EU AI Act Article 50, Explained for Businesses

Article 50 of the EU AI Act is the transparency rule. It says providers and deployers of generative AI must mark synthetic output so it can be identified as machine-made, and must tell people when they are dealing with AI.

The part that surprises most business owners is reach. Article 50 is not limited to companies based in Europe. If your AI output lands in front of EU users, the rule can apply to you, even from the United States or the United Kingdom.

This guide explains what Article 50 says in plain English, who it covers, whether it touches US and UK firms, what it makes you do, why Claude now watermarks its text, and the timelines and penalties. It is general information, not legal advice.

What is Article 50 of the EU AI Act?

Article 50 of the EU AI Act is the article that sets transparency obligations for certain AI systems. It is the part of the law about being open when content or a conversation involves AI.

The rule has two main jobs. It makes AI systems disclose when a person is interacting with a machine. It also makes providers of generative AI mark their synthetic output so it can be detected as artificially generated.

Article 50(2) is the marking duty. It says the output of a generative AI system must be marked in a machine-readable format and be detectable as artificially generated or manipulated. Machine-readable means a computer can spot the mark, not just a person.

The law names several ways to do this. Recitals list watermarks, metadata, cryptographic methods for proving provenance, logging, and fingerprints as acceptable techniques. The point is that AI output should carry a signal a system can read later.

• Interaction disclosure: tell users when they are talking to an AI, such as a chatbot.
• Synthetic-output marking: mark AI-generated text, image, audio, and video so machines can detect it.
• Deepfake and public-interest text labels: disclose certain AI-made images, audio, video, and news-style text.
• Machine-readable: the mark must be readable by software, not only visible to a human.

Who does Article 50 apply to?

Article 50 applies to providers and deployers of specific AI systems whose output or interaction reaches people. A provider builds or supplies the AI system. A deployer uses it in the course of business.

Providers carry the marking duty. If you make or supply a generative AI system, you must mark its synthetic output as machine-generated in a machine-readable way. This is the obligation Article 50(2) places on the model maker.

Deployers carry disclosure duties. If you run an AI chatbot, you must tell users they are talking to a machine. If you publish deepfakes or AI-generated text on matters of public interest, you must disclose that too, with some exceptions the law spells out.

So both sides of the chain are covered. The company that builds the model and the company that puts it in front of customers each have their own part to do. Many businesses are deployers without realizing it.

• Providers: must mark synthetic output as machine-generated and detectable.
• Deployers of chatbots: must disclose that users are interacting with AI.
• Deployers of deepfakes: must disclose AI-generated or manipulated images, audio, and video.
• Deployers publishing AI text on public-interest matters: must disclose it, subject to the law's exceptions.

Does the EU AI Act apply to US and UK companies?

Yes. The EU AI Act can apply to US and UK companies when their AI output is used in the EU. The law has extraterritorial reach, so where you are based matters less than where your output lands.

The test is not your headquarters. The test is whether the AI system's output is put into service or used inside the EU. If EU residents receive your AI-generated content or use your AI feature, Article 50 can reach you.

Picture a US SaaS company with an AI writing feature used by customers in Germany and France. Its AI output reaches EU users, so the transparency duties can apply. A US address does not create an exemption.

Picture a UK marketing agency that publishes AI-generated articles and images to an audience across the EU. Because that synthetic content is consumed in the EU, the marking and disclosure expectations can follow it.

This mirrors how the EU's data law, the GDPR, reached companies worldwide. If you serve EU users, EU rules tend to come with them. The safe assumption for any company with EU customers is that Article 50 is in scope.

• The trigger is EU use of the output, not the company's location.
• A US SaaS tool serving EU customers can be in scope.
• A UK publisher or agency reaching EU audiences can be in scope.
• Having no EU office does not remove the obligation.

What Article 50 requires you to do

Article 50 requires you to make AI use visible in two directions: to the machines that process content and to the people who receive it. The duties differ by role, but the theme is the same.

If you are a provider of generative AI, mark synthetic output in a machine-readable way. That means embedding a signal a system can detect, such as a watermark or signed provenance metadata, so the content reads as artificially generated.

If you run a chatbot or voice assistant, disclose the AI. Users should know they are interacting with a machine, in a clear and timely way, unless it is already obvious from the context.

If you publish deepfakes or AI-generated text on matters of public interest, label them. The law expects a disclosure that the image, audio, video, or article was artificially created or manipulated, with narrow exceptions.

Keep it practical. The law asks for solutions that are effective, interoperable, robust, and reliable as far as is technically feasible. It does not demand perfection, but it does expect a real, current effort.

• Mark AI-generated output so software can detect it (the provider duty).
• Disclose AI interaction to users of chatbots and assistants.
• Label deepfakes and, where required, AI-generated public-interest text.
• Use current, state-of-the-art techniques such as watermarks or provenance metadata.

Why Claude now watermarks its output

Claude now watermarks its output because Anthropic added text watermarking and C2PA provenance to meet Article 50(2)'s machine-readable-marking duty. Anthropic announced the change on August 11, 2026.

Two techniques do the work. Claude embeds an invisible, machine-readable watermark into generated text by biasing token choice during sampling. For supported files such as .png, .jpg, and .svg, it attaches signed provenance metadata built on the C2PA standard.

The marking is set at the model level, so no product surface can switch it off. It covers Claude models released on or after August 2, 2026 across the API, claude.ai, Claude Code, and other surfaces, for free and paid users alike.

The reach is global, not EU-only. Article 50 was the trigger, but Anthropic chose to apply the marking worldwide, so every user everywhere gets marked output. We break down the mechanics in our Claude watermark explainer and the wider AI watermarking guide.

One caution matters for your team. A watermark shows text may have been processed by Claude, not that AI authored it. People use Claude to proofread, translate, and reformat human writing, and that gets marked too, so a mark is not proof of AI authorship. Public detection tooling is still forthcoming, so you cannot verify the text mark yourself yet.

• Text: an invisible, machine-readable watermark added during generation.
• Files: cryptographically signed C2PA provenance metadata on supported formats.
• Applied at the model level, so no surface can turn it off.
• Rolled out globally, driven by Article 50 but not limited to the EU.

Timelines and penalties

Article 50's transparency obligations became enforceable on August 2, 2026. The Commission's Code of Practice on transparency of AI-generated content took effect the same day and is recognized as an adequate way to show compliance.

There is a short transition for existing systems. Under the AI Omnibus provisional agreement of May 2026, generative AI systems already on the market before August 2, 2026 have until December 2, 2026 to meet the machine-readable marking duty. Verify how this applies to your systems.

Penalties are set by Article 99, not Article 50 itself. Breaches of the Article 50 transparency obligations sit in the tier of up to €15 million or 3% of total worldwide annual turnover, whichever is higher. Confirm the current figures against the primary source before you rely on them.

That tier is below the top band. The highest fines, up to €35 million or 7% of turnover, are reserved for banned practices under Article 5, not transparency breaches. Even so, the transparency tier is large enough to take seriously.

• August 2, 2026: Article 50 transparency obligations enforceable; Code of Practice in effect.
• December 2, 2026: transition deadline for pre-existing generative systems to meet marking (verify).
• Penalty tier: up to €15 million or 3% of global annual turnover, whichever is higher (Article 99).
• SMEs and startups: the fine is capped at the lower of the amount or the percentage.

What businesses should do now

Start by finding where you use generative AI, because you cannot disclose what you have not mapped. Most compliance gaps we see are not refusals to comply. They are AI usage nobody wrote down.

Across the content-automation routines we run on our own portfolio, the pattern is consistent: AI shows up in more workflows than the team expects, from support replies to product copy to images. The first fix is always an inventory, not a policy.

Then confirm your vendors mark their output, add clear user-facing AI disclosures where you deploy chatbots or publish synthetic content, and keep records of provenance. Signed provenance metadata is the file half of this, and we cover it in the C2PA content provenance guide.

Use the table below to sanity-check common situations. It is a rule of thumb, not legal advice, and edge cases need a professional. When you want the maps kept current as your stack changes, that is the kind of AI-disclosure and provenance wiring we build in an AI workflow audit.