Get weekly AI video rankings, evaluation updates, and industry news delivered to your inbox.
The regulation covers text, images, video, and audio, and reaches beyond European borders
As of August 2, 2026, a business in Chicago that generates a realistic-looking video of a public figure speaking and distributes it to European audiences must label that content as AI-generated and embed a machine-readable watermark in the file. The EU AI Act's transparency provisions are now in force, and the geographic scope is broader than many non-European companies anticipated.
The rules apply to deployers of AI systems, including those who only use but do not build them, and they extend to any organization whose AI-generated content targets or reaches EU customers, regardless of where that organization is headquartered. A US company serving European markets falls under the same obligations as a company based in Berlin.
What the rules require
The regulation covers synthetic text, images, video, and audio designed to look real. Content that meets that threshold must carry a visible disclosure and a machine-readable marker, such as a digital watermark, so that detection tools can identify it as artificially produced.
The realistic-looking standard does the regulatory work. Only content designed to appear truthful or lifelike triggers the labeling requirement. That distinction matters for artistic and clearly stylized work, which may fall outside the mandate's scope, though the line between stylized and realistic will require case-by-case interpretation.
AI-generated media depicting real people saying or doing things they never said or did is explicitly covered. The regulation's stated purpose is to combat deceptive synthetic media and protect democratic integrity, language that points toward political content and public figures as the primary concern.
The label fatigue problem
Some industry groups warn that mandatory labeling across a wide range of content will produce the opposite of the intended effect. If every AI-assisted image, from a marketing graphic to a news photo, carries a disclosure, users may begin ignoring labels entirely, making them least effective precisely when they matter most.
That concern has a precedent in cookie consent banners, which became so common under EU privacy rules that research consistently found users clicking through them without reading. Whether AI labels follow the same trajectory depends partly on how platforms implement them and whether the machine-readable watermark layer, rather than the visible label, becomes the primary enforcement mechanism.
Enforcement and the watermark layer
The visible label addresses human viewers. The embedded watermark addresses automated detection systems, and that dual-layer approach anticipates a future where content moderation is partly algorithmic. Platforms and regulators can, in theory, scan content at scale for watermarks without relying on deployers to surface disclosures voluntarily.
What remains untested is how reliably those watermarks survive common transformations such as screenshot crops, format conversions, and social media compression. The regulation mandates that markers be embedded and detectable, but the technical standards for durability are not specified in the source materials available here.
The first enforcement actions under the transparency provisions will test whether the watermark requirement has teeth, and the first cases brought by EU regulators will set the practical boundaries of what detectable actually means.
Get weekly AI video rankings, evaluation updates, and industry news delivered to your inbox.