Helsinki: Europe’s AI transparency obligations became binding on 2 August 2026, and every chatbot, voice assistant and synthetic media tool serving users in the European Union must now make clear that a machine, not a person, produced what appears on the screen.
Article 50 of the Artificial Intelligence Act imposes four duties. Providers of systems that talk to people must tell users they are dealing with an AI unless that fact is obvious. Providers of generative tools must mark synthetic audio, image, video and text in a machine-readable format. Deployers of emotion recognition or biometric categorisation must inform the people exposed to them. And deployers publishing deepfakes, or AI-generated text on matters of public interest that no editor reviewed, must label the output plainly.
The last duty carries the sharpest edge for newsrooms, campaign teams and public bodies. The Commission’s guidance says the disclosure has to be visible or audible to an ordinary reader without any detection tool, which rules out a watermark buried in file metadata. It also applies whether or not anyone intended to deceive, and whether or not the depicted person exists.
Timing gives the moment its peculiar shape. Lawmakers spent much of 2026 pulling the Act’s harder deadlines apart. The Digital Omnibus on AI, which Parliament adopted in June and the Council confirmed at the end of that month, moved high-risk obligations for stand-alone Annex III systems to 2 December 2027 and for AI embedded in regulated products to 2 August 2028. Transparency did not move. August 2 arrived on schedule for the one chapter that touches consumers directly.
Companies have had warning. The Commission published draft guidelines on transparency for AI-generated content in May, alongside a code of practice on marking and labelling that firms can sign to demonstrate good faith. The Commission’s own explainer spells out where the obligations attach to a provider and where they land on the business deploying the tool, a distinction that matters enormously for the many European firms that build products on models trained elsewhere.
Enforcement is the softer half of the picture. Fines for breaching Article 50 reach fifteen million euro or three percent of worldwide turnover, whichever bites harder, but the money only moves once national authorities exist to collect it. Several member states missed the August 2025 deadline to designate market surveillance authorities and lay down penalty regimes, leaving a rulebook that applies uniformly and an enforcement map that does not.
Technical standards lag as well. Marking synthetic content in a robust, interoperable, machine-readable way remains an unfinished engineering problem. Watermarks survive one platform and vanish on the next. Provenance credentials work well until a screenshot strips them. European standardisation bodies are still drafting, which leaves compliance officers reading guidance rather than specifications.
The bet Brussels has made is worth stating plainly. Transparency does not judge whether an AI system is good, safe or accurate. It only insists people know what they are looking at. Regulators believe that knowledge changes behaviour, that a labelled political deepfake persuades less than an unlabelled one, and that a chatbot which admits its nature earns more appropriate trust.
Evidence for that belief is thin. Research on food labels and cookie banners suggests disclosure fatigue sets in quickly, and a label seen a hundred times a day may stop registering at all. Europe will run the experiment regardless, and by next summer the first enforcement decisions should show whether the market treated August 2 as a deadline or as a suggestion.




