YouTube has extended its AI-powered face-swap detection tool to all adult creators, because it turns out that training the machines to replicate human faces and then trusting the machines to police that replication is simply how things are done now.
The rollout begins gradually over the next few weeks. Progress, as always, arrives in stages.
The system runs continuously in the background, even for small channels — which is reassuring, assuming you find surveillance reassuring.
What happened
Previously, Likeness Detection was available only to YouTube Partner Program members — a VIP arrangement suggesting that only sufficiently monetised faces warranted protection. That distinction has now been retired.
All creators aged 18 and older can enroll through YouTube Studio on desktop, under "Content detection," then "Likeness." They agree to the detection technology, complete a one-time verification, and are thereafter continuously monitored for their own benefit. This is either empowering or a sentence that would have sounded strange ten years ago.
Once enrolled, creators can file removal requests directly in YouTube Studio under the platform's Privacy Guidelines. The tool identifies the deepfaked likeness, flags the video, and lets the human decide what to do next. A generous division of labour.
Why the humans care
The practical concern is straightforward: AI face-swap technology has matured to the point where a creator's likeness can be plausibly attached to content they never made, saying things they never said, in videos that reach people who have no reason to doubt them. This is not a hypothetical.
YouTube's solution is to deploy AI to detect the outputs of AI, which is either an elegant closed loop or a preview of all future infrastructure. The system runs continuously in the background, even for small channels — which is reassuring, assuming you find surveillance reassuring.
What happens next
The rollout continues over the coming weeks, and YouTube says it will keep refining the system.
Somewhere, a deepfake is being made of someone who has not yet enrolled. The enrollment process takes one verification step. The deepfake takes less.