Unique Research · Apps · December 14, 2024
Deepfakes are everywhere. Generative AI’s commoditization has flooded the web with synthetic fakes: identity platform Sumsub says global deepfake volume grew 4× from 2023 to 2024, and in 2024 deepfakes accounted for 7% of all fraud — impersonation, account takeover, and complex social engineering among them.
Meta’s contribution is Meta Video Seal — announced Thursday, open-sourced for integration into existing software — which can embed hard-to-notice watermarks in AI-generated video clips. It complements Meta’s other watermark tools: Watermark Anything (re-released today under a more permissive license) and Audio Seal.
“We built Video Seal for more effective video watermarking — especially detecting AI-generated video and protecting originality,” Meta AI research scientist Pierre Fernandez told TechCrunch.
It is not first: DeepMind’s SynthID and Microsoft also watermark video. Fernandez argues many existing methods fall short — weak under the video compression common on social platforms; inefficient at scale; insufficiently open or reproducible; or adapted from image watermarking, which is suboptimal for video.
Beyond watermarks, Video Seal can hide provenance information for later revelation. Meta claims robustness to blur, crop, and popular codecs. Trade-off: visibility vs. tamper resistance; heavy compression and heavy edits can alter or erase the mark.
Adoption is the bigger hurdle — especially for shops already on proprietary stacks. Meta is launching Meta Omni Seal Bench, a public leaderboard comparing watermark methods, and will run a watermarking workshop at ICLR.
“We hope more AI researchers and developers bake in some form of watermarking — and to move faster with industry and academia,” Fernandez said.
— END —