Overview of the Code of Practice
AI Act Adam Radivojevic ● Covered by 2 sources
The EU's Code of Practice for big AI models is now fully mapped out, covering copyright, safety and transparency rules. New models must comply by August 2025, but real enforcement waits until 2026 — so there's a grace period, but the clock's already ticking.
Based on reporting by AI Act, Adam Radivojevic — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Brussels has finally put real meat on the bones of the AI Act's general-purpose AI provisions, and the newly detailed Code of Practice reads less like abstract policy and more like an operations manual for anyone building frontier models. The framework, developed with the EU AI Office, breaks compliance into three chapters — transparency, copyright, and safety and security — each stacked with specific commitments and measures rather than vague principles. Providers aren't legally forced to adopt the Code itself, but doing so is clearly meant to be the easy path; going your own route means proving compliance some other way, with less certainty about how regulators will judge it.
The timeline matters more than people realize. Rules for GPAI models kick in August 2, 2025, meaning anything released after that date has to comply from day one. But the AI Office won't start throwing its enforcement weight around — information requests, model access demands, recalls — until a full year later, August 2, 2026. Older models get even more breathing room, until 2027. That gap is deliberate: it gives labs time to build out documentation and safety infrastructure alongside regulators rather than under immediate threat of penalty.
The substance leans heaviest on the safety and security chapter, and for good reason — this is where systemic-risk models live. Signatories must draw up a
Read more about this at: AI Act