Deepfake, the EU Closes the Game: Declare AI or Disappear

The date that will make deepfake creators tremble

Imagine posting a video on TikTok and finding out that if you don't declare it was generated by AI, you risk hefty fines or content removal. From June 10, 2026, this is no longer science fiction: the European AI Office has published the Code of Good Practice on Transparency, implementing Article 50 of the AI Act. Obligations for labeling, multi-level marking, and steganographic watermarking for providers. In practice, if you use AI to create content, you must lay your cards on the table. Or disappear.

What changes for you (and your feed)

If you are a company producing deepfake videos for marketing, or a creator using Midjourney for covers, get ready: every AI-generated content must have a visible mark (e.g., an 'AI' icon) and an invisible one (steganographic watermark) that withstands screenshots and compression. It's like putting an indelible stamp on every digital artwork. And no, you can't remove it: the law provides for penalties up to 7% of global turnover.

Steganographic watermarking: the hidden signature

The code introduces a two-level system: clear labeling for the user (e.g., 'This video was generated by AI') and invisible technical marking, detectable only by authorized software. It's like writing your name in invisible ink: no one sees it, but authorities do. For AI model providers, the obligation takes effect immediately; for distributors (e.g., social media), there is a 6-month adjustment window.

Practical impact: from theory to bureaucratic boredom

Reading the new guidelines is as fun as cleaning tile grout with a toothbrush. But it's necessary. Companies will need to update their content generation systems, train employees, and implement detection tools. Citizens, on the other hand, will gain transparency: no more fake videos of politicians saying things they never said. At least, in Europe.

Exceptions: when AI does not need to be declared

Not all AI-generated content must be labeled. If the content is clearly artistic, satirical, or parodic, or if the AI intervention is minimal (e.g., automatic filter correction), marking can be avoided. But beware: the line is thin. Better to over-declare than risk a fine.

What to do now: a checklist to avoid being caught off guard

• Check if your AI tools (e.g., ChatGPT, DALL-E, Synthesia) already support the required watermarking.
• Update company policies to include the labeling obligation.
• Train the team on how to apply visible and invisible markings.
• Monitor the European AI Office guidelines: detailed technical specifications will arrive by the end of the year.

Don't wait for a violation notice. June 10, 2026 is tomorrow, in regulatory terms. And if you think it's a hassle, think about how much worse it is to see a deepfake of your boss announcing mass layoffs. Transparency is boring, but necessary.