Starting Sunday, August 2, 2026, companies in the European Union must label AI-generated content that is made to look real (The Guardian). The rules apply to new AI systems entering the EU market from that date. AI systems that already exist get four extra months to comply.

These obligations come from Article 50 of the EU's AI Act, a law the European Commission describes as the first complete set of rules for AI in the world (European Commission). The Act was first agreed in December 2023. To help companies prepare, the Commission published guidelines on July 29, 2026 (European Commission), following a Code of Practice on labelling AI content released June 10, 2026 (European Commission). A second draft of that Code had been published March 5, 2026 (European Commission).

AI-made text, images, video, and audio designed to look truthful must carry a visible mark showing it was made by AI. The content must also contain a digital watermark, which is a hidden code inside the file that computers can read to confirm the content is artificial. Think of the visible mark like a warning label on a cigarette pack — something you can see right away. The watermark is more like a serial number stamped inside the product, invisible to your eye but detectable by machines. Text on topics of public interest must also be labeled as AI-made if no human editor reviewed it. The rules do not apply to content people create for personal use. There is an exemption for clearly artistic, satirical, and fictional works. Companies are encouraged, but not required, to label AI content that existed before these rules.

The Commission has created black-and-white AI labels that anyone can use, though companies may design their own. Companies that fail to comply can be fined up to €15 million or 3% of their total worldwide revenue.

Green MEP Sergey Lagodinsky, who helped negotiate the AI Act, said the transparency rules are about protecting both customers and democracy.

The broader context here involves a disagreement over how far these rules should reach. The Computer and Communications Industry Association (CCIA Europe), an industry group, argues that guidelines published by the EU in July 2026 went beyond what the original 2024 AI Act intended when it came to defining a deep fake. A deep fake is a piece of AI-made content, usually video or audio, designed to look like a real person said or did something they did not. Boniface de Champris, CCIA Europe's AI policy lead, represents the group's position on the matter.

This disagreement touches on the legal foundations of the AI Act itself. The Commission's August 2025 guidelines on banned AI practices, hosted on its AI Act Service Desk domain, laid additional groundwork for how the rules are interpreted and enforced (European Commission). The path from the original 2024 law to the July 2026 guidelines shows the ongoing process of interpretation that companies must navigate.

For affected companies, the difference between new and existing AI systems creates two timelines. Companies launching new AI products on the EU market from August 2, 2026 must comply right away. Those with older systems have until late 2026 to add visible markings and digital watermarks to their AI content. The Commission's ready-made black-and-white label lets companies start immediately or design their own, as long as the labeling and watermarking requirements are met.

In my view, the tension between the Commission's broader definition of deep fakes and the industry's reading of the original 2024 law is an open regulatory question. How enforcement authorities draw the line between authentic-looking AI content and the exemptions for artistic or satirical works will shape how companies across the EU's digital market approach compliance. The financial penalties, reaching up to 3% of global revenue, give regulators significant leverage in any disagreement over whether labeling is adequate or how far the rules extend.

The implementation date comes about two and a half years after the initial political agreement on the AI Act in December 2023. The staggered timeline, giving older systems extra months to comply, recognizes the technical difficulty of adding digital watermarks to AI systems already in use.