For most of the last two years, everyone in the AI industry treated August 2, 2026 as one monolithic deadline — the day the EU AI Act's toughest rules would finally bite. Then the Digital Omnibus simplification package split that date in two, and a lot of companies read the headlines about "deadlines being delayed" and quietly stood their compliance programs down.
That was a mistake. The single most broadly applicable obligation in the entire Act — Article 50 — was not delayed. Neither was the power to fine you for breaking it.
What got delayed, and what didn't
The Digital Omnibus, on which the European Parliament and Council reached provisional agreement on May 7, 2026, deferred the headline high-risk regime. The standalone high-risk systems listed in Annex III — the category most enterprises were racing to certify — moved from August 2, 2026 all the way to December 2, 2027. That is a genuine sixteen-month reprieve, and it is the deadline most people meant when they talked about "the August 2026 deadline."
But Article 50 sits in a class of its own. It is not a high-risk provision and it is not tied to Annex III. It is a horizontal transparency obligation that applies to specific categories of AI regardless of risk classification — and it takes effect exactly as written on August 2, 2026.
The cleanest way to hold the distinction in mind: the part of the Act that was hardest to engineer for the largest number of companies — transparency — is the part that stayed on schedule, complete with fines.
What Article 50 actually requires
The duties split between providers (who build the systems) and deployers (who put them to use):
- Article 50(1) — AI interaction disclosure. Providers of systems meant to interact directly with people — chatbots, voice assistants, conversational agents — must design them so a user is told they're dealing with an AI, no later than the first interaction. The only exception is narrow: where it's already obvious to a reasonably observant person that they're talking to a machine.
- Article 50(2) — machine-readable marking. Providers of generative AI must mark their outputs — text, audio, image, video — in a machine-readable format that's detectable as artificially generated. The marking must be "effective, interoperable, robust and reliable as far as technically feasible."
- Article 50(3) — emotion recognition and biometric categorization. Deployers must inform the people subjected to these systems.
- Article 50(4) — deepfakes and public-interest text. Deployers who use AI to produce or manipulate a deepfake must disclose it's artificial. Deployers publishing AI-generated text on matters of public interest must do the same — unless a human reviewed the content and holds editorial responsibility for it.
Notice the structure: Article 50 creates a chain. Providers embed the marking at the point of generation under 50(2); deployers make it visible to the public at the point of publication under 50(4). A failure at either end breaks the chain, and both ends become enforceable on the same day.
The disclosure has to be real
The European Commission published its final guidance on these transparency requirements on July 20, 2026 — giving companies less than two weeks to get ready. And the guidance is blunt about what does not count.
A statement buried in a terms-of-service page does not satisfy Article 50(1). A metadata watermark on its own does not satisfy it. Calling your bot a vague "assistant" does not satisfy it. The information has to be perceivable in the interaction itself, at first contact. If you deployed a customer-service chatbot years ago and never thought of it as "AI Act scope," you are squarely in scope now.
The enforcement switch flips on the same day
Transparency rules on paper mean little without consequences, and this is the part the "everything's delayed" narrative most badly distorts. The Act's penalty provisions — Article 99 (penalties applied by member states) and Article 101 (fines on general-purpose AI model providers, levied directly by the Commission's AI Office) — become operative on August 2, 2026.
For breaches of operator obligations such as Article 50, the fine is up to €15 million or 3% of total worldwide annual turnover, whichever is higher. Supplying incorrect or misleading information to authorities carries its own tier of up to €7.5 million or 1%. In other words, August 2 is not merely when the transparency rules apply — it's when the regulator gains the standing to fine you for ignoring them.
The timeline at a glance
| Obligation | Who | Status on Aug 2, 2026 |
|---|---|---|
| Prohibited practices (Article 5) | All | In force since Feb 2, 2025 |
| GPAI model obligations | GPAI providers | In force since Aug 2, 2025 |
| Article 50 transparency duties | Providers & deployers | Applies — not deferred |
| Article 50(2) marking, systems already on market | Generative-AI providers | Transitional grace to Dec 2, 2026 |
| Article 50(2) marking, new systems | Generative-AI providers | Applies from Aug 2, 2026 |
| Fining power (Art. 99 / 101) | Enforcement | In force — not deferred |
| Annex III high-risk obligations | High-risk providers/deployers | Deferred to Dec 2, 2027 |
The one piece of breathing room inside Article 50 itself: generative systems already on the market before August 2, 2026 get a transitional window to December 2, 2026 to bring their machine-readable marking into conformity. Anything launched on or after August 2 must mark from day one. Treat that four-month window as a short bridge, not a reprieve.
Who needs to act now
The reach here is broader than the high-risk rules it's so often confused with, because Article 50 keys off what a system does, not how risky it is. If you're a generative-AI provider, a chatbot operator, a marketing or media team publishing synthetic images or video, or a newsroom running AI-assisted text on public-interest topics, none of your obligations benefit from the Annex III deferral. The editorial-review carve-out for public-interest text is meaningful — but you have to be able to document the human review to rely on it.
The Bottom Line
The Digital Omnibus did something genuinely useful: it gave the high-risk regime the runway it needed. But it did so without touching the two things that make August 2, 2026 a hard date for the widest set of organizations — the Article 50 transparency obligations and the enforcement power behind them. If your business talks to customers through a bot, generates synthetic media, or publishes AI-written content to EU audiences, the deadline did not move for you. Inventory your systems by Article 50 sub-paragraph, build first-contact disclosure into every conversational interface, and document your marking and review decisions — because the fines are now real, and they arrive on the same day as the rules.
This article is for general information and does not constitute legal advice.
