The EU AI Act's Transparency Rules Explained: What Article 50 Actually Requires

If you've noticed several major AI companies rolling out watermarking and disclosure features within days of each other, that's not a coincidence, it's a regulatory deadline. Here's what Article 50 of the EU AI Act actually requires, in plain language.

TL;DR: Article 50 of the EU AI Act requires AI systems to mark generated content in a machine-readable way, so it can be identified as AI-made, and requires certain AI-generated content (like deepfakes) to be disclosed to users. Its core transparency obligations took effect on August 2, 2026, which is why Anthropic, Google, and Meta all confirmed compliance measures around that date. Penalties for non-compliance can reach into the tens of millions of euros or a percentage of global annual turnover, and businesses building products on top of AI models carry their own compliance responsibility, separate from the underlying model provider's.

What Article 50 actually requires

Article 50 is the transparency section of the broader EU AI Act, and its core requirement is straightforward in principle: content generated or substantially modified by AI needs to be marked in a way that lets people and systems identify it as AI-made. In practice, this has taken the form of machine-readable watermarks in generated text and signed provenance metadata in generated files like images, the approach Anthropic detailed in its recent announcement, covered in our post on Claude's new watermarking system. The article also covers disclosure requirements for certain categories of AI-generated content, like synthetic audio, video, or images that could be mistaken for real depictions of people or events, often referred to as deepfakes.

Why this became urgent in early August 2026

The EU AI Act rolled out obligations in phases rather than all at once. The core transparency and high-risk system obligations under Article 50 took effect on August 2, 2026, which created a hard compliance deadline. Anthropic, Google, and Meta have all confirmed participation in the EU's associated Code of Practice on Transparency of AI-Generated Content, a voluntary framework companies can adopt to demonstrate compliance with the underlying legal requirements. That's why watermarking announcements clustered so closely together around the same date.

What's at stake for non-compliance

The penalties attached to the EU AI Act are substantial enough to function as a genuine forcing function rather than a soft guideline. Depending on the specific violation, fines can reach into the tens of millions of euros or a percentage of a company's global annual turnover, whichever is higher, a structure similar to how GDPR penalties work. For large AI companies operating globally, the practical result is that compliance measures built for the EU tend to get rolled out worldwide rather than region-locked, since separating EU and non-EU content streams is often more complex and risky than applying one global standard.

What this means if you're building on top of AI models

This is the part that gets missed in a lot of the coverage: compliance responsibility doesn't stop with the model provider. If you're a business using Claude, Gemini, or another AI model inside your own product, you're generally responsible for independently assessing what Article 50 requires of your specific product or service, not just relying on the underlying model provider's compliance. That can mean needing to disclose AI involvement in your own product that you previously didn't, even if the AI provider you're using has already implemented watermarking on their end.

What to actually do about it

1. Check whether your product falls within scope. Article 50's requirements are most directly relevant to products that generate or substantially modify content presented to end users, so audit your specific use case rather than assuming it does or doesn't apply.
2. Don't rely solely on the underlying model's watermarking. As covered in our post on Claude's watermarking rollout, the mark can be imperfect or absent depending on how content is edited or combined, so it isn't a substitute for your own disclosure practices where required.
3. Watch for guidance from your specific AI providers. Expect further documentation and detection tooling to roll out over the coming months as labs continue refining their compliance approach.
4. Consult legal counsel for your specific situation. This overview is general information, not legal advice, and the practical requirements for your product depend on factors specific to your business and jurisdiction.

FAQ

Does Article 50 only apply to companies based in the EU? No. It applies to AI systems and products offered to users in the EU, regardless of where the company is headquartered, which is part of why global companies like Anthropic, Google, and Meta have all had to respond.

Is watermarking the only requirement under Article 50? No. Watermarking and machine-readable marking address the labeling requirement specifically; the article also includes disclosure obligations for certain categories of synthetic content, like deepfakes, that go beyond simple technical marking.

What happens if a company doesn't comply? Penalties can reach into the tens of millions of euros or a percentage of global annual turnover, depending on the specific violation, which is a significant enough risk that most major AI companies have moved quickly to demonstrate compliance.

Does this affect AI models used only for internal, non-public work? The requirements are primarily focused on content generated for or presented to end users, so purely internal use cases may fall outside the most direct disclosure obligations, but this is exactly the kind of scope question worth confirming with legal counsel for your specific situation rather than assuming.

Are other countries introducing similar AI transparency rules? AI regulation is moving in a similar direction in several jurisdictions, though specific requirements and timelines vary significantly by country, so it's worth tracking developments in whichever markets your product actually operates in rather than assuming EU rules are the global standard.

Last updated: August 24, 2026. AI regulation is a fast-moving area; this post is general information, not legal advice, and specific compliance requirements should be confirmed with qualified legal counsel.