Museums keep doing this. Funders keep paying for it. Nobody learns.

In my previous piece on AI governance in cultural institutions, I argued that museums, galleries, heritage sites, and arts organizations have something to teach us about AI governance, because their authority rests on community trust rather than market position or legal mandate. Cultural institutions have been forced to develop what I’ve been calling legitimacy governance: not just asking whether a technology meets legal compliance requirements, but whether the communities they serve will still trust them if they deploy it.

When that hard-earned knowledge goes ignored by funding bodies (who built the conditions that make failure almost inevitable) and policy frameworks (that arrive too late and too abstract), institutions are left holding the bag.

After my last post, a designer of interactive museum experiences sent me a LinkedIn DM that captures what this looks like on the ground. The implementations he encounters, he said, tend to be “either on the side of correct and boring, or experiential in ways that feel one-off, and disconnected from long-term stewardship thinking.” When he tries to move institutions toward concrete action, they hesitate. He wanted to know why I think that is.

Both failure modes have the same origin, and it predates AI by decades

For roughly two decades, from the mid-1990s through December 31, 2020, Adobe Flash was the default infrastructure for interactive digital content in museums. Institutions like the Smithsonian and the V&A used Flash for educational interactives, audio tour supplements, children’s programming, accessibility resources, you name it.

Just like today, the grant programs funding digital interactive work throughout the 2000s rewarded novelty and deployment. Flash was the dominant platform, so Flash got built. Funding applications never required evaluation of platform dependency, succession planning, or any answer to what would happen when the technology became obsolete. The incentive to plan for what came next didn’t exist.

Adobe announced in July 2017 that Flash would reach end-of-life on December 31, 2020. Despite three and a half years’ notice, most cultural institutions spent that time doing nothing. By the deadline, the Smithsonian was still hastily archiving Flash-dependent sites, with some content already rendering as blank pages (Schmitz Fuhrig, 2021).

Institutions weren’t negligent so much as they were understaffed, underfunded, and operating inside a funding and governance ecosystem that had never once required them to think about platform dependency. Every generative AI tool built on a third-party API an institution doesn’t control repeats that same vulnerability.

This pattern isn’t unique to cultural institutions. I sit at the intersection of tech startup marketing and cultural institution board governance, and I’ve watched the same dynamic play out in both worlds. McKinsey’s 2025 State of AI survey found only one third of companies have achieved enterprise-wide scaling of new technologies. The rest stay stuck in what the industry calls “pilot purgatory”: the pilots work, but when the pilot glow fades, nobody owns the system, nobody owns the decision to scale, and the initiative quietly stalls (McKinsey, 2025). The difference is that a multinational caught in pilot purgatory has runway to burn. A cultural institution doesn’t.

The incentives are working exactly as designed

Across major public and private funding mechanisms for cultural institutions, grant applications are structured around a predictable set of criteria: reach, innovation, demonstrable impact within the funding period. These are not unreasonable questions, but the problem is what they systematically reward and what they don’t.

The funding criteria are built to reward novelty. “First-ever AI-powered gallery guide” is exciting. “Community-embedded digital program, now in its third year of iteration” isn’t sexy enough to get the big bucks. Institutions optimize accordingly. Curators and executive teams aren’t lazy or lacking in moral fibre, they’re just beaten down by a broken incentive structure.

Take for example, the Local Government Association’s analysis of UK cultural funding. It found that the fragmented, competitive grant structure forces cultural bodies to chase multiple small funding pots on timescales that make — in the LGA’s own words — “proper consultation and community engagement difficult” (Local Government Association, 2024). A funding system that acknowledges, in its own research, that its structure prevents the outcomes it claims to want.

Remember the “correct and boring” implementations my LinkedIn friend mentioned? Community engagement requirements, mandatory for many grant applications, are frequently satisfied by the cheapest available intervention. Walk through enough cultural institutions and you will encounter the detritus of “community engagement” requirements: the costume rack where you can take a selfie in period dress, the sticky notes on the gallery wall asking “how does this art make you feel?” Academics Steven Hadley and Eleonora Belfiore have identified the structural logic precisely: engagement requirements, as currently designed, risk being “co-opted into a perpetuation of the status quo” rather than producing genuine shifts in how institutions relate to their publics (Hadley and Belfiore, 2018). The curatorial expertise required to assemble a rigorous exhibition is not the same expertise required to design meaningful participation programs. Conflating them (or assuming one person can cover both with fifteen percent of their time and a craft supply budget) produces cringey results.

Grant cycles also produce what might be called governance debt: the accumulated obligations every technology deployment creates. Someone must own the system. Someone must update it as the technology evolves. Someone must field complaints when it goes wrong. Someone must decide when and how to retire it responsibly. The British Museum has the budget to staff those decisions. A regional photography archive doesn’t.

What good looks like: Start with a stewardship question, not a technology question

The Cleveland Museum of Art launched what became ArtLens in 2013 around a specific problem: how to help visitors build a self-directed relationship with a collection that can feel overwhelming to people who don’t already feel confident in museum spaces. That question, rather than any particular technology, has driven continuous iteration for over a decade. An NEA-funded study with 438 paired visitor surveys found that visitors who engaged with the system spent significantly more time in the galleries, felt more confident about their understanding of the collection, and reported that the museum felt more approachable (Ridenour, 2019). The technology changed across those iterations, but the underlying mission didn’t. CMA also made its code open source, sharing it with institutions including the Smithsonian. That kind of infrastructure thinking produces transferable value, in a way that “innovative digital experiences” deployed to tick a grant box never could.

What’s infuriating is that the structural conditions described in this article (the short-cycle funding, the novelty bias, the governance frameworks that arrived a decade too late) actively prevent more institutions from codifying and sharing what they’ve learned about long-term technology stewardship.

We need to solve the problem upstream

The first thing to change is arts funding criteria. Funding bodies are not passive observers of the novelty problem; they engineered it and they sustain it. Redesigning grant criteria to require applicants to demonstrate a full lifecycle governance plan would change institutional behavior without requiring institutions to act against their own survival interests. Sustainability requirements deserve equal weight to innovation and reach.

The second change we need is around AI procurement standards. The EU AI Act, which went into force in August 2024, requires transparency for some public-facing AI systems, such as chatbots, and much heavier accountability for systems it classes as high-risk. For those high-risk systems, Article 9 requires risk management to be maintained not just at launch but across the full lifecycle of a deployment (European Parliament, 2024). Most cultural deployments fall below that threshold. The OECD AI Principles include accountability as a core pillar; AI actors remain responsible for the systems they deploy and their impacts over time (OECD, 2019). Both frameworks articulate the right principles, but the Act reserves them for high-risk systems, and neither envisages the conditions in which they are achievable for institutions that have never been funded for the ongoing work of governance.

Current procurement frameworks, where they exist at all, evaluate AI systems at the point of acquisition. They do not require institutions to demonstrate lifecycle governance capacity before deployment: who will own this system, who will maintain it, who will audit it for drift from its original purpose, how it will be responsibly retired when its useful life ends.

Sector-agnostic AI procurement standards should mandate lifecycle governance planning as a precondition for deployment. This will put good pressure on funding bodies to start requiring applicants to demonstrate how they will sustain, not just launch.

Cultural institutions didn’t create the conditions that make governance failure almost inevitable, but they absorb all the public blame. Decades of documented and public failures make this argument unusually legible, but this is not a niche problem about arts funding. It is a preview. Every under-resourced public institution operating under short-cycle funding pressure (schools, local health services, social care providers) faces the same structural conditions. Most of them are now deploying AI.

References