established formal indicators to measure those benefits systematically.”
Essentially, it seems that there’s a lot of individual AI exploration going on, but less organisational muscle has been built to actually note and log the results of this exploration in a more systematic and objective way.
This tracks quite closely with what I’m seeing. Part of my work with The Audience Agency, during last year’s Let’s Get Real: AI, and this year’s (ongoing) Let’s Get Real: Social Impact + AI, involves working with our cohorts on how we’ll actually build up practical forms of experimental design into our projects with AI, so we can create some genuine organisational learning which is free (or at least, reduced) from individual subjectivity.
That said, my observation is that individual subjectivity actually works well to initially explore AI and make informed decisions about a person’s own relationship towards it. I’ve been advocating for this type of playful exploration for several years (see this article from Arts Professional in 2023), but this category of usage is often distinct from organisational alignment around AI practices, which is when more mindful experimental design becomes important.
Good experimental design for AI ideally begins with an actual hypothesis being tested, builds in strategies to reduce various forms of bias (particularly confirmation bias and selection bias), is mindful of how the data is being received (surveys, A/B testing, interviews, scoring, website analytics), is repeatable to an extent (especially when using updated AI models the following year), and has an agreed standard prior to the experimentation phase about how the results are interpreted and what standard is being set for different conclusions.
Adding to this, I also try to emphasise with clients that experiments around AI don’t need to be defacto prototypes about how an organisation might begin using or adopting AI. It’s absolutely viable to create experiments which explore how AI might affect your work negatively, introduce additional bias or harm, and potentially create evidence for why choosing not to integrate those processes into your organisation is advised. Regardless though, the experiment should be free of pressure to ‘prove’ anything; results will often be mixed, have some surprising failure or success cases, and require further interpretation or exploration before solid conclusions can be reached.
A great outcome from an experiment might be that another experiment (with different data and design) is necessary before the underlying question or hypothesis can be resolved. The goal to create a recurring, nimble practice of experimentation (whether these individual experiments take place over days or months), rather than the occasional flagship organisational experiment where its results are probably over-extrapolated from or have too much stakeholder input to be free from competing interests about what it should prove.
To learn a more about how experimental design can be built into AI exploration within cultural organisations, the full report from last year’s Let’s Get Real: AI is full of case studies exploring that question.
2) The majority of museums don’t have an AI policy or guidelines for use
From the report: “When asked about existence of internal AI policy, strategy or guidelines, only 7% of museums reported the implementation of a formal plan, while 15% reported the use of informal guidelines, pilot projects or internal practices. The majority of museums, 55%, reported having no AI policy, strategy or guidelines. For 22% of museums, a strategy is currently under development.”
Since I encountered it, I’ve taken a lot of inspiration from Rachel Coldicutt’s article ‘FOMO is not a strategy’ (I suggest you read it in full if you have time: Read here). Although the original article doesn’t discuss AI FOMO in the explicit context of AI policy production, I personally have seen parallels within my work; I’ve observed a worry that, without a formal policy document (which their peer-organisations might already be known to have), some senior members of the organisation might feel it is adrift in some way and particularly vulnerable to AI’s more likely dangers.
This may or may not be true in reality. Although I’m happy to help organisations develop an AI policy if they’ve decided they need one, my initial question with that strand of work is always “What has been your journey to deciding you need a formal policy?,” and “If this policy is created, is there genuine organisational energy to engage staff with it and keep the policy sufficiently updated?”. AI moves quickly and if policy can’t keep up, it risks becoming functionally redundant quite fast. Because of this, staff guidelines are sometimes produced instead - either as a bridge to a full policy, or because they’re generally seen as easier to communicate and have the ability to be updated more quickly as concerns and usage patterns change over time.
Either approach can work. It can also be perfectly suitable to have neither, if the organisation take a different approach - perhaps agreeing on general principles for use, or (often for smaller organisations) concluding that the existing procedures for discussing work with their team will be sufficient for addressing AI-related issues, opportunities or concerns.
Generally speaking though, I tend to find that larger museums usually conclude that some formal policy work is necessary and helpful, at which point it becomes productive to discuss what sort of outputs are realistic, how staff will be involved (to feel ownership over the policy and avoid the sense that it is implemented entirely top-down), and what mechanisms will be in place to adjust the policy or make it a suitably ‘living document’ which can adapt and respond to potentially rapid capability changes within AI. Following that, a more bespoke policy design roadmap can be drawn up.
If you’re interested in learning more about AI policies and how they might be built, I’d signpost you to the ‘Developing an AI Policy’ section of Arts Council England’s Responsible AI Practical Toolkit
3) The report is full of interesting examples of specific AI use by different museums
From the report: “Respondents reported, among others, the use of handwritten-text recognition, semantic collection search, named-entity recognition, object identification, automated metadata generation, alt-text production, archival photograph colorization, AI-assisted coding, visitor analytics, geographic information systems (GIS), and archaeological predictive modelling.”
Among discussing the most common uses of AI within the museums surveyed (understanding and analysing text, creating content, researching information, analysing images or videos, speech understanding, visitor engagement, data analysis, personalised recommendations, etc), there were some interesting individual examples being highlighted.
One which I hadn’t previously encountered came from the Fermanagh County Museum which “described using AI to create visual content for folklore storytelling, enabling it to share medieval and prehistoric stories through reels on its website and social media. The museum noted that this would not previously have been possible within its budget.” [Note: I had a search for these assets online out of curiosity, but couldn’t confidently locate them. I’m going to reach out to Fermanagh County Museum and see if they’re comfortable sharing or providing more details]
Another example, from the Wellcome Collection in London “described work focused on using AI “to aid discovery and use of our collections”, which includes using language and vision models to support handwritten-text recognition, semantic search, named-entity recognition and collection discovery.”
As an aside, if testing areas like handwriting recognition from images feels like it could be interesting to explore, I generally direct clients to begin initial experimentation with a wide range of off-the-shelf models. A pretty good starting point is to use a website like https://openrouter.ai/, load about £10 into it, and then test a range of different samples (assuming they’re public domain) against numerous AI models (both the major frontier ones and also smaller, potentially open-source options). This can be an interesting and efficient way to test dozens of models at once on the same set of data.
4) 87% of the 403 museums within the report don’t offer any learning opportunities related to AI for staff, and 86% don’t allocate any funding at all towards AI activities
This is a bit difficult for me to discuss because I financially support myself by running independent AI training and experimentation sessions for cultural organisations, so I have an obvious vested stake in the perceived value of AI learning opportunities, and it would be disingenuous to not immediately acknowledge that. With that on the table though, I was still struck by just how high the 87% figure was, especially given the wide discourse on how AI and synthetic generative media is affecting how our society is absorbing information and deciding what is trustworthy.
This prerogative is mentioned later in the report: “One recurring concern relates to the credibility of the museum and public trust. The Metropolitan Museum of Art (United States) noted “pushback and negative feedback from some visitors”as a challenge encountered in response to AI-related activities. Respondents noted that museums occupy a particular position as trusted providers of cultural or historical information, a role that could be undermined by inaccurate, poorly sourced or insufficiently transparent AI-generated content.”
While stretched financial resources (and stretched time) are undoubtedly going to be factors which limit training provision for museums, I suspect another tension might exist around defining exactly what ‘AI training’ refers to, and whether an educational provision which feels appropriate can be located.
It’s not difficult to find an abundance of ‘AI literacy’ programmes online - especially those produced by the same technology companies which are creating the latest AI models and have a financial interest in their use - but that creates an unavoidable issue with an overly positive framing and a subtle assumption that AI adoption is inevitable, shifting the question away from ‘if’ and more towards ‘how’ AI is going to be used within a museum. That’s obviously quite problematic, especially for a museum which is already skeptical about how AI will interact with its processes, and if it wants to develop the ability to be critical (and perhaps even represent its interests against AI) just as much as any technical strategy with a specific tool or suggested workflow.
My own approach towards AI within cultural settings takes the perspective that AI literacy shouldn’t be defined as just the tools and tech an organisation’s staff are going to use themselves. Equally important to understand is how information could be processed or wholesale generated by AI before it arrives to an employee, how external agencies may be using AI within their processes which affect the results the museums use, how audiences may have changing expectations about the providence of information due to the ease at which AI can produce realistic alternatives, or even how scams, deepfakes and bots might significantly affect individual employees or the organisation more widely.
The report itself touches on this idea: “a small but important group of respondents expressed reservations about framing capacity-building solely around increased AI adoption. Comments included “None. We don't use it and don't want to”and a request for guidance on “how to not encourage the use of AI in the museum industry.” These responses indicate that capacity-building may also need to support museums in critically evaluating whether and when AI is appropriate, rather than assuming that adoption is always the desired outcome.”
5) Museums are surprisingly familiar with open-source tools
From the report: “Of museums that indicated using AI, 51% use commercial off-the-shelf AI tools, 38% use open-source AI tools and 26% use AI functionalities integrated into existing museum software. In addition, 21%reported using custom-developed AI tools, and 18% indicated that their AI tools have been developed in partnership with universities or companies.”
Although we should take into account that there’s likely some selection bias in this sample (those more engaged with AI are probably more likely to complete a survey about it), it still struck me as surprising just how familiar this sample seems to be with open source AI tools. Often when I work with organisations in this space, there seems to be less awareness of this distinction than that number might suggest.
If you’re less familiar with the difference between open and closed source AI, this Wikipedia page explains the fundamentals fairly well
Generally speaking, the perceived advantage of using closed-source AI models is that they tend to be the cutting-edge (at least for now), with most dominant AI models (like ChatGPT, Claude and Google Gemini) falling into this category. They’re also usually better integrated into apps and web services - e.g., OpenAI’s new (at time of writing) Astra model scores highly for computer use functionality, being able to jump between different programmes and web browsers to complete a particular type of work. The issues for using these models often revolve around cost (they require subscription payments or usually more expensive pay-as-you-go APIs), environmental tracking (these models are often criticised for making it difficult to determine exact energy use), political control (we’ve seen the USA restrict certain models for periods of time from being accessed by other countries) and the perceived potential for ideological drift (given that these models are produced in China and the United States, there is concern that they might gradually reflect pressures from those governments which affect their output).
It depends on the type of work you’re looking to do with an AI model, but as of September 2026 there’s a good argument to be made that you can achieve comparable quality (or at least, sufficient quality) by using open source models instead for some tasks. While these can theoretically be downloaded locally and live entirely on your own server, you can also access them online and you have a wider choice of providers, which typically reduces their cost substantially. This tends to allow much greater environmental tracking, data security (when used locally), more stable models (as they don’t change until they’re specifically updated), and the ability the more fully customise their outputs and refine their capabilities.
If you want to start experimenting with these yourself, on your own computer, I tend to recommend https://www.jan.ai/ - you download the initial program, then you can use which open source models to install and attempt to run. If you’re not sure which models might work for you, you can use websites like https://www.canirun.ai/ and https://willitrunai.com/ to help you narrow down your best options.
