Are AI Watermarks Legally Required? The EU AI Act, Explained Plainly
Yes — in the European Union. Article 50 of the EU AI Act requires companies that build generative AI systems to mark AI-generated images, video, audio, and text in a machine-readable, detectable way. The deadline is August 2, 2026, with a grace period until December 2, 2026 for tools already on the market. This is exactly why Gemini, ChatGPT, and other major AI tools have been rolling out watermarking systems over the past two years.
If you've noticed every major AI image generator suddenly adding watermarks — visible stars, invisible signatures, metadata tags — this isn't a coincidence or a branding trend. It's regulatory compliance, and the law behind it has real teeth.
The law: EU AI Act, Article 50
The EU AI Act (Regulation 2024/1689) is the European Union's comprehensive AI regulation. Article 50 specifically covers transparency obligations for generative AI, and it requires two distinct things from providers (the companies that build AI systems like Google, OpenAI, and Midjourney):
- Machine-readable marking — embedding a technical signal (like SynthID or C2PA metadata) directly into AI-generated content so automated detection tools can verify it
- Detectability — ensuring that marking survives common modifications like compression, cropping, and format conversion
Separately, deployers — the people or companies actually publishing AI-generated content, especially deepfakes — have their own disclosure obligation: clearly informing viewers when realistic synthetic content is AI-generated or manipulated.
Why this matters globally, not just in Europe
The AI Act applies to any organization whose AI systems or AI-generated content reach EU audiences — regardless of where the company is headquartered. Since Google, OpenAI, and Midjourney all serve European users, they're all in scope. This is why the watermarking you see on Gemini, the C2PA metadata in DALL-E images, and similar mechanisms aren't optional features — they're global rollouts driven by EU compliance, even if you're using these tools from outside Europe.
Timeline: what happens when
What the technical requirement actually looks like
The draft Code of Practice makes clear that no single watermarking method is considered sufficient on its own. It calls for a layered approach combining:
- Machine-readable provenance metadata — standards like C2PA, embedded directly into the file, recording who created the content and with which AI system
- Pixel-level invisible watermarking — a signal woven into the image data itself (this is exactly what Google's SynthID does) that must survive compression, cropping, and format conversion
- Fallback fingerprinting or logging — where marking techniques fail or get stripped, providers must maintain other ways to trace content back to its AI origin
This explains why Gemini specifically uses two layers — the visible star for human disclosure and invisible SynthID for machine detection — rather than relying on just one.
What happens if a company doesn't comply?
The penalties are substantial: fines of up to €15 million or 3% of total annual worldwide turnover, whichever is higher. For a company the size of Google or OpenAI, 3% of global revenue is a number large enough to take seriously — which is exactly why these companies started building watermarking infrastructure years before the actual enforcement deadline.
Does this law apply to you personally, as a user?
This is the part that gets confused most often. Article 50's watermarking obligation applies to the AI providers — not to individual users who simply generate or download an image. You are not legally required to keep a watermark on an image you personally generated and own.
Where individual obligation does kick in is the separate deployer disclosure rule: if you create a realistic deepfake-style image or video — content that could falsely appear to show a real person, place, or event — and you publish it, you may have a personal obligation to disclose that it's AI-generated, independent of whatever watermark the AI tool already embedded.
Removing a visible logo from your own ordinary AI-generated image (a landscape, an illustration, a product mock-up) is not what this law is targeting. Publishing a realistic fake of a real person without disclosure is. The law cares about deception risk, not about whether your portfolio piece has a small star in the corner.
Does the law require AI companies to make watermarks impossible to remove?
No — and this is a subtlety worth understanding. The Code of Practice requires watermarking to be robust against common, incidental modifications (compression, resizing, format changes) — not necessarily unbreakable against a determined adversarial attack designed specifically to defeat it. This is part of why claims like the SynthID bypass story we covered separately generate so much debate: the law's robustness bar and "theoretically unbreakable" are two different standards.
Quick comparison: who has which obligation
| Actor | What they must do | Example |
|---|---|---|
| AI providers | Embed machine-readable, detectable marking in all generated content | Google, OpenAI, Midjourney |
| Deployers publishing deepfakes | Clearly disclose realistic synthetic content depicting real people/events | A marketing team using an AI avatar of a real spokesperson |
| Ordinary individual users | No specific watermarking obligation for personal content | You, generating a personal AI illustration |
What this means practically, today
If you generate an image with Gemini or another major AI tool, you'll keep seeing visible and invisible watermarks regardless of where you live — this is now baked into the products globally as a compliance choice, not a regional feature. If you own the image and want a clean export without the visible logo for a presentation, portfolio, or personal project, that remains a separate, legal, and simple task.
Need to remove a visible AI watermark from your own image?
Free, auto-detection for Gemini and Midjourney. No account needed.
Try WatermarkOff free →Related reading
Sources: EU AI Act Article 50 (artificialintelligenceact.eu), European Commission Code of Practice on AI-Generated Content, Latham & Watkins AI Act Update (May 2026). This article reflects publicly available regulatory information as of its publication date and is not legal advice.