The Story
As of August 2, 2026, a chunk of the EU AI Act that had been sitting on the calendar for two years finally switched on. It’s called Article 50, and if you’ve heard it described as “the deepfake law” or “the AI labeling rules,” that’s basically right — though it’s a little broader than that.
Here’s the plain-English version of what it now requires.
If you build a chatbot, you have to tell people they’re talking to an AI. Not buried in a terms-of-service page nobody reads. The disclosure has to be clear, at the point of contact, in language a normal person understands. The Commission’s guidelines are blunt about this: hiding “you’re chatting with a bot” in the fine print doesn’t count.
If you make a deepfake — and the law defines that broadly as AI-generated or manipulated image, audio, or video that looks authentic — you have to label it as artificial. What’s interesting is that there’s no requirement that you intended to deceive anyone. Realistic synthetic media triggers the obligation on its own, even if nobody’s being tricked and even if the “person” in the image doesn’t exist. Audio deepfakes get specific treatment in the Commission’s guidelines: the recommended approach is to open with a plain, spoken disclaimer that the content is AI-generated.
And then there’s the piece that matters most technically. Providers of generative AI have to embed machine-readable marks — think watermarks or metadata — so that synthetic content can be detected as synthetic by software, not just by a human squinting at a label. The whole idea is that a detector, a platform, or a fact-checker should be able to run a file through a tool and get an answer: real, or generated.
On enforcement: this falls primarily to national market-surveillance authorities in each member state, with the Commission’s AI Office holding jurisdiction only over a narrow slice — mainly cases tied to general-purpose AI model providers and very large platforms. The penalty for getting this wrong is up to €15 million or 3% of a company’s total worldwide annual turnover, whichever is higher — though for smaller companies the calculation flips, and the lower of the two figures applies. That’s the transparency tier, by the way — the Act’s harshest fines (the €35M / 7% ones) are reserved for outright banned uses, not labeling slip-ups. Still, 3% of global revenue is a real number for a company like Meta or Google.
One more detail worth knowing, because it shows the rules aren’t quite as instant as the headline suggests. Generative systems that were already on the EU market before August 2 get a grace period: they have until December 2, 2026 to meet the machine-readable marking requirement. Anything launching fresh after August 2 has to carry the marks from day one, no runway.
So that’s the shape of it. Chatbots announce themselves, synthetic media wears a label, and the label is supposed to be readable by machines, not just eyeballs.
The Takeaway
Here’s the thing I keep coming back to, and it’s not whether the goal is good. Transparency about synthetic media is obviously a reasonable thing to want. The question is whether the mechanism can actually deliver it — and that’s where I get honest with you.
The soft spot is the machine-readable mark. Watermarks and embedded metadata are wonderfully clean in theory: the file carries a signature, a detector reads it, done. In practice, those marks are fragile. Screenshot a watermarked image and the metadata is gone. Re-encode a video through a couple of platforms and the signal degrades. Crop, compress, re-record off a screen — a lot of the ways synthetic media actually spreads are exactly the ways that quietly strip the mark. The Commission clearly knows this; the guidelines even say the marks should survive cropping and alteration “as far as technically feasible.” That last phrase is doing an enormous amount of work. It’s an admission, written into the rule itself, that the tech isn’t fully there yet.
So you get a gap. On one side, a regulation that says AI content must be detectable. On the other, detection tools that a motivated bad actor can defeat with a screenshot. The people who’ll comply cleanly are the big, legitimate providers who’d probably label their content anyway. The people producing malicious deepfakes — the actual problem the law is aimed at — are also the people most willing to route around a watermark. That’s the uncomfortable shape of it.
Which is not the same as saying the law is pointless. It isn’t, and here’s why. Even an imperfect mark changes the default. When labeling is legally mandatory for every serious provider, “unlabeled” starts to become a weak signal in itself — a reason to look twice. And there’s a second effect that’s easy to miss from inside the US tech bubble: the “Brussels effect.” Because 3% of worldwide turnover is on the line, a global company doesn’t build one system for Europe and a lawless one for everyone else. It’s cheaper to label everywhere. The EU has done this before — GDPR cookie banners and privacy defaults quietly became the planet’s baseline. Article 50 is a decent bet to do the same for AI disclosure. The EU writes the rule; the world inherits the plumbing.
This connects to something this blog has been circling for a while. When we wrote about how researchers are teaching image models to draw people who actually touch each other, the subtext was that synthetic media is getting good enough that the seams are vanishing. And when we looked at why “AI agent identity” suddenly became a hard problem, the theme was the same one hiding under Article 50: as AI systems get more capable and more autonomous, the boring infrastructure question — how do we know what we’re actually looking at, or talking to? — stops being boring. Provenance is becoming the load-bearing problem of this whole era. Article 50 is the first serious attempt by a major government to legislate an answer, and the fact that it leans on watermarks that don’t fully hold tells you how early we still are.
So what do I actually take from it? Two things, and they don’t cancel out. The direction is right — mandatory disclosure is a sane response to a real problem, and forcing every legitimate provider to label their output raises the floor for everyone. But the enforcement mechanism is running ahead of the technology it depends on. Machine-readable marks are only as good as their survivability, and right now that’s the weak link. The honest read isn’t “this fixes deepfakes” and it isn’t “this is toothless theater.” It’s that the EU has planted a flag on a principle before the tools to defend that flag have fully arrived — and whether the watermarking catches up is the thing actually worth watching, more than the €15 million headline.
This article is for informational purposes only.
Photo: ALEXANDRE LALLEMAND / Unsplash
댓글 남기기