EU AI Act’s Transparency Rules Are Now Law. Here’s Exactly What Changed Yesterday

Yesterday was not the day the EU AI Act took effect. That day was August 2, 2024, when the regulation entered force after publication in the Official Journal. Yesterday was the day that companies operating in the EU stopped being able to describe their AI systems as compliant with EU regulation while ignoring the transparency requirements. As of August 2, 2026, every chatbot deployed in the EU must tell users it is an AI at first contact. Every AI-generated deepfake must carry a visible disclosure. Every AI-generated image, video, or audio clip from a new generative AI system must embed machine-readable watermarks. Every automated article published on a matter of public interest without a human editor’s review must display a disclaimer. The obligations are not aspirational. They are enforceable, with penalties reaching €15 million.

One distinction matters for reading coverage of this week’s enforcement milestone accurately: the obligations that took effect on August 2, 2026 are specifically the transparency requirements under Article 50 of the EU AI Act. The high-risk AI system obligations — covering AI used in employment, credit, education, biometrics, and essential services — were pushed back to December 2027 by the Digital Omnibus on AI, signed in July 2026. The transparency rules and the high-risk rules are separate tracks of the same regulation, and only one track became enforceable yesterday.

What Article 50 Requires

Chatbot and AI Interaction Disclosure

Article 50(1) of the EU AI Act requires that providers of AI systems intended to interact directly with natural persons design those systems so that users are clearly informed they are interacting with AI at the time of their first interaction. That obligation sounds simple; the Commission’s guidance on what it requires in practice is more demanding. A reference buried in terms and conditions is not sufficient. A metadata marker that users typically do not see is not sufficient. Vague references to an ‘assistant’ or technically accurate but obscure language like ‘this system uses LLMs’ are explicitly not sufficient. The draft guidelines published by the European Commission on May 8, 2026 specify that the disclosure must be ‘perceivable in the interaction itself’ — meaning visible, audible, or otherwise perceptible to the user at the point of first contact, in plain language that a reasonably informed person can understand. The requirements are stricter when the intended audience includes children, elderly people, or persons with disabilities. As the European Commission’s official AI regulatory framework makes clear, the only narrow exemption is where the AI character of the system is ‘obvious from the context and circumstances’ — and the Commission’s guidelines test that exemption against a reasonably well-informed, observant person from the actual audience, not against a technically sophisticated observer.

Deepfake and Synthetic Content Labelling

Article 50(3) and 50(4) require that deployers who use AI to create or manipulate images, audio, or video to produce realistic synthetic content — deepfakes — must disclose that the content has been artificially generated or manipulated. This obligation applies to deployers, not just providers: the company that uses a generative AI tool to produce a synthetic video must label it, regardless of whether the AI tool provider has itself already watermarked the output. The disclosure must be ‘clear and distinguishable’ and placed so that users can see or hear it at the point of engaging with the content. The Commission’s draft guidelines specify that the labelling should be persistent — not just shown at the start of a video and then disappear — and that the same disclosure requirement applies to synthetic audio, including voice clones, in addition to video.

AI-Generated Public-Interest Text

Article 50(4) extends the labelling requirement to AI-generated text published on matters of public interest — news articles, political commentary, policy analysis, public health information — where the content was generated with significant AI involvement and published without meaningful human editorial review and oversight. The carve-out for human editorial involvement is important: AI-assisted journalism that involves a journalist’s active review, editing, and editorial responsibility does not trigger the labelling requirement. The obligation targets specifically the case where an automated system generates and publishes content on public-interest topics without human editorial gatekeeping. Given the deployment of AI content generation tools across news publishing, political campaigning, and public health communication, this is one of the most commercially significant transparency obligations in Article 50.

Machine-Readable Watermarking

Article 50(2) requires providers of generative AI systems to embed machine-readable markers in their outputs — images, audio, video, and text — so that the AI origin of content can be detected by downstream systems. Visible labels and machine-readable markers are separate requirements: visible labels communicate AI origin to human viewers; machine-readable markers allow automated systems to detect AI content in contexts where human viewing is not involved, such as platform content moderation systems and search engine indexing. Providers whose generative AI systems were already on the EU market before August 2, 2026 benefit from a grace period until December 2, 2026 to implement the machine-readable marking capability — a concession the Digital Omnibus established to allow retrofitting rather than requiring system redesign. Systems newly placed on the EU market from August 2 must comply from that date without grace period.

What the Digital Omnibus Changed

The Digital Omnibus on AI, signed July 8, 2026 by the co-legislators and published in the Official Journal, made two significant changes to the EU AI Act’s implementation timeline that affect how August 2 should be understood. First, it pushed the enforcement date for high-risk AI systems in Annex III — standalone AI systems used in employment, credit, education, biometric categorisation, and essential services — from August 2, 2026 to December 2, 2027. For systems embedded in regulated products under Annex I, the deadline was extended further to August 2, 2028. These extensions, granted in response to sustained pressure from European industry, reduce the August 2 enforcement action from the comprehensive AI regulation milestone it was originally positioned as to a transparency-specific enforcement milestone. Second, the Digital Omnibus added a new prohibition to Article 5: AI-generated non-consensual intimate imagery is now explicitly prohibited under the EU AI Act, adding to the prohibited practices that have been in force since February 2025. As explored in our earlier coverage of the state-level AI legislation landscape across the US in 2026, the non-consensual intimate imagery prohibition addresses a category of AI harm that multiple US states have also moved to regulate, reflecting an emerging global consensus on this specific use case.

Technical Challenges: Watermarks That Can Be Stripped

The State of the Technology

The AI Weekly analysis of the August 2 obligations noted a tension the Commission acknowledges in its own guidance: the watermarking mandate has outpaced the technology it regulates. Watermarks can be stripped. Metadata can be lost in file format conversions. Machine-written text is notoriously difficult to detect reliably with current technology. The EU AI Office is developing technical standards for watermarking that are intended to be robust — resistant to stripping, interoperable across different platforms and tools, and reliable across content types — but those standards are not yet finalised. Companies required to embed machine-readable markers from August 2 are doing so against a technical specification that is still being developed, using current watermarking capabilities that the market has available rather than the future-proof standard the regulation anticipates.

The Label Inflation Problem

A separate concern raised by CCIA Europe’s AI policy lead and others is the risk of label inflation. The EU AI Act’s Article 50 obligations were originally designed to flag deceptive content — deepfakes intended to mislead viewers about events or people’s words. With the deceptive-intent test largely removed from the final text, the labelling obligation applies to AI-generated content broadly, not specifically to deceptive AI content. The practical concern is that if a stock photo produced with AI assistance carries the same visible disclosure label as a politically motivated deepfake of a politician saying something they never said, users learn to ignore the label — the same attention fatigue that makes cookie consent banners ineffective. Whether the AI content label becomes a meaningful signal or a meaningless ubiquitous badge depends significantly on how consistently and conspicuously it is applied across the enormous volume of AI-generated content entering the EU market.

Who Is Affected and How

Providers

Providers of AI systems — companies that develop and place AI products on the EU market — face the primary obligation under Article 50(2) for machine-readable marking. Any provider offering a chatbot, generative image tool, voice synthesis service, AI writing assistant, or AI video generation system to EU users or operators must design their systems to produce machine-readable marked outputs and to disclose the AI nature of the system at first user interaction. Non-EU providers whose systems are available to EU users are in scope: the EU AI Act applies to systems placed on the EU market regardless of where the provider is headquartered. This is the same extraterritorial reach that characterises GDPR, and it produces the same enforcement challenge: the EU can pursue providers in its own territory efficiently, but enforcement against non-EU providers depends on cooperation and market leverage.

Deployers

Deployers — companies that take AI systems provided by others and deploy them to end users — face the primary obligation under Article 50(3) and (4) for visible disclosure and deepfake labelling. A company that uses a third-party generative AI tool to produce content for its customers is a deployer and bears the obligation to ensure that content is visibly disclosed as AI-generated. Platform operators deploying AI-powered chatbots, HR technology using AI in candidate interactions, financial services using AI in customer communications, and news publishers using AI for content generation are all deployers with disclosure obligations that apply from August 2. The Commission’s guidance explicitly states that labels applied automatically by the AI platform provider do not substitute for the deployer’s own disclosure obligation — the deployer must independently ensure that users receive the required disclosure, not rely on the provider to have done it.

The Brussels Effect in Real Time

The enforcement of the EU AI Act’s Article 50 transparency obligations is an early test of whether the Brussels Effect documented in the Thomson Reuters Foundation’s analysis of nearly 3,000 global companies will materialise in the AI content disclosure domain specifically. The commercial logic of the Brussels Effect applies with particular force to generative AI content: companies that produce AI-generated images, audio, and video for global audiences face the same binary choice as companies that faced GDPR — maintain separate labelled and unlabelled content streams for EU and non-EU audiences, or apply the EU’s disclosure requirements globally and operate one content pipeline. For companies producing AI content at scale, the operational cost of separate pipelines typically exceeds the cost of universal disclosure. The result, as with GDPR, is likely to be widespread adoption of EU-style AI content disclosure globally, driven by operational economics rather than legal obligation. The intersection of EU regulatory obligations with the broader enterprise AI governance challenge is explored in our reporting on how the Anthropic export control episode shaped the regulatory relationship between AI developers and governments — in both cases, the regulatory framework shapes AI deployment practices beyond the jurisdiction that enacted the rule.

What Happens Next

The immediate practical consequence of August 2 is the activation of national market surveillance authorities’ enforcement powers. In each of the 27 EU member states, the designated authority can now initiate proceedings against providers or deployers of AI systems that fail to meet Article 50 obligations. The enforcement priority in the early months is expected to be visible, high-profile cases — a major platform’s chatbot that fails to disclose its AI nature, a widely circulated deepfake without required labels — rather than systematic screening of all AI content. That enforcement pattern is consistent with how GDPR’s early enforcement was handled: high-visibility cases establish precedent and signal that the obligation is real, before capacity is built for broader compliance monitoring. The December 2 deadline for existing generative AI providers to implement machine-readable marking gives a specific near-term milestone that affects every provider whose system was on the EU market before August 2. After December 2, the technical obligations fully apply across the market without grace periods.

Why It Matters

The EU AI Act’s Article 50 transparency obligations are the first enforceable global rules requiring that AI-generated content be disclosed to the humans who encounter it. The obligations are imperfect — the watermarking technology is still developing, the label inflation risk is real, and the exemptions create ambiguity that will generate enforcement disputes — but they establish a legal baseline that did not previously exist. Every deepfake circulating without a label in the EU is now an infringement, not just a concern. Every chatbot that allows users to believe they are talking to a human without disclosure is now a violation, not just a design choice. That legal baseline, however imperfect its initial technology and however variable its early enforcement, is the foundation from which more precise and technically mature requirements will develop over the next two to five years as the standards are finalised, the case law accumulates, and the enforcement authorities build the capacity for systematic oversight.

Sources

EU AI Act Article 50 full text (Regulation (EU) 2024/1689). European Commission draft guidelines on Article 50 scope and application, published May 8, 2026 (digital-strategy.ec.europa.eu). AI Weekly, July 31, 2026. Technology.org, July 17, 2026. Greenberg Traurig legal analysis, June 8, 2026. Bratby Law Article 50 guide, June 12, 2026. EU AI Act official transparency rules overview (artificialintelligenceact.eu). Digital Omnibus on AI, signed July 8, 2026.