Europe’s Filter Bubble Gets a Label
Brussels has a fondness for deadlines that arrive quietly and detonate loudly. August 2nd is one of them. On that date, Article 50 of the EU’s Artificial Intelligence Act begins to bite, and with it comes a rule that will touch rather more of daily life than the marquee provisions on “high-risk” systems that have dominated headlines since 2024.
The rule is simple to state and maddening to apply: content generated or meaningfully altered by artificial intelligence must say so, visibly, inside the content itself. Not in a caption. Not in a footer. Not buried in metadata—that particular obligation, aimed at the big model-makers, has been pushed back to December 2026. From this week, a synthetic face on a shopfront advert, an AI-polished hotel photo on a travel blog, or an AI-narrated video must carry a mark that a viewer can see or hear without special tools.
Non-compliance is not cheap. Regulators can levy fines of up to €15m or 3% of global annual turnover, whichever is larger, and there is no warning shot built into the law. In Germany, the Bundesnetzagentur—the same body that already polices telecoms—has been handed the job of market surveillance, meaning any citizen who spots an unlabelled synthetic image can simply report it.
Who counts, and who doesn’t
The Act splits the world into “providers”—the OpenAIs and Midjourneys of the world, who build and sell the systems—and “deployers,” everyone else who puts those systems to commercial use. A marketing agency generating product shots, an online retailer using synthetic models, a YouTuber building AI thumbnails: all deployers, all on the hook. Purely private, non-monetised use is exempt, but the threshold for “commercial” is generously low. A creator earning modest sums from Instagram already qualifies.
Corporate structure offers no shelter, either. Responsibility sits with the legal entity that publishes the content, regardless of whether the actual generation was outsourced to a freelancer or an agency.
The deepfake test nobody expected
The Act’s definition of a “deepfake” is broader than most executives assume. Brussels has clarified—in guidance published on July 20th—that the content need not depict a real, identifiable person, place or object. It merely has to resemble something that could plausibly exist in the real world convincingly enough to mislead. A photorealistic, entirely invented human face used as a stock model therefore requires labelling, even though no real individual was copied. Intent to deceive is irrelevant; only the capacity to mislead matters.
The exemption is narrow: obviously fantastical imagery—a mythical creature soaring over a European skyline, say—needs no disclosure, because no reasonable viewer would mistake it for reality. Everything in between, including routine retouching, sits in a grey zone that lawyers are still mapping. Removing a stray rubbish bin from a background with AI tools is generally treated as ordinary editing; conjuring an entire fictitious location is not. Where exactly the line falls between the two is, for now, a matter of professional judgment rather than settled law.
Text gets an easier ride
Written content fares better than images, at least on paper. The labelling duty for AI-generated text applies only to material published on matters of “public interest”—politics, health, consumer affairs—and even then, a human editorial review that takes responsibility for the final copy removes the obligation entirely. A spellchecker does not count as review; running a second AI model over the text does not count either. Advertising copy, product descriptions and internal communications fall outside the rule altogether.
Where the label has to sit
The regulation itself only requires that a disclosure be perceptible “without technical aids” at first contact with the content—a deliberately vague standard. A voluntary code of practice, going further than the law strictly demands, suggests concrete symbols: “AI-Generated” for wholly synthetic material, “AI-Modified” for edited content, placed directly in the visible frame of an image or video, repeated after interruptions in longer videos, and rendered audibly in sound content. Because the code is voluntary, businesses may design their own labelling—at their own legal risk if a court later finds it insufficient.
Predictably, industries are pushing back. Broadcasters worry that an audible disclaimer before every AI-assisted clip will wreck programme flow. Photographers ask, reasonably, where routine retouching ends and “a new image” begins. Regulators have offered guidance, not certainty, and litigation will likely do the rest of the definitional work.
The bigger picture
Germany’s notoriously active system of competitor lawsuits (Abmahnungen) means the transparency rule is likely to be tested in court quickly, and often, well before the December metadata deadline adds a second layer of obligation for the model-makers themselves. Labelling, it is worth stressing, is not a shield: content that infringes personality rights or copyright remains illegal whether or not it carries an “AI-Generated” tag.
For a law aimed chiefly at the platforms and publishers of political and consumer discourse, its practical reach extends to a rather more mundane cast of characters—online retailers, travel bloggers, and, as one German lawyer put it, “every 15-year-old making a bit of money on Instagram.” Whether that cast notices in time is the open question Brussels has, for now, left to enforcement.
About Techvox: Techvox is the English-language technology and AI desk of Swissvox, an independent Swiss media platform funded solely by its readers — free of party affiliation, advertisers, or state support. Techvox covers technology and AI developments with a particular focus on Switzerland and Asia. Readers who want to support independent tech journalism can do so through Techvox’s membership model.