Every chatbot, deepfake and AI-altered photo shown to European consumers must carry a clear disclosure starting Sunday, when Article 50 of the European Union’s Artificial Intelligence Act takes effect. The rule targets machine-made disinformation. Companies that skip the label face fines up to €15 million (roughly $16 million) or 3% of global turnover, whichever is higher.
Brussels wrote this duty into the Act back in 2024 and gave companies a two-year runway to prepare for it. The European Commission finalized its compliance guidelines on July 20, just thirteen days before the deadline. A separate reform pushed the law’s tougher high-risk AI rules back to December 2027, so the labeling duty now stands as the Act’s first full-scale enforcement test.
Four Categories of Content Now Need an AI Tag
European officials have watched generative tools get cheaper and more convincing for years, and they are direct about why the label rule exists.
Generative AI enables disinformation to be created at unprecedented scale, tailored to specific audiences, and disseminated with remarkable speed.
An EU official made that case in a statement announcing the rules, adding that the technology is making it harder for people to tell what is real from what is synthetic. The goal, the official said, is to protect the public’s ability to trust what they see, hear and read.
The European Commission’s own breakdown of four separate transparency obligations splits that broad goal into specific duties by content type.
| Disclosure Category | Who Must Act | What The Rule Requires |
|---|---|---|
| Direct interaction (chatbots) | Providers of conversational AI systems | Tell users up front they are talking to a machine, unless that is already obvious |
| AI-generated content | Providers of image, audio, video or text generators | Mark outputs in a machine-readable format that signals artificial origin |
| Emotion recognition and biometric categorization | Deployers running such systems | Inform any person exposed to the system that it is in use |
| Deepfakes (image, audio or video) | Deployers publishing synthetic media | Disclose that the content is artificially generated or manipulated |
| AI-generated text on public-interest matters | Deployers publishing such text | Disclose that the text was artificially generated or manipulated |
Each duty attaches to whoever puts the AI system to use, whether that is the company that built the chatbot or the business that deployed it to talk with customers.
A Two-Year Countdown Ends With Guidance Two Weeks Old
The timeline stretches back two years. Brussels wrote Article 50 into the Act when it became law and staggered enforcement across a multi-year rollout instead of activating every provision at once.
- August 2024: The EU’s Artificial Intelligence Act enters into force.
- August 2, 2025: Article 99’s penalty tiers, up to €35 million or 7% of global turnover, become enforceable.
- June 29, 2026: The Council of the EU gives final approval to a Digital Omnibus package that delays high-risk system duties under Annex III.
- July 20, 2026: The European Commission adopts final guidelines on Article 50 transparency obligations, less than two weeks before they apply.
- August 2, 2026: Article 50’s labeling duties for chatbots, deepfakes and synthetic content take effect.
- December 2, 2026: A separate new ban on AI tools that generate non-consensual intimate imagery becomes enforceable.
- December 2, 2027: The delayed Annex III high-risk system obligations are now due to apply.
The Commission used nearly the entire window before publishing final instructions. Its finalized guidance adopted July 20 gave companies just thirteen days to align their systems with the government’s reading of the rule before fines became possible.
Brussels Delayed Its Harder AI Rules to 2027
A month before the guidelines landed, the Council of the European Union approved a Digital Omnibus package that simplifies parts of the AI Act. A Gibson Dunn analysis of the deal found it pushed the Annex III deadline to December 2027, a 16 months extension for high-risk system obligations such as registration and conformity checks.
Article 50 sat outside that extension. Its labeling duties kept their original August 2 date, so transparency enforcement now begins well before the Act’s risk-management framework for high-risk systems, which will not bind companies until December 2027.
That sequencing makes Article 50, once just one piece of a much larger rollout, the AI Act’s first full-scale enforcement test.
Who Has to Comply, and Who Is Exempt?
The duty falls on companies and professionals who deploy AI tools commercially inside the EU, not on individuals experimenting at home. Businesses that build, deploy or publish AI-generated content for customers, audiences or professional purposes must disclose it. Personal, non-commercial use of the same tools stays outside the law’s reach.
- Companies and professionals – anyone building, deploying or publishing AI-generated content for commercial or professional purposes inside the EU must disclose it under Article 50.
- Private individuals – people using AI tools in a purely personal, non-professional capacity are not covered, according to the EU’s own description of the rule.
- Small and mid-sized businesses – for specified administrative infractions, their penalty calculation flips to whichever figure between the flat sum and the turnover percentage is lower, rather than the higher figure large firms face.
An EU official framed the practical rollout as starting with chatbots, making it plain to users when they are talking to a machine, before enforcement attention shifts toward deepfakes, which lawmakers have called a greater concern.
The Price of Getting It Wrong
That mid-tier fine mentioned above is the one Article 50 falls under. A breakdown of the three separate penalty tiers under Article 99 shows the top bracket, reserved for outright bans on practices like social scoring, running up to €35 million or 7% of turnover.
The bottom tier, for supplying regulators with incorrect information, caps at €7.5 million or 1% of turnover. Regulators size each fine against both the flat euro figure and the percentage of revenue, then apply whichever number is larger for big companies.
Compliance Readiness Lags the Calendar
A gap remains despite the two-year runway. A Cloud Security Alliance research note tracking the Digital Omnibus rollout found most organizations still unprepared, with 78% not having taken meaningful compliance steps as of April, months after the Commission signaled the transparency deadline would hold.
Part of the readiness problem is that the format itself only firmed up recently. Commission guidance points companies toward a standardized visual marker, a plain “AI” tag localized as “KI” in German or “IA” in French and Italian, meant to accompany labeled material. A voluntary Code of Practice on Transparency of AI-Generated Content, published alongside the July 20 guidelines, is meant to help providers apply the marking consistently across platforms and languages.
Regulators in 27 member states now carry the job of checking whether that marker actually shows up where it should.
Frequently Asked Questions
Does the Labeling Rule Apply to AI Content Made Before August 2026?
No. Article 50(4) of the AI Act does not impose a retroactive labeling duty, so images, audio, video or other synthetic content created or altered before August 2, 2026 do not need to be marked after the fact. Only content generated or manipulated from that date forward falls under the disclosure requirement.
What Legally Counts as a Deepfake Under the AI Act?
Article 3(60) of the Act defines a deepfake as AI-generated or AI-manipulated image, audio or video content that resembles real people, objects, places, entities or events and would falsely appear authentic or truthful to a viewer. The definition applies even when nobody intended to deceive anyone and even when the content does not depict an actual identifiable person.
Are Ordinary People Who Use AI Tools at Home Affected?
No. The EU has said the obligations apply to content produced for professional or business purposes, and that individuals using AI tools in a purely personal, non-professional capacity fall outside the rule. Someone generating an image for their own use does not need to label it; a business publishing the same image to customers does.
Does the Rule Cover AI-Written Text as Well as Images?
Yes. Beyond chatbots and deepfakes, Article 50 also requires disclosure when AI-generated text is published to inform the public on matters of public interest, covering AI-assisted news articles, analysis and similar written content distributed to European audiences. That duty sits alongside, and is separate from, the image and video marking rules.
