The EU AI Act played it safe. Cultural institutions are paying for it.

This is the third article in a series on AI governance in cultural institutions. The first two pieces are Arts Institutions as AI Governance Laboratories and Museums Keep Doing This. Funders Keep Paying For It. Nobody Learns. This one operates on the regulatory text.

The EU AI Act came into force in August 2024. It is the most comprehensive attempt to govern artificial intelligence in the world, binding across EU member states, with a risk taxonomy designed to scale obligations to consequences, and enforcement mechanisms that give it teeth no previous framework had.

It was preceded by the OECD’s AI Principles in 2019, which established accountability across the full AI lifecycle as a foundational commitment, and by the EU High-Level Expert Group on Artificial Intelligence (AI HLEG)’s Ethics Guidelines for Trustworthy AI, also from 2019, which explicitly named cultural context as ethically relevant.

In 2021, UNESCO’s Recommendation on the Ethics of Artificial Intelligence followed, naming museums, galleries, libraries, and archives as AI deployment contexts requiring governance support. It articulated a deployer accountability principle, that responsibility for AI systems should rest with actors “corresponding to their role in the life cycle,” which includes the institutions deploying these tools, not just the companies building them (UNESCO, 2021, paras. 42-43).

By the time the AI Act was finalized, European legislators had three successive international frameworks articulating the same core commitments (deployer accountability, lifecycle governance, cultural impact) to build on.

But for an entire category of AI deployment, public-facing systems operated by cultural institutions in their capacity as custodians of collective identity, the AI Act’s most stringent requirements do not apply.

The institutions most likely to fail their communities publicly are the ones the Act leaves ungoverned.

The people who wrote the Act had UNESCO and the OECD principles on the desk, telling them what responsible AI governance required. They chose to narrow the scope anyway.

This article names three structural faults in the EU AI Act’s treatment of cultural institutions, and proposes the interventions that would correct them.

How the Act classifies risk, and what falls through

The EU AI Act organizes AI systems into a hierarchy of obligations. At the top: prohibited systems. Immediately below: high-risk systems, defined by Annex III’s eight domains: biometrics, critical infrastructure, employment decisions, credit scoring, law enforcement, border control, educational access, administration of justice. These systems face the Act’s most stringent requirements: lifecycle risk management under Article 9, transparency documentation under Article 13, mandatory human oversight under Article 14, continuous post-market monitoring under Article 72 (European Parliament, 2024).

Then: limited risk. Chatbots and deepfakes. Transparency obligations only (tell users they’re talking to a machine).

Then: minimal risk, like spam filters and video games. No mandatory obligations.

This is where most AI deployed by cultural institutions lands. A museum’s AI-powered collection guide: minimal risk. A heritage site’s visitor analytics system: minimal risk. A gallery’s generative content tool: probably limited risk, triggering a disclosure requirement and nothing else.

Annex III’s eight domains share a logic: AI harms individuals when it makes consequential decisions about their rights, life outcomes, or freedoms.

Cultural institutions’ AI deployments produce a different category of harm: collective damage to community trust, cultural legitimacy, and the institutional relationships that give museums, archives, and heritage bodies their social function.

The Act’s Recital 4 acknowledges that AI can affect “societal, economic, cultural and democratic transformations” (European Parliament, 2024). No Annex III domain covers public-facing AI deployments by cultural or heritage institutions in their capacity as custodians of collective identity. The taxonomy recognized eight categories of individual risk and missed an entire class of collective harm.

First: The taxonomy doesn’t recognize collective harm

When the British Museum published promotional images in January 2026 featuring an AI-generated character examining artifacts while dressed in outfits derived from the cultures those artifacts represented, no individual rights were violated in any legally cognizable sense. The harm was collective: an institution already dogged by accusations of cultural theft had used AI to automate cultural appropriation. The museum complied with community demands, apologized, and announced it would develop internal guidelines. Reactive governance, after the fact, because no binding framework required anything else.

When Christie’s, an auction house whose sales shape what enters and exits the cultural commons, held its first AI art auction in March 2025, the sale was legally unimpeachable. Over 6,000 artists signed a petition demanding cancellation, objecting to models trained on their work without consent.

These incidents don’t register as risks the Act was designed to address. The Act’s concept of harm requires an identifiable individual, an adverse decision affecting their interests, a traceable mechanism of injury. Collective harm to community trust is structurally invisible to this framework.

This is not a marginal omission. Cultural institutions are custodians of collective meaning, cultural memory, and democratic participation in the fullest sense Hannah Arendt meant when she described the public realm as constituted through shared human action (Arendt, 1958). The Act positioned those decisions outside its accountability framework. They belong inside it.

Second: The deployer-provider split misallocates accountability

Cultural institutions are almost never AI providers in the Act’s sense. They are deployers, organizations using foundation models built and maintained by large technology companies. The General Purpose AI (GPAI) provisions in Chapter V impose obligations on providers: copyright compliance, training data transparency summaries, systemic risk assessment for the most capable models (European Parliament, 2024).

The deploying museum inherits none of those obligations for minimal-risk deployments. The result is an accountability gap the Act does not resolve: when a cultural institution deploys a generative AI tool and that tool produces outputs trained on artists’ work without consent, the museum bears the full reputational and community cost of the deployment. The provider bears the formal regulatory obligation. The accountability sits with OpenAI. The legitimacy damage lands on the Rijksmuseum.

UNESCO’s 2021 Recommendation articulated the correct principle: responsibility for AI systems should rest with actors corresponding to their role in the life cycle (UNESCO, 2021, paras. 42-43). The EU Act, in limiting the most stringent chain-of-accountability provisions to high-risk systems, functionally exempted the cultural deployer from the accountability framework UNESCO told EU governments to build.

Gabriel Menotti’s argument that neural networks function as institutional archives — accumulating and remixing cultural materials at scale through the same extractive logic that colonial-era museums used on physical artifacts — lands with particular force here (Menotti, 2025).

When a museum integrates a foundation model for visitor engagement, it is entering into a partnership with an entity whose foundational act mirrors the institution’s own historical complicity in cultural extraction.

Many cultural institutions have spent decades reckoning with provenance, repatriation, the ethics of what they hold and how they acquired it. The Act does not require them to apply that same reckoning to what they deploy.

Those GPAI obligations flow from provider to downstream developer. They do not flow to the community standing in front of the exhibit. A museum has no regulatory leverage and no contractual right to demand meaningful training data transparency from a foundation model provider. Whatever the provider publishes, the institution cannot verify. If the community asks, the institution cannot answer.

The litigation landscape confirms the harm caused by this structural fault. Over 70 copyright infringement lawsuits have been filed against AI providers since 2023. Artists, publishers, record labels, and news organizations are pursuing the companies that built and trained models on their work without consent. None of that litigation reaches the cultural institutions deploying those models. The accountability chain stops at the provider, in court just as in the Act.

Article 25 addresses responsibilities along the AI value chain, but its provisions only engage at the high-risk threshold (European Parliament, 2024). The institution that made the community-facing decision is the party with no mandatory accountability for what follows.

Third: Lifecycle governance stops short

For high-risk systems, Articles 9 through 15 create comprehensive lifecycle governance requirements. Risk management systems must be established, implemented, documented, and maintained not just at deployment but throughout the system’s operational life. Human oversight must be baked in, logs must be kept, post-market monitoring must be continuous (European Parliament, 2024).

The Act’s lifecycle governance requirements embody a clear principle: if this system matters enough, you are accountable for it indefinitely. But cultural institutions’ deployments don’t trigger those requirements.

In a previous piece I documented the tendency to treat AI deployment as an event rather than infrastructure. The Act offers no corrective for institutions operating below the high-risk threshold. The Council of Europe acknowledged in March 2025 that cultural institutions lack the infrastructure for adequate AI governance (Council of Europe, 2025). The OECD AI Principles commit signatories to accountability over time (OECD, 2019). The Act itself recognized lifecycle governance as the minimum condition for responsible deployment, but drew the threshold in a place that leaves most cultural deployments on the wrong side of it.

This is increasingly visible to practitioners working across AI governance. There is a recognized shift underway from principle-based frameworks toward operational governance: deployable oversight, auditability requirements, impact assessment processes that keep pace with deployment. UNESCO and the OECD established the principled foundations. The binding framework that should have operationalized three generations of principled scaffolding instead granted cultural institutions an exemption.

Three interventions

These faults are architectural. They require structural responses.

The first is taxonomic. Article 7 lets the Commission add new high-risk use cases, but only within the eight areas Annex III already lists. Adding a new area takes an amendment to the Act itself, and Article 112 provides the route: the Commission must review the Annex III list every year, and report by August 2028 on whether new areas should be added (European Parliament, 2024). The evidence is building: the British Museum and Christie’s episodes show the kind of collective harm at stake, and the sector’s own surveys show most institutions have no AI governance policy to prevent it (Chang and MacDonald, 2025). Adding a domain, public-facing AI deployed by publicly-funded cultural and heritage institutions in their capacity as custodians of collective identity, would trigger proportionate lifecycle governance obligations without imposing enterprise compliance costs. The Commission’s guidelines on the practical application of the high-risk classification, expected in 2026, represent an opportunity to name this fault formally; the Article 112 review is where the case for a new area should be made.

The second is procedural. For GPAI deployments in cultural, educational, and public heritage contexts, deployer-side community impact assessments should be mandated as a condition of public procurement eligibility. This closes the provider-deployer accountability gap without requiring deployers to possess technical expertise they don’t have. A community impact assessment asks institutional, not technical questions: Whose materials are plausibly implicated in this model’s training data? What consent processes existed? What community input shaped the deployment decision?

The third follows directly from the lifecycle governance gap, and it is the most politically durable of the three. Public funding frameworks for AI adoption in cultural institutions should mandate lifecycle governance planning as a precondition for deployment. Before a grant is awarded or a procurement contract signed, applicants should demonstrate who will own this system, who will maintain it, how it will be monitored for drift, and how it will be responsibly retired. The Act created this requirement for high-risk systems because its architects understood it was the minimum condition for responsible deployment. The same logic applies wherever public institutions are deploying AI with public money.

This intervention matters because it routes around the political conditions that make the first two difficult. UNESCO’s framework assumed member state governments would build the scaffolding. The more likely outcome is 27 divergent national approaches, or more accurately, 27 divergent non-approaches, since the recommendation is non-binding and enforcement is nonexistent. Governments navigating austerity and culture wars are not, in practice, going to prioritize governance infrastructure for museums. Creative Europe and structural funds flow through EU mechanisms that operate independently of whether any particular member state government is culturally engaged or politically willing. That’s the mechanism that survives hostile political conditions.

What to do while we’re waiting for the fixes

The EU AI Act’s risk taxonomy is genuinely innovative, but its focus on individual harm will fail to predict or prevent collective harm in public-trust deployments. It will do so precisely where the relationship between AI governance and democratic participation is most at stake.

Article 112 obliges the Commission to keep revisiting Annex III. The evidence is accumulating.

Any amendment process will be slow. The governance vacuum is immediate.

The cultural sector has the knowledge to act. The early AI missteps we’ve seen follow predictable patterns that experienced practitioners and governance-literate institutions have already mapped.

What is missing is collective action: shared standards developed by and for cultural institutions.

That means governance playbooks that travel across borders in the way that individual institutional good practice currently doesn’t, sector-led frameworks that create the evidence base for the Annex III amendment argument and demonstrate the governance capacity that makes it harder to ignore. The regulatory fix is worth fighting for. While that fight proceeds, the sector should build what governments have failed to provide.

References