The EU AI Act's Transparency Rules Are Live: What Article 50 Means for Your AI Products and Content

If you spent the summer reading that the EU had delayed its AI rules, you got half the story. Europe did push back the heaviest parts of the AI Act. It did not push back EU AI Act Article 50 — the transparency rules that govern chatbots, AI-generated images and video, deepfakes, and AI-written text. Those rules have applied since August 2, 2026, and they reach well beyond companies headquartered in Europe.

For a US SaaS company with European users, a brand generating marketing content with AI, or a startup shipping a generative feature, this is no longer a future compliance item. This guide explains what Article 50 requires, who it covers, the one deadline that still gives some breathing room (December 2, 2026), and how the rules connect to the IP questions you are already managing around AI.

What changed this summer — and what did not

The EU adopted a simplification package known as the Digital Omnibus on AI, which entered into force on July 27, 2026. Its headline effect was to postpone the AI Act’s high-risk system requirements: obligations for standalone high-risk systems now apply from December 2027, and those for AI embedded in regulated products from August 2028. That delay generated most of the coverage, and a lot of companies read it as a general pause.

The only relief granted on transparency was narrow: generative AI systems already on the market before August 2, 2026 have until December 2, 2026 to implement the machine-readable marking requirement. Everything else in Article 50 is live now, and systems launched on or after August 2 get no grace period at all.

The four Article 50 duties in plain English

Article 50 is short, but it creates four distinct obligations that fall on different parties. The first job is to work out which ones apply to you.

Obligation	Who it falls on	What it requires
AI interaction disclosure	Providers of AI that talks to people (chatbots, voice assistants)	Tell users they are dealing with an AI system unless that is obvious from context
Machine-readable marking	Providers of generative AI (text, image, audio, video)	Embed detectable markings so synthetic output can be identified as AI-generated
Emotion recognition / biometric categorisation notice	Deployers of those systems	Inform the people exposed to the system
Deepfake and public-interest text disclosure	Deployers publishing the content	Disclose deepfakes and AI-generated text on matters of public interest, unless a human editor has reviewed it and taken responsibility

Two terms do the heavy lifting. A provider is the party that develops an AI system or places it on the market under its own name or trademark. A deployer is the party using the system in a professional capacity. Many businesses are both: if you put your brand on a customer-facing assistant built on a third-party model, you are likely acting as a provider for that assistant, and if your marketing team publishes AI-generated video, you are a deployer of the tool that made it.

Yes, this applies to US companies

The AI Act applies to AI systems placed on the EU market and also to providers and deployers located outside the EU where the system’s output is used in the Union. A US company whose customer-facing chatbot or content-generation tool reaches EU users is within scope today — no European subsidiary, server, or office required.

That extraterritorial reach is the same pattern that made the GDPR a global compliance issue. If your product has meaningful European usage, assume Article 50 applies and work backwards from there.

The guidance and the Code of Practice

Two implementation documents arrived in July 2026. On July 20, the European Commission adopted non-binding interpretive guidelines on Article 50. On July 31, the EU AI Office published a voluntary Code of Practice on Transparency of AI-Generated Content, covering the technical marking obligations and deployer labelling. Neither is legally binding, but both are the clearest available signal of how regulators will read the rules, and signing or aligning with the Code is likely to be the most defensible posture if a national authority comes asking.

Expect some variation: enforcement will run through multiple national authorities, and their interpretations may not perfectly align. Documenting how you reached your compliance decisions matters as much as the decisions themselves.

Where Article 50 meets your IP strategy

Article 50 is a transparency rule, not a copyright rule. But it lands in the middle of the IP questions every AI-using business is already working through, and it is worth connecting the dots.

Labelling and ownership pull in the same direction

In the US, purely AI-generated output generally cannot be protected by copyright without sufficient human authorship — a line the Supreme Court left in place when it declined to hear Thaler v. Perlmutter in 2026, as we covered in AI and Your IP in 2026. The record-keeping you need to show which content was AI-generated for Article 50 purposes is the same record-keeping that tells you which assets you can actually claim to own. Build one system and use it for both.

Training data rules diverge sharply across borders

The rules on training AI with copyrighted material now differ in a way that matters for any company operating in more than one market:

• United States: courts have found that training on lawfully acquired works can be fair use, while pirated sourcing carries serious liability. The analysis is fact-specific and still evolving.
• European Union: commercial text-and-data mining is permitted unless the rightsholder has opted out, and general-purpose AI model providers have carried obligations on copyright policy and training-content summaries since August 2025.
• United Kingdom: in a report published on March 18, 2026, the UK government abandoned its previously preferred option of a broad commercial text-and-data mining exception with an opt-out. The existing exception remains limited to non-commercial research, so commercial training on protected UK content still generally requires a licence, with further policy decisions deferred.

The practical upshot: a training-data strategy that is defensible in the US may not be defensible in London or Brussels. If you train or fine-tune models, map your data sources against each market you serve.

Your Article 50 action plan before December 2

Most companies can get to a defensible position in a focused sprint. Work through these steps in order.

• ☐  Inventory every EU-facing AI touchpoint — chatbots, voice agents, content generators, image and video tools, and any internal tool whose output gets published externally.
• ☐  Assign a role to each system — provider, deployer, or both — and record the reasoning.
• ☐  Confirm conversational interfaces disclose that they are AI at the start of the interaction, unless that is genuinely obvious from context.
• ☐  Verify machine-readable marking on generative outputs, either natively or through your model vendor, and get written confirmation of what the vendor actually provides.
• ☐  Add deepfake and public-interest labelling to your publishing workflow, and document where human editorial review applies.
• ☐  Update vendor contracts to allocate responsibility for marking, detection, and cooperation with regulators.
• ☐  Align with the Commission guidelines and the Code of Practice, and decide whether signing the Code makes sense for your business.
• ☐  Tie the AI-content log to your IP records, so the same data tells you what must be labelled and what you can claim to own.

The bottom line

The EU’s AI rulebook did slow down this summer, but not where most companies interact with it. Article 50 is in force, it applies to non-EU businesses whose AI output reaches Europe, and the last grace period closes on December 2, 2026. Treat it as a cross-functional project: product owns disclosure, engineering owns marking, marketing owns labelling, and legal owns the record that ties it all to your IP position. If you need help mapping your AI products against Article 50 and the diverging US, EU, and UK training-data rules, that is exactly the kind of cross-border review we run for clients.

---

Disclaimer: This article provides general information about IP strategy and should not be relied upon as legal advice. IP laws vary significantly by jurisdiction and every business situation is unique. Consult qualified counsel about your specific circumstances.