The EU's Big AI-Transparency Deadline Hits August 2. The Part That Actually Matters Quietly Slipped to December.
On August 2 the EU AI Act's Article 50 transparency rules take effect — and what actually arrives is a chatbot telling you it's a chatbot. The version in the headlines, sweeping machine-readable labeling of AI-generated content, is not what goes live: that provision (50(2)) was quietly deferred to December 2 for every genAI system already on the market. Live Aug 2 is cheap disclosure (chatbot ID, deepfake disclosure, emotion-recognition notice); the hard provenance piece — the one that would let anyone tell real from synthetic — slipped four months, and even then nobody's solved how to label a half-AI-edited video.
On August 2, the European Union's first broad AI-transparency law takes effect, and you'll know because a chatbot will tell you it's a chatbot. That's most of what actually arrives. The version in the headlines — sweeping labeling of AI-generated content across the internet — is not what goes live that day. The single hardest and most consequential obligation, the one requiring AI systems to stamp their output with machine-readable "this was made by AI," quietly got pushed to December for every system already on the market. The EU is shipping the cheap, visible part of its transparency regime on schedule and deferring the part that would actually let you tell real from synthetic.
This is regulatory theater in the precise, non-conspiratorial sense: the easy obligation lands on the announced date, the hard one slips, and the coverage conflates them into "AI transparency arrives August 2." The law itself is real and enforceable — Article 50 carries fines up to €15 million or 3% of global turnover — so this isn't a paper tiger. It's just that the thing arriving on the marquee date is a disclosure banner, not a provenance system, and those are very different levels of ambition.
Here's the split, precisely, because almost every writeup gets it wrong. Live on August 2: chatbots and conversational systems must tell users they're AI (Article 50(1)); deployers must disclose deep fakes (50(4)); emotion-recognition and biometric-categorization systems must notify the people subject to them (50(3)). Deferred to December 2, for any generative system already on the market before August: the provider obligation to mark synthetic output as machine-readable AI-generated content (50(2)). That deferral came out of the Council and Parliament's provisional agreement in May, and it's the reason the widely-repeated line — that content labeling goes live August 2 "regardless of the Omnibus delay" — is simply not accurate. Disclosure wasn't delayed. Incumbent content-marking was, by four months.
Why does the split matter? Because the two halves are not close in either difficulty or value. Telling a user that ChatGPT is AI is trivial to implement and changes almost nothing — you already knew. Machine-readable provenance is the part that would actually matter: it's what lets a platform, a newsroom, or a court distinguish a real video from a synthetic one at scale, automatically, without a human squinting at artifacts. That's the obligation that got deferred, and it got deferred for the obvious reason — it's technically hard and nobody has fully solved it. The EU shipped the part that's easy because it's easy, and postponed the part that's hard because it's hard, which is a understandable way to run a rollout and a dishonest way to headline one.
And even December doesn't resolve the real problem, which is definitional. The EU published a voluntary Code of Practice in June with a set of uniform icons and labels for AI-generated content. But the hard question isn't the icon — it's the threshold. How does a platform mark a video that was color-corrected by AI versus one where the subject never existed? "AI-assisted edit" and "AI-generated" are ends of a spectrum, and virtually all real media is going to land somewhere in the murky middle as AI tools get baked into every editor. The icons don't tell you where on that spectrum the label becomes mandatory, and that ambiguity is exactly where the first genuinely messy enforcement fight is going to happen.
So watch two things. The first Article 50 enforcement action will almost certainly hit the easy obligations — a chatbot that didn't disclose, or an undisclosed deep fake — because those are cheap to prove. The one that matters is whether the December 2 marking deadline for incumbents actually holds, or slips again, because the EU has now demonstrated it will defer the hard part when the hard part is inconvenient. My read: August 2 is a real milestone and a smaller one than billed — the EU gets an enforceable disclosure regime now and kicks the provenance problem, the thing that actually addresses a synthetic-media internet, down the road, where a body that just deferred it once may well defer it again. What would change my mind is concrete: the December marking deadline holding, plus an enforcement action with actual teeth behind it. Until then, the world's first sweeping AI-transparency law arrives as a chatbot introducing itself — the easiest promise it could have made, and very close to the least anyone needed.