The European Union’s rules on labelling AI-generated content stop being guidance and start being law on August 2, 2026. From that date, companies that build generative AI systems must mark the images, audio, video and text those systems produce so the output can be detected as artificial, and companies that publish deepfakes or AI-written text on matters of public interest must disclose that to the people who see it.
The duties land on two groups at once: the providers that develop and supply the systems, and the deployers that use them in their own products, services and publishing. Geography offers no shelter. A provider or deployer established anywhere in the world is covered whenever its AI system’s output is used inside the EU, according to the Commission’s questions and answers on the rules. Breaches carry fines of up to €15 million or 3% of worldwide annual turnover, with proportionality applied to smaller firms.
Enforcement runs mainly through national market surveillance authorities in each member state rather than through Brussels. The EU’s AI Office keeps a narrower role, covering systems built on general-purpose models by the same company and systems embedded in the very large platforms and search engines already designated under the bloc’s content rules. The European Data Protection Supervisor polices AI used by the EU’s own institutions.
Two days before the deadline, the Commission reported that about 190 organisations had signed the voluntary code of practice built to demonstrate compliance with the marking and labelling duties. Eighty-three signed the providers’ chapter, including Anthropic, Google, Meta, Microsoft (MSFT ), Mistral, OpenAI, Cohere, Aleph Alpha, Black Forest Labs and Synthesia. A further 152 signed the deployers’ chapter, a list that runs well outside the AI industry: Getty Images (GETY ), Lenovo, Lufthansa, Iberdrola and the jeweller Bulgari among them. About half the signatories are small and recently founded companies, the Commission said.
What the rules require
The transparency tier of the EU AI Act, the bloc’s 2024 law that regulates AI systems according to the risk they pose, sets out four distinct duties:
- Disclosure on interaction. Chatbots, AI agents and avatars must be designed so people are told they are dealing with a machine from the start of the first exchange, unless that is already obvious to a reasonably observant person.
- Machine-readable marking. Synthetic audio, image, video and text must carry marks that are effective, robust and interoperable enough for the content to be detected as AI-generated or manipulated downstream.
- Visible deepfake labels. Deployers must disclose AI-generated or manipulated images, audio and video that resemble real people, places or events and would falsely appear authentic. A machine-readable mark alone does not satisfy this; the disclosure has to be perceivable without special tools.
- Labels on public-interest text. AI-generated or AI-altered text published to inform the public on politics, public health, justice, security, the environment, consumer safety or comparable subjects must be labelled, unless a person reviewed the substance and someone holds editorial responsibility for it.
The exemptions matter as much as the duties. Personal, non-professional use falls outside the law entirely, so an individual posting an AI-made image sits beyond its reach while an advertising agency doing the same does not. Deepfakes inside evidently artistic, satirical or fictional works get a lighter form of disclosure that must not spoil the work. Spellchecking, grammar correction and other assistive editing do not trigger marking, and short strings, source code, machine-to-machine outputs and intermediate steps in closed production pipelines are out of scope.
Existing systems get a narrow reprieve. Generative AI systems already placed on the EU market before August 2, 2026 have until December 2, 2026 to meet the machine-readable marking requirement. Content generated before the rules apply does not have to be labelled retroactively, though the Commission encourages it where feasible. Everything else in the transparency tier applies from day one, regardless of when a system reached the market.
How companies show compliance
The code of practice, published on June 10, 2026 and since assessed as adequate by the Commission and the AI Board, is the shortcut. Signatories can point to its measures to show they meet the marking and labelling duties in every member state at once, instead of defending a bespoke approach to 27 different regulators. Non-signatories keep the same legal obligations and must prove their own methods are equivalently adequate, case by case, and can expect more requests for information about how they mark and label.
The EU has also published a free set of icons deployers can use: a basic AI mark, a “fully AI-generated” variant for content made end to end by a model, and a “partially AI-modified” variant for a real photograph altered into a deepfake. Using them is optional and does not by itself establish compliance, but they come with placement rules — visible at first exposure, and still visible when content is reshared or downloaded. That last requirement is the practical engineering problem, and it is why Google’s own signing of the code leaned on watermarking and provenance work rather than on interface badges alone.
Where the industry is pushing back
Industry groups that supported making deepfakes visible have spent the run-up arguing the scope grew past what the law intended. The Computer and Communications Industry Association says the guidelines the Commission adopted on July 20, 2026 stretched the definition of a deepfake well beyond the 2024 text. “The label was meant to flag deceptive content,” CCIA Europe’s AI policy lead Boniface de Champris told the Guardian, arguing that a landscape in an advert now sits in the same category as a manipulated political speech. He expects the visible change to show up less on social platforms, which already label AI content, than in advertising, film and publishing.
Meta made a version of the same argument while joining the code, saying in its July 28, 2026 announcement that transparency measures should provide clarity rather than more confusion, and warning against content carrying so many overlapping labels that they overwhelm people. Sergey Lagodinsky, the Green MEP who helped negotiate the law, framed it to the Guardian as democracy protection as much as consumer protection, and said industry complaints about burden have tended to fade once implementation arrives.
Platforms have been moving in this direction commercially for a while, from provenance watermarks to the user controls TikTok introduced for AI-generated content in feeds. What changes on August 2, 2026 is that the disclosure becomes a legal duty enforced by national regulators, and it extends to sectors that have never thought of themselves as AI companies — including newsrooms and marketing teams publishing text that readers increasingly cannot distinguish from human writing.
The next concrete step belongs to the signatories. The AI Office will launch two task forces in September 2026, one for providers and one for deployers, to share implementation practice and refine the marking and labelling measures, and the code remains open for new signatures. Member states’ market surveillance authorities take over from there, and the broader machinery of the AI Act moves with them.
