What Is a Deepfake Video?

Deepfake videos are created by training a generative AI model on images of a target person, then mapping that person's facial geometry onto each frame of a source video. The result is a video that appears to show the target person performing acts they never performed. The process now takes minutes with consumer software. In 2023, Sensity AI reported over 95,000 deepfake videos published online; nearly all depicted non-consensual intimate content. ScanErase's detection system scans for face embeddings across video content — finding deepfake videos that traditional hash-matching systems miss because each AI rendering is technically unique.

Key facts about this term

1. Each deepfake video render is technically unique Unlike copied files, each AI-rendered deepfake is a new unique file. Traditional content moderation systems that rely on file hashes cannot detect them. Biometric face embedding detection is required.
2. Source videos are often taken from legitimate platforms Perpetrators often source the base pornographic footage from adult platforms and then overlay the victim's face. This is why the resulting deepfake may not be detectable by the source platform's hash database.
3. Deepfake videos are explicitly covered by the TAKE IT DOWN Act Congress specifically addressed deepfake videos in the legislative text. Platforms cannot argue that AI-generated content is not 'real' content to avoid their removal obligations.