EU's AI labeling rules take effect Sunday — and the public-interest text clause leads the way

On Sunday, EU AI Act Article 50 takes effect, and one of its least-discussed clauses requires AI-written public-interest text without meaningful human editorial oversight to carry a "machine-made" notice. The text clause is what will show up first in news feeds, press releases, and government communications, ahead of the deepfake and chatbot rules that have dominated coverage.

From 2 August 2026, providers and deployers of AI systems that generate text, images, audio, or video for professional or public-interest use must attach machine-readable labels and watermarks, per the European Commission's Article 50 guidelines. The Commission has paired the rule with a Code of Practice on AI-generated content that defines the editorial-oversight test narrowly: a person with judgment and accountability has to materially shape the output, not just approve it.

Press releases spun by an LLM, government bulletins drafted and posted without a named human sign-off, and bulk newsroom output where a model writes the lede and a desk editor clicks publish are all in scope. A column where a reporter uses a model for grammar help is not. The line is whether a human with editorial accountability materially shaped what got published.

The carve-outs are narrower than the public debate implies. Purely personal or non-professional use is out of scope. So is work that is artistic, creative, satirical, or fictional. A novel chapter drafted with an LLM, a satirical column clearly labeled as such, or a personal blog post does not trigger the rule. A press release under a corporate letterhead does. A government bulletin explaining a new policy without a named author does.

Existing systems have a longer runway. AI systems already on the market have until 2 December 2026 to come into compliance, a delay confirmed by Latham & Watkins' May 2026 update on the EU's revised AI Act timetable. That grace period covers deployers' installed systems, not new public-interest text going out from Sunday. The two deadlines are easy to conflate; they are not the same.

Enforcement is built on fines, scaled to the size of the non-compliant provider. The Commission frames the rule as a literacy tool, not a censor: as generative AI makes audience-targeted disinformation cheap, citizens need a way to read what was made by a machine. Labels, not litigation, are the first line of defense.

Industry self-labeling is already ahead of the rulebook. TikTok reports it has tagged more than three billion pieces of content with AI markers, per the AFP-syndicated report carried by Digital Journal. Meta runs an "AI Info" tag on Instagram and Facebook. Google has signed the EU's Code of Practice on AI-generated content and is working with unnamed technology companies on the underlying digital tagging.

The companies most exposed are the long tail. TikTok and Meta already run the labeling infrastructure, even though their stacks predate the rule. The mid-sized PR firm, the regional newsroom that uses a model to draft its weekday roundup, the trade association publishing a position paper under a generic "communications team" byline — these are the deployers that need a documented human-in-the-loop workflow by Sunday, or face the same fine regime as the platform giants.

Named voices are already splitting on the operational load. Google's Karen Massin, quoted in the AFP-syndicated coverage, warns of "regulatory complexity" and overlapping labels that "confuse the people these rules are meant to help." Ashley Casovan of the International Association of Privacy Professionals concedes implementation will be "very, very difficult" but argues the world will adapt. The disagreement is about friction, not direction.

For a reader, the practical filter is short. From Sunday, a public-interest article with a machine-made notice, an "AI Info" tag, or a watermark is doing what the rule asks. An article with no label and no traceable human author on a contested public-interest claim is now the thing to read with more caution. The rule does not make AI text false. It makes the absence of a human visible.