# By 2025 > Source: > Published: 2026-08-03 11:41:25+00:00 # By 2025 ## What the labeling rule actually says The EU isn't asking for a generic disclaimer on every chatbot response. It's targeting what the legislation calls "deepfakes" and "AI-generated or manipulated content" — think synthetic images, audio, video, and text presented as real. If a piece of content is AI-generated and could reasonably be mistaken for human-made, it needs to be clearly marked. There are two layers to this: **Disclosure to end users**— you must clearly state that content is AI-generated, not just bury it in a terms-of-service page.** Machine-readable labeling**— content should be marked in a way that's technically detectable, either through metadata or other embedded signals. That second part is the one that's going to make engineers sweat. It's not enough to write "(AI)" in the caption. The label has to travel with the file, survive screenshots and re-uploads, and remain readable by automated systems. ## Why this is harder than it sounds From a technical standpoint, the enforcement pipeline is still full of holes. Image provenance standards like C2PA exist, but adoption is patchy. The biggest model providers are already embedding watermarks, but open-weights models running locally don't go through that pipeline at all. If I generate an image on a consumer laptop using a small open model, there's no practical way to force a metadata label into it. The EU seems aware of this. The rules will mostly land on large platforms and high-risk AI systems, not individual uploaders. Platforms like YouTube, X, and TikTok will have to build moderation flows to detect unlabeled synthetic content and flag it. That pushes the burden to companies with engineering teams, which is probably the right call. But it also means the enforcement will look wildly inconsistent across different types of content. ## What's good about it I actually think the labeling requirement is one of the more practical pieces of the AI Act. It doesn't try to ban synthetic media — it just makes the synthetic part visible. That preserves free expression while giving viewers context. For someone following a viral video claiming to show a politician saying something inflammatory, knowing it's AI-generated changes the entire analysis. The machine-readable part is also smart, even if it's underdeveloped. If metadata standards become mandatory in the EU, we'll see those standards creep into global tools — the same Brussels effect that gave us GDPR-style privacy buttons everywhere. That could lead to a de facto global watermarking standard for AI content. ## The catch There's a real tension between labels and creativity. A harmless AI-meme with a funny cat face suddenly needing a formal transparency notice feels like overkill. And for text content, the line between "AI-assisted" and "AI-generated" is blurry. If I use an LLM to fix spelling but write every sentence myself, does that need a label? The EU's guidance suggests no, but the boundary is fuzzy in practice. Enforcement will also be messy. Auditors can't actually verify whether a piece of text was written by a human or an AI. You can't watermark prose reliably the way you can with images — it's trivially easy to strip or rephrase. So the text-labeling part may remain aspirational for a while. Still, the EU is the first major regulator to push these requirements into law. The rest of the world will be watching how the enforcement mechanics shake out, and I suspect the technical community will end up building the detection tools that make it feasible. Whether that's through metadata, watermarking, or behavioral analysis, the next couple of years will be interesting. [Next Paying artists for AI training: enough to change minds? →](/en/news/4840/)