Article 50 starts tomorrow: AI transparency on a double calendar
Tomorrow, 2 August 2026, Article 50 of the AI Act, Regulation (EU) 2024/1689, starts to apply. The omnibus, Regulation (EU) 2026/1744, split the calendar in two: systems placed on the market from 2 August must meet the transparency duties immediately, while systems already on the market before that date have until 2 December 2026. If you ship anything with AI next week, the second date does nothing for you.
Four duties, in plain terms
Article 50 covers four situations. People must be told when they interact with an AI system, which reaches every customer-facing chatbot and voicebot. Synthetic content, meaning AI-generated text, images, audio and video, must carry a machine-readable marking. Deepfakes, realistic AI images, audio or video of people, places or events, must be labelled. And AI-generated text published to inform the public on matters of public interest needs a disclosure.
Run down that list against your own stack: the support chatbot, the marketing content pipeline, the product demo videos, the drafting tools in the comms team. Most companies find themselves on it at least twice.
The double calendar decides your deadline
The dividing line is the date a system is placed on the market. From 2 August 2026 onwards, the duties apply from day one, with no transition. For systems from before 2 August 2026, you have until 2 December 2026 to close the gaps. Launches planned for August and September need the labelling built in before release, because there is no catch-up window. For the existing estate, four months is enough for an orderly fix, and too little for a project that starts in November.
What machine-readable looks like in practice
The law asks that generated content be marked so that software can detect it as artificial. In implementation work this usually means two threads. First, what your model provider already ships: many generators attach provenance metadata or a marking to their output, and your job is to know whether that is switched on, and for which formats. Second, what your own pipeline does with it: exports, compression and CMS uploads can strip metadata, so a marking that existed at generation can be gone by publication. Test the full path from prompt to published file, and write down where the marking survives and where it dies.
Picture Monday morning, 3 August. A product manager asks whether the chatbot released on Friday tells users it is a bot. If nobody can answer inside a minute, use the list below.
What we recommend now
The Monday test: five questions, and two weeks to answer them.
- Where can a person talk to our AI, and does the interface say so at first contact?
- Which tools generate text, images, audio or video, and does the output leave our systems with the machine-readable marking intact?
- Could anything we produce pass as a deepfake, and who labels it before it goes out?
- Does the company publish AI-drafted text to inform the public, and who owns the disclosure?
- For every system placed on the market before 2 August, who runs the remediation plan that lands before 2 December 2026?
Within two weeks: name one owner for Article 50, inventory the chatbots and generators including AI features inside vendor SaaS, test whether markings survive your publishing pipeline, and add a transparency check to the release gate so nothing ships without it. The systems you launch from tomorrow set the standard the older ones must reach by December.
Sources: Regulation (EU) 2026/1744 (Digital Omnibus on AI) · Regulation (EU) 2024/1689 (AI Act)