The AI Office Can Fine You Now

A lot of people read the EU AI Act delay headlines this summer and concluded the whole thing slipped a year. That’s not what happened.

The delay applies to one tier of the Act, the high-risk use cases in Annex III: hiring tools, credit scoring, medical devices, and a defined list of similarly sensitive categories. General-purpose AI is a separate track, and it did not move. The AI Office’s enforcement power over providers of general-purpose AI models activated on schedule on 2 August 2026. Fines under that power reach up to 3% of global annual turnover or 15 million euro, whichever is higher. The underlying obligations were not new that day, they had already been in force since 2 August 2025 for new models. What changed on 2 August 2026 is that the AI Office can actually act on them, not just note that they exist.

The same day, Article 50’s transparency duties took effect as planned, with one narrow exception: generative AI systems already on the market before 2 August 2026 have until 2 December 2026 to add machine-readable marking to their output. Everything else in Article 50 applies now, and it is the part of the Act most likely to touch a small team’s actual product, whatever tier the rest of your system lands in.

Provider or deployer, and why it matters

A GPAI provider is the organization that develops a general-purpose model and places it on the market, the foundation model maker. A deployer is anyone using an AI system in the course of their own operations, which covers most teams calling somebody else’s API to build a feature. The 3% fine is aimed at providers. If your product calls Claude, Gemini, or any other foundation model through an API rather than training and distributing your own, that specific number is not really about you, you’re a deployer of somebody else’s GPAI model.

Article 50 does not respect that distinction the same way. If your product talks to a user, generates content, or produces synthetic media, the transparency duties land on you whether you built the underlying model or you’re three API calls deep into someone else’s.

Who the 3% number is actually aimed at

“Whichever is higher” between 3% of global annual turnover and 15 million euro is worth reading twice, because it doesn’t mean the same thing for every provider. For a genuinely large foundation model maker, 3% of global turnover can run into the billions, so the fine scales with the size of the company being fined. For a small provider, 3% of a modest revenue number is usually well under 15 million euro, so the flat 15 million figure is the one that applies. The formula is built to keep the largest providers from treating a fine as a rounding error. A team small enough to worry about the 15 million floor at all is, almost by definition, not the target this clause was written for, though the transparency duties below apply regardless of size.

What Article 50 actually asks for

Three things, none of them exotic:

• Chatbot disclosure. If someone is talking to an AI system, they need to know that clearly, before or as they start, not three paragraphs into a terms-of-service document. A support widget styled to look like a live human agent needs a label before the first message, not a disclaimer buried in a footer link.
• AI content marking. Content your system generates needs a marker identifying it as artificially generated, wherever that’s technically feasible. A tool that auto-drafts marketing copy or blog posts needs that marker attached before the output leaves your system, not added later if someone asks.
• Deepfake labelling. Image, audio, or video content that could pass as a real person, place, or event needs a clear label saying it was generated or manipulated. A face-swap or voice-clone feature needs that label the moment its output could be mistaken for real footage.

None of these require a legal team to interpret on their own. They require you to actually open your own product and check whether the disclosure exists anywhere a user would notice it, and whether the marker gets attached every time, not sometimes.

What to do this quarter

Walk your own product against that list before a regulator, a journalist, or an annoyed user does it for you. If you’re a deployer calling an existing model, that walkthrough is most of the actual work, an afternoon with your own product open in a browser tab. If you’re a provider of your own general-purpose model, the fine is now a live risk rather than a future one, and that changes how urgently you should be documenting what your model does and does not do.

Keep a simple record of what you checked and when, even a dated note in your own repo is better than nothing, because “we looked at this in November 2026” is the kind of statement you’ll want to be able to make later if anyone ever asks.

A scanner can help with part of this and no more than that part. AIVory Guard runs inside the JetBrains IDE and can flag the code-level evidence, a call to a model with no disclosure step anywhere near it, a content-generation path with no marker attached before the response goes out. It does not produce your technical documentation, and nothing in this post is legal advice. If any of this applies to you at real scale, get an actual lawyer, not a blog post.

The delay news made it easy to relax about the whole Act this summer. The enforcement mechanism didn’t take the summer off.