YouTube will expand its new “Similarity Detection” technology that identifies AI-generated content such as deepfakes to people in the entertainment industry, the company announced Tuesday.
The technology works similarly to YouTube’s existing Content ID system, which detects copyrighted material in videos uploaded by users, allowing rights holders to request removal or share in the video’s revenue.
Similarity detection does something similar, but on simulated faces. This feature is intended to protect creators and other public figures from having their identities used without their permission. This is a common problem for celebrities who find their likeness being used for fraudulent advertising.
The technology was first made available to a small group of YouTube creators in a pilot program last year, and then expanded to a wider audience this spring, including politicians, government officials, and journalists.
YouTube says the technology is now available to people in the entertainment industry, including talent agencies, management companies, and their own celebrities. The company has received support from major agencies such as CAA, UTA, WME, and Untitled Management, who provided feedback on the new tool.
Celebrities don’t need to have their own YouTube channel when using similarity detection tools.
Instead, the feature scans AI-generated content to detect visual matches to the enrolled participant’s face. Users can request that the video be removed for violating the privacy policy, submit a copyright removal request, or do nothing. YouTube says it won’t remove all content because its rules allow parody and satirical content.
In the future, the technology will also support audio.
Relatedly, YouTube is also asserting similar protections at the federal level, supporting Washington, D.C.’s NO FAKES law. This regulates the use of AI to reproduce an individual’s voice or visual likeness without permission.
The company has not yet disclosed how much AI deepfake takedowns the tool has managed so far, but said in March that the amount of takedowns was still “very low.”
