The short version
Since 2 August 2026, the transparency rules in Article 50 of the EU Artificial Intelligence Act have been in force. Four things now have to be made clear to the people on the receiving end of AI.
- An AI that talks to you has to say so. Chatbots, voice agents and AI assistants must make clear that the other side is a machine, unless that is already obvious.
- AI-made content has to be marked. The company that provides a generative system must embed a machine-readable marker in AI-generated audio, image, video and text so it can be detected downstream.
- Face and emotion systems have to be disclosed. Anyone deploying emotion recognition or biometric categorisation must tell the people exposed to it.
- Deepfakes and AI-written public-interest text have to be labelled. Here the disclosure is for human eyes, not just for software.
Two of these duties sit with the provider that builds the system. Two sit with the deployer that uses it. In practice most brands are deployers, which means the visible labelling duty is theirs.
Why a European law matters in the US and Canada
The AI Act follows the audience, not the head office. It applies to providers and deployers wherever they are established if their AI systems are placed on the EU market or if the outputs of those systems are used in the EU. A North American brand running an EU-facing campaign, a chatbot that serves European visitors, or a product page reachable from Berlin is inside the scope. The turnover figure that sets the penalty ceiling is worldwide, not European.
This is the same pattern the GDPR established. Most companies did not build one privacy standard for Europe and a looser one for everywhere else, because maintaining two standards costs more than meeting the stricter one. Expect the same to happen with AI disclosure.
What it costs to ignore
Breaching the transparency obligations can trigger fines of up to EUR 15 million or 3 percent of worldwide annual turnover, whichever is higher. Lower ceilings apply to small and medium enterprises. Enforcement sits with national market surveillance authorities, with the European AI Office coordinating.
The dates
- 20 July 2026. The European Commission adopted its final guidelines on Article 50. They are non-binding, but they are what authorities will reach for.
- 2 August 2026. The obligations apply, including to systems already on the market.
- 2 December 2026. Generative systems placed on the market before 2 August 2026 have until this date to implement the technical marking and detection measures.
- Content made before 2 August 2026. No retroactive labelling for images, audio and video, where the relevant date is the date of generation. For text on matters of public interest the relevant date is publication, so older material published on or after 2 August 2026 is caught.
What a real label looks like
This is the part that separates compliance from theatre. The guidance requires the information to be clear, distinguishable from the content around it, and given at the latest at the point of first exposure. That rules out most of the habits companies have relied on.
- A line in the page footer, far below the image, does not qualify.
- A clause in the terms and conditions does not qualify.
- A watermark rendered in a tone that vanishes into the picture does not qualify.
- A disclosure that only appears after the click, the scroll, or the end of the video does not qualify.
- A chatbot that admits it is AI only when directly challenged does not qualify.
Alongside the rules, the EU AI Office produced a voluntary Code of Practice on transparency, which the Commission and the AI Board have assessed as an adequate way to demonstrate compliance. It proposes a standard visual mark reading "AI", localised as "KI" or "IA" where appropriate, and a split between content that is fully AI-generated and content that is AI-assisted. Signing is optional, and not signing does not remove the underlying duty, but non-signatories will have to persuade national authorities that their own approach is adequate.
What this means if you are a brand
Treat it as an inventory problem before a legal one. Map every place AI touches a customer: support agents, product imagery, video, voiceover, ad copy, review summaries, synthetic spokespeople, translated content. Decide where your line between fully AI-generated and AI-assisted sits, write it down, and apply it the same way every time.
Then test the pipeline. Machine-readable marking is worthless if your CMS, image processor or ad platform strips it on the way to publication, and many do. Put the human-visible label where the eye actually lands, in the local language. Make chatbots introduce themselves in their first message. Keep a record of what you marked and why.
The commercial case points the same way as the legal one. Getty Images found in a 2024 survey of more than 30,000 people that close to 90 percent want to know when an image is AI-generated. Bazaarvoice's 2025 research put AI-written reviews at the top of shoppers' specific AI concerns. Labelling clearly tells a suspicious audience that everything you have not labelled is real, which is a signal manufactured proof cannot imitate. The brands that will find this uncomfortable are the ones whose imagery has quietly been synthetic all along.
What this means if you are a shopper
You have gained a reference point. Until now, whether a company told you an image or a voice was machine-made was entirely its own choice, which made undisclosed AI close to undetectable for anyone without forensic tools.
A label is not a warning. A labelled AI illustration on an article is perfectly legitimate; the label simply tells you to read the image as a picture of an idea rather than a photograph of the product that will arrive. What deserves your attention is the gap between behaviour and duty: glossy synthetic imagery with nothing marked, a watermark you can barely see, a support agent that dodges the question of whether it is human. And because the rollout is uneven, a missing label is a reason to look for a second source, not a verdict.
The North American backdrop
There is no US federal equivalent, but the same conduct is already actionable on other grounds. The FTC's Operation AI Comply, launched in September 2024, has produced a run of deception cases over AI capability claims, including May 2026 settlements totalling 930,000 dollars with three marketing companies over a service sold as an AI "Active Listening" tool that, per the complaints, resold email lists instead. The FTC's rule on consumer reviews and testimonials, effective October 2024, already bans AI-generated fake reviews. The SEC's Cyber and Emerging Technologies Unit, created in 2025, covers AI claims made to investors. Several states have legislated on synthetic media directly, and the Department of Justice has treated fabricated AI capability as criminal fraud. In the United Kingdom there is no equivalent statute either, but existing regulators apply existing duties to the same behaviour.
For a brand working across these markets, the practical answer is a single disclosure standard set to the strictest requirement, which today is the EU's.
Where this sits in the bigger picture
Labelling laws address the symptom that is easiest to legislate: whether a thing was made by a machine. They do not address whether a claim is true, whether a review came from a real buyer, whether an endorsement was paid for, or whether a badge means anything. Those remain the harder problems, and they are the ones this initiative was set up to examine.
This explainer summarises publicly available law and guidance for general information. It is not legal advice. Sources: Regulation (EU) 2024/1689 Article 50; European Commission guidelines on transparency obligations, adopted 20 July 2026; FTC press materials, September 2024 to May 2026; Getty Images survey, 2024; Bazaarvoice research, 2025. Written by Emir Ramic.