Please Be Advised
Two of the world’s most influential regulatory bodies picked the same day — August 2, 2026 — to tell you the same thing: you have a right to know when the thing in front of you came out of a machine.
The EU AI Act’s Article 50 transparency requirements kicked in. People must be informed when they’re interacting with an AI system. Synthetic content has to be marked in a machine-readable way. Deepfakes need disclosure. California’s AI Transparency Act, SB 942, landed the same day, narrower: it covers provenance rather than conversation, requiring large generative AI providers to label the images, video, and audio their systems produce and to run a free detection tool anyone can check content against.
The shared date is not a coincidence, and I should say so before someone else does. California’s original deadline was January 1, 2026. AB 853 moved it to August 2 specifically to line up with Brussels. What did converge, and wasn’t coordinated, is the theory of the problem. Two jurisdictions that agree on almost nothing else in tech policy both landed on the same instinct: disclosure is the fix.
I want to believe that. It’s a clean idea. A label. A watermark. A little notice that says “this is AI” before you proceed, the same way a cigarette pack says “this will kill you” before you light up. Informed consent, applied to conversation.
But think about the last time a disclosure actually changed your behavior.
Cookie banners have been mandatory in Europe since 2018. Nearly a decade of “this site uses cookies, click to accept.” How many people read them. How many people even see them anymore, versus reflexively clicking through the way you’d swat a fly. The disclosure exists. The informed part of “informed consent” quietly evaporated somewhere around year two, once the click became muscle memory instead of a decision.
That’s the risk sitting inside these two laws, arriving on the same day, with the same theory of the problem. Disclosure works when the thing being disclosed is legible enough that knowing about it changes what you do next. “This coffee is hot” works because you can feel hot coffee and adjust your grip. “You are talking to an AI” only works the same way if talking to an AI still feels meaningfully different from talking to a person.
And that’s exactly the ground shifting under these laws as they take effect. The better the systems get, the less the disclosure does. A clunky chatbot from 2019 announcing itself was almost redundant — you knew in the first sentence. A frontier model in 2026, given a label that says “AI,” might still fool you into forgetting the label within thirty seconds, because the conversation itself doesn’t give you any texture to hang the warning on.
There’s a version of this that works. It’s the version where disclosure isn’t a one-time click-through but something structurally embedded — a persistent visual marker, a watermark in the content itself, something that doesn’t rely on a moment of attention at the start of the interaction and then disappears from working memory. The EU’s machine-readable marking requirement for synthetic content is closer to that model than a chatbot’s opening disclaimer, and California’s latent provenance requirement is the same bet made narrower. Metadata persists. A sentence you read once and forget does not.
The deepfake disclosure piece is the hardest test case, because deepfakes are specifically engineered to defeat exactly the kind of pattern-recognition that makes disclosure work in the first place. A label on a video only helps if the label travels with the video everywhere it’s shared, stripped, re-uploaded, and re-shared again. That’s an infrastructure problem, not a UX problem, and neither law fully solves it by itself.
So the real test of these laws isn’t whether they pass, or even whether companies comply. It’s whether, five years from now, “AI disclosure” means something closer to nutrition labels — genuinely informative, occasionally acted on — or closer to cookie banners: universal, technically compliant, and functionally invisible.
The problem these laws are trying to solve is that the interface has become indistinguishable from the real thing. It’s worth asking whether a label can fix a problem whose entire premise is that you can no longer tell the difference.
Sources: EU AI Act, Article 50 ↗ · California SB 942 ↗, as amended by AB 853 ↗