From August 2, in the EU, if an AI talks to you, it has to say "I am AI" first. Not politeness. Law.
You see a video that looks like a politician said something outrageous. Before, you had to guess if it was real. From August 2, the EU rule changes: any AI-generated deepfake must carry a label and a machine-readable marker so platforms and content systems can auto-detect it. Chatbots and AI assistants must tell the user up front, "you are talking to AI."
What exactly expanded
The European Commission announced on July 31 that, from August 2, it expands the scope of the AI Act. Previously the law targeted "high-risk" uses (hiring, credit, policing). This round pulls the everyday interaction layer of general AI into scope:
- Chatbots and AI assistants must clearly state at the start of interaction that the user is talking to AI.
- Deepfake content must be labeled and carry a machine-readable marker for automatic detection.
- General AI model providers: the European AI Office can directly enforce and hold accountable on these rules.
In plain terms: regulation moved from "gate the use" to "gate the identity." No matter where the AI is used, if it speaks or generates convincing fakes, it must show its badge. That is a "mandatory self-introduction" threshold for every AI product facing EU users.
What it means for AI builders
If you ship to the EU market, three things to add:
First, a disclosure line at the conversation entry - "You are interacting with AI." Not buried in collapsed terms. Visible at the start.
Second, image, audio, and video generation pipelines need visible watermarks plus machine-readable metadata (C2PA-style standards are becoming the de facto norm). Unlabeled content can be taken down or fined in the EU.
Third, general-model providers serving the EU now face the European AI Office as a direct enforcer. Compliance is no longer "write a white paper" - it is a real, accountable obligation.
Cold water: labeling is not a silver bullet
Real limits:
First, machine-readable markers can be stripped on purpose. Watermarks and metadata can be erased; the "dark web" version of a deepfake still circulates. The law governs compliant platforms, not bad actors who deliberately evade.
Second, "told it is AI" does not lower all risk. Users knowing the other side is AI does not stop them from being swayed by AI-generated false arguments - labeling fixes awareness, not discernment.
Third, compliance cost gets passed on. Small teams adding C2PA, reworking flows, and running audits face another dev and legal bill. Big labs have compliance teams; small teams may be priced out of the EU market.
Fourth, cross-border friction is inevitable. One EU standard, another for other regions - global teams maintain multiple labeling logics, raising complexity.
What this means for you
If you build AI products (especially overseas-facing): the EU wave is a compliance bellwether. Even if your main market is outside Europe, C2PA watermarks and AI identity disclosures are becoming the "international default." Bake the labeling logic into architecture now; retrofitting per market later costs more.
If you make content or run a media account: you will see more "AI-generated" labels. That is good - at least the right to know returns. But do not relax just because "it tagged AI." The label shows origin, not truth.
If you are just a user: after August 2, EU-zone AI chats get more "honest" about identity, and deepfake videos more likely flagged. You get one more layer of protection. But whether you hit pause is still your own hand.
The EU approach replaces "pre-review" with "forced transparency" - it does not block what you build, but requires you to make clear what it is. The direction is debated, but at least it moved "should AI identify itself" from industry self-discipline to legal baseline.