Article 50 became enforceable on Sunday. Here is the new baseline.

This analysis first appeared as the weekly Praxikon newsletter.

<h2>The AI Act became enforceable on 2 August</h2> <p>Last Sunday, Article 50 entered into application. In concrete terms: an AI system that interacts directly with people must say so, unless it is obvious. Providers of generative systems must mark output in a machine-readable way. Deployers must label deepfakes, as well as AI-generated text published to inform the public, unless a human has taken editorial responsibility. Anyone using emotion recognition or biometric categorisation must inform the people exposed to it.</p> <p>One transition period remains: only the machine-readable marking (Article 50(2)), and only for systems placed on the market before 2 August, until 2 December 2026. Everything else applies now. Enforcement sits with national market surveillance authorities; fines can reach 15 million euro or 3% of global turnover.</p> <h2>Also sharp since Sunday: GPAI enforcement</h2> <p>From the same date, the Commission, through the AI Office, can genuinely act against providers of general-purpose AI models: requesting documentation, running model evaluations and requiring corrective measures, with the same penalty ceiling. For most organisations the practical side is simpler: you are not a model provider, but you can now hold your supplier to the transparency information they are required to provide. Take that question into your next supplier conversation.</p> <h2>190 organisations signed the Code of Practice</h2> <p>Just before the deadline, on 31 July, the Commission published the list of initial signatories of the Code of Practice on transparency of AI-generated content: over 190 organisations, of which 83 signed Section 1 (providers, including Anthropic, Google, Microsoft, Mistral and OpenAI) and 152 signed Section 2 (deployers, including Getty Images, Lenovo and Lufthansa; some signed both sections). The Commission and the AI Board assessed the Code as adequate. Signatories can therefore rely on it to demonstrate compliance with Article 50(2), (3) and (5). Signing later remains possible; the list is updated on an ongoing basis.</p> <h2>Action for this week</h2> <p>The deadline week is over; this is the evidence week. Record this week what you have actually put in place: which chatbots and AI applications carry a disclosure, who is responsible per application, and where your labelling policy lives. The first information requests from supervisory authorities will go to what is publicly visible, and the difference between "arranged" and "demonstrably arranged" is exactly that file.</p> <p>To see how visible that difference is from the outside, have a look at our <a href="https://www.praxikon.com/en/posts/article-50-field-test-dutch-chatbots">field test of ten Dutch chatbots</a>: what a visitor (and therefore a supervisor) does or does not get to see today.</p> <p>You receive this weekly AI Act update because you signed up with Praxikon or used one of our tools or templates. You can unsubscribe with one click at the bottom.</p> <p>Until next Tuesday.<br/>Zahed Ashkara<br/>Praxikon</p>