A new rule from the European Union became enforceable on August 2, and it changes how any business showing AI-made content to European customers has to operate. The rule, part of the EU's AI Act, says that AI-generated deepfakes, meaning fake images, video, or audio that look like a real person, place, or event, must carry a clear label. AI-written text on public interest topics that no human has reviewed needs one too.
This is not a small compliance footnote. Companies that ignore it can be fined up to fifteen million euros, or three percent of their entire worldwide revenue, whichever number is bigger. And the rule reaches far beyond Europe's borders: any company anywhere in the world falls under it the moment its AI content is visible to someone sitting in the EU.
The label itself has real rules attached. It has to be something a person notices right away, not a line buried in the terms and conditions, and not just invisible file data that only a machine can read. For video and images, that means a visible marker, and for audio, a spoken disclaimer at the start.
Companies must also embed invisible, machine-readable tags into the content itself. Anthropic has already confirmed its Claude models will carry these hidden markers going forward, with older models being updated to match.
Here is where it gets complicated for anyone hoping this simply solves the deepfake problem. Studies on how well people spot fake content on their own are not encouraging. One well-known study found people correctly told real from fake media apart only about half the time, which is essentially a coin flip.
Other research has found the number can be even lower, and in some cases people rated AI-made faces as more trustworthy than real ones. Labels are supposed to fix that gap, but there is a real risk they backfire.
If people get used to seeing an AI label on content, they may start treating anything without a label as automatically real. That is a problem because the people making fake content to deceive others, by definition, are not going to voluntarily label it.
There is also a technical hole in the plan. The invisible tags meant to mark AI content can be stripped out just by taking a screenshot, converting a file format, or re-uploading to a social platform that scrubs metadata. That happens to be exactly how most fake content spreads online in the first place.
For any business using AI to generate marketing images, product photos, ad voiceovers, or written content aimed at EU customers, the practical takeaway is simple: label it clearly and visibly, and don't rely on invisible tags alone. Don't assume a compliance checkbox solves the trust problem with your own customers, either.
The rule is now live, and Europe is watching for violations from day one.