AI-made images, audio and video, and some AI-written public-interest text, must now be clearly labeled and carry invisible, machine-readable watermarks in the EU.
That rule comes from Article 50 of the EU AI Act, which entered into force on 2 August 2026 and has moved into enforcement. It requires companies that provide or deploy AI systems to disclose when content is generated or manipulated and to embed an invisible mark that tools can read.
Why this matters: until now, synthetic content often reached people with no clear sign it wasn’t real. The new duty forces platforms, service providers and creators to make synthetic material visible to users and detectable by machines, shifting responsibility from consumers to producers.
How it works, simply: think of the watermark as a barcode only software can scan. It does not visibly change an image or audio file, but it gives automated systems a reliable flag to spot AI outputs.
What changes now: regulators are already checking documentation and deepfake labeling, and noncompliant firms risk large fines, up to €15 million or 3% of global turnover. Practically, this creates a new compliance burden and an industry need for common watermarking standards and detection tools. It also won’t stop every fake overnight, hidden marks can be removed and detection is an arms race.
The open question is whether watermarks and enforcement will actually make synthetic content safer, or just push bad actors into harder-to-detect tactics. That will decide if the rule becomes a real brake on deception or just another compliance checkbox.