Update, : Since this article was published, the Digital Omnibus on AI was formally adopted and entered into force on 27 July 2026. Obligations for stand-alone high-risk AI systems (Annex III) now apply from 2 December 2027 rather than 2 August 2026. The Article 50 transparency obligations were not delayed: they have applied since 2 August 2026, so the disclosure and sign-off steps below are already in force. The rest of the article appears as originally published.
The EU AI Act’s full enforcement regime applies starting 2 August 2026, including for any museums, galleries, archives, and heritage sites that have deployed a third-party AI system in a professional or public-facing context.
If your institution uses an AI tool for visitor engagement, collection search, accessibility, fundraising analytics…you are a Deployer. Deployers carry enforceable legal obligations and the fines are going to be brutal.
This piece covers what Deployer status actually means, your most likely risk zones, and what your institution should be doing right now. Everything you need to know, in the time it takes to eat a sandwich.
In the UK, Canada or the US? Don’t skip this article
Can EU regulatory authorities compel a Toronto museum to pay a fine? Probably not. But you’re still going to experience real changes:
- Contractual pressure from EU partners. If you co-produce, tour, or collaborate with EU institutions, AI Act compliance clauses are already appearing in partnership and vendor agreements. You could get locked out of deals or forced to retrofit compliance mid-relationship.
- Reputational and press exposure. If your AI tool causes harm to EU visitors or EU-origin communities, the story writes itself. Public tolerance for AI governance failures at cultural institutions is already very low.
- EU-based vendors could end service or change terms. If you’re using EU-based AI tools or platforms, those vendors are themselves subject to the Act and may impose compliance requirements downstream on you.
This is how GDPR played out. UK and US institutions that ignored it found themselves locked out of data-sharing arrangements with EU partners until they caught up.
Taking on the four steps I present later in this article will keep you in good standing with your EU partners, and honestly? They’re just good practice, even if your region doesn’t have similarly rigorous requirements.
How the EU AI Act defines “Deployer” and risk levels
The EU AI Act defines two key legal roles in the AI supply chain: Provider and Deployer. A Provider builds or sells the AI system. A Deployer puts it to work.
If your institution licensed a third-party chatbot, deployed an LLM-backed collection search tool, or integrated an AI accessibility product, you are the Deployer. The Act’s obligations attach to that role regardless of who wrote the code. Don’t assume that liability sits with the vendor, because that is not a legal position that will hold under scrutiny.
The Act organises AI systems into four risk tiers: Unacceptable, High, Limited, and Minimal. Not all museum AI sits in the same risk tier. A basic FAQ chatbot that gives opening hours and ticket prices is probably minimal risk; no direct regulatory obligations there, just voluntary codes of conduct.
What applies to every Deployer, regardless of risk tier: Article 50 transparency obligations. If your institution deploys AI in a public-facing context, you must disclose when content is AI-generated, label synthetic media, and ensure visitors know when they’re interacting with an AI system rather than a human. These obligations apply to all systems, from minimal to high risk, from 2 August 2026.
Some AI tools museums are deploying fall under high risk classification. I’ll tell you how to recognise those applications below.
Why risk is amplified in small to midsize institutions
Cultural institutions have compounding exposure factors that most other organisations deploying AI do not:
- Public trust mandate. When AI fails in a museum, harm transfers to communities, cultural memory, and institutional legitimacy.
- Multiple flavours of data. Visitor behaviour data, membership records, school group registrations, etc. The same personal data streams your team handles carefully under GDPR flow through AI tools with far less scrutiny.
- Limited legal and technical capacity. Small-to-midsize institutions frequently lack the in-house expertise to audit their vendors, and the false assumption that vendors handle compliance is endemic.
- Scope underestimation. AI that interfaces with visitors is making representations on behalf of the institution. The risk classification follows what the tool actually does, not what the vendor calls it or how good your intentions were when you deployed it.
What you need for day-to-day operations
You need up to four things, in order of difficulty. Steps 1–3 are good governance practice for all systems and legally required for high-risk ones. Step 4 is specifically required if you have high-risk systems.
- A documented approval process. You need to be able to demonstrate that AI-generated content was reviewed before going live. In practice for a small museum, you need a simple sign-off log. You can do this right now. Create a shared spreadsheet with these columns: content item, AI system used, whether it was fully AI-generated or AI-assisted, reviewer name, date reviewed, approved/flagged.
- A complete register of every AI tool currently in use. Check whether your marketing or fundraising team has bought something without central IT’s approval. Once you have the list, classify each tool by risk tier. This is a helpful risk classification guide, albeit from a vendor.
- An audit trail. If your AI system generates accessibility content or informs what a visitor is told about the collection, you must be able to reconstruct how an output happened. If you publish AI-generated text intended to inform the public, you must either disclose it as AI-generated or ensure it went through substantive human editorial review. Records must be retained for at least six months. Your sign-off log (from Step 1) is your evidence that the review happened and is half of what you need. The other half depends on whether your vendor gives you access to their logs, which most don’t by default. Go down your list of AI tools and check whether each contract includes a clause requiring the vendor to retain and provide you with access to system logs on request, going back at least six months.
If your contract doesn’t include a log access clause, ask your vendor for a data processing addendum; most will have one. If they won’t provide log access at all, treat that as a serious compliance risk. You have a decision to make.
- And possibly, a Fundamental Rights Impact Assessment. For high-risk AI systems specifically, Deployers must conduct a Fundamental Rights Impact Assessment (FRIA) before deployment under Article 27.
In a museum or heritage context, high risk applications include:
- Biometric systems that identify people from recorded footage
- AI used to evaluate learning outcomes in formal education programmes your institution runs
Many common museum tools like fundraising platforms, collection search, and chatbots sit in a grey zone. The risk classification depends on what the tool is actually doing with data.
If you think you need an FRIA, I recommend this guide, jointly produced by the Danish Institute for Human Rights and the European Center for Not-for-Profit Law. It takes you through the process step by step.
Don’t let this be you
You’re a Collections Manager at a mid-sized museum. Your digital team licenses an AI tool that auto-generates audio descriptions for visually impaired visitors. Descriptions go live with minimal curatorial review. One misattributes the cultural origin of a work held under contested provenance. A visitor from that community complains publicly.
A regulator comes knocking. They find that there’s no disclosure in place labeling the audio guide as AI-generated content, and you can’t produce any evidence that a human checked the output before it went live.
That’s an Article 50 violation. Fines go up to €15 million or 3% of total worldwide annual turnover.
Know that nothing is on fire, yet
You have time to get this right (2 months, 14 days, if you’re reading this on 19 May). You just need lightweight, sensible governance steps, and to introduce your team to habits like using the sign-off log.
Realistically, this is going to play out exactly like GDPR. Museums will be invisible on the enforcement radar for the first few years, while enforcement goes after a few high-profile cases.
But use this window wisely. The better informed you are, the less likely you’ll get into hot water by signing a contract with the wrong vendor.
22 May update
The European Commission published draft guidelines on high-risk classification on 19 May 2026, with a public consultation open until 23 June 2026.
You may have heard about the Digital Omnibus Regulation, which pushes back the enforcement timeline for high-risk AI systems to 2 December 2027 at the earliest.
Your core Deployer classification and transparency obligations are unchanged. The Omnibus buys time on high-risk compliance specifically, but it does not remove the requirement, and it hasn’t been formally adopted yet. Don’t build your planning around a delay that isn’t final.
There’s a next level of strategic and reputational steps your institution should be taking alongside these compliance obligations, in the context of cultural sector lobbying to the European Commission. See my latest Policy Diary →
