Any deepfake, any AI-written article, any synthetic voice that could pass for real needs a label in Europe. TikTok already flags billions of posts. Google’s watermark has tagged 100 billion images. Critics say the new rules already reach further than lawmakers intended.
The EU’s AI transparency rules took effect Sunday, 2 August, requiring companies to label AI-generated content that is realistic enough to pass for authentic. The requirement applies to images, video, audio, and text created or manipulated by AI that could be mistaken for human-made content. Therefore, companies, including social media platforms, must mark deepfakes. Otherwise, they may face fines that can reach €15 million. “Generative AI enables disinformation to be created at unprecedented scale, tailored to specific audiences, and disseminated with remarkable speed,” an EU official told Le Monde.
What’s Happening & Why It Matters
The Definition of Deepfake
The rules distinguish between general AI-generated content and deepfakes. A deepfake, under Article 50 of the AI Act, means realistic AI-generated or AI-manipulated images, audio, or video. It imitates a real person’s appearance, voice, or actions in a way that could lead people to believe they said or did something that never happened. Any AI-generated or manipulated content that realistically depicts a person, object, place, or event, and could be mistaken for authentic, must carry a clear label.
The rules extend beyond entertainment content too. If AI creates or manipulates news articles, political content, or public health information intended to inform the public, disclosure is required. However, this does not apply if a human editor has reviewed the material and taken editorial responsibility for the final publication. That carve-out gives newsrooms and publishers a clear compliance path. Of course, this is provided human oversight happens before publication.
Companies Preparing for Months
The compliance groundwork started well before Sunday’s deadline. More than 190 organizations, including Google and Meta, signed a voluntary code of practice supporting the transparency requirements by the end of July. TikTok says it requires creators to label realistic AI-generated media and has already labeled more than 3 billion pieces of content. Google’s SynthID watermarking tool has tagged more than 100 billion images. It has tagged 60,000 years’ worth of audio content, according to the company.
Enforcement won’t rely on a single universal method. Any technique that allows for easy detection of synthetic content is acceptable, whether that’s a visible watermark, embedded metadata, or a clear on-screen label. Henna Virkkunen, the European Commission’s executive vice president for tech policy, described the balance the rules aim to strike. AI is a “transformative technology” capable of “extraordinary benefits,” while the most advanced systems could create risks “on an entirely new scale.”
The Enforcement Gap
Existing platform policies haven’t proven foolproof, even with labeling requirements already in place. For example, TikTok’s own creator-labeling rules haven’t stopped a wave of AI-generated videos featuring fake doctors dispensing dubious health advice from slipping through. As TF covered in Gemini AI Feature Pulled From Google Earth Within Days, Google’s own watermarking system could be bypassed in some cases, according to BBC Verify testing. This undermines confidence in watermarks as a reliable detection method even from the companies building them.
Industry groups are already countering on how the Commission interpreted its own law. The Computer and Communications Industry Association argues guidance issued in July expanded the definition of a deepfake beyond what lawmakers intended when they passed the AI Act back in 2024. “The label was meant to flag deceptive content,” said Boniface de Champris, CCIA Europe’s AI policy lead, suggesting the current guidance catches content that was never designed to deceive anyone.
TF Summary: What’s Next
The transparency obligations apply as of 2 August, with fines of up to €15 million for non-compliance. Additional AI Act provisions take effect in 2027, per the law’s staggered rollout schedule. The Commission’s code of practice is open for additional signatories beyond the current 190-plus organizations. CCIA Europe’s objection to the deepfake definition’s scope is unresolved.
MY FORECAST: Expect the first meaningful enforcement action within six months, against a smaller platform or AI tool provider that hasn’t signed the voluntary code. This will happen rather than against Google or Meta. Both of those companies have already built compliance infrastructure. The bigger unresolved problem is technical, not legal. Watermarks that can be bypassed, the failure BBC Verify documented with Google’s own system, mean the label requirement creates a paper compliance trail. However, it does so without solving the underlying detection problem. Watch for the CCIA’s scope objection to resurface in a formal legal challenge. This is given how de Champris stated the current guidance as exceeding what the 2024 law authorized.
Related Stories
- Gemini AI Feature Pulled From Google Earth Within Days
- Meta’s Muse Image Consent Backlash
- EU Fines Google €890 Million for Favouring Its Own Apps