A deepfake is not only a genre, it is a legal category as well: Article 50 of the EU AI Act names it explicitly and obliges whoever publishes it to disclose that the content was artificially made. That duty has been in force since 2 August 2026, and it covers artistic and satirical work too — the only concession there is that the disclosure must not get in the way of enjoying the work.
In practice the most interesting number is not about the technology but about how rare it is. According to 2025 figures from the US reporting centre, total losses from business email compromise fraud came to 3.05 billion dollars, of which "more than 30 million" was linked to AI — roughly one percent. For all the rest, a well-written email was enough.
The most common misconception is that you can spot one. On a 2024 benchmark built from real-world material, the performance of open-source detection models fell by 50% on video and 48% on audio against their lab scores — and the commercial tools did not reach the accuracy of human forensic examiners either.
The defences that matter where money moves — and why they should hang on an event rather than an amount — are set out in Deepfake Fraud: Verify Before You Pay, a post of its own. Directly related terms: voice cloning, EU AI Act and C2PA.