THE LATENCY GAP
Content moderation runs on a review pipeline measured in seconds to hours. Real-time face-swap now runs at video framerate — under 40 milliseconds per frame. The attacker outpaces the defender by three orders of magnitude.
WHY FACE-SWAP GOT FAST
Until 2023, deepfakes required minutes of GPU time per second of output. Diffusion-based and GAN-distilled models now run in a single forward pass on consumer hardware. The same compression that made image generation instant made impersonation instant.
THE DETECTION ARMS RACE
Deepfake detectors are themselves neural networks trained on examples of fake video. Every public detector becomes a training signal for the next generation of generators — adversaries fine-tune until they pass the classifier. This is why no detector stays accurate for long.
SECTION 230 AND THE LIABILITY SHIELD
Under US law, platforms like Kick are not liable for user-generated content. The 1996 Communications Decency Act §230 was written for text forums; it now shields a livestream platform from a real-time impersonation of a real person to millions of viewers.
WHERE THE LAW BITES BACK
Right-of-publicity statutes (varying by state) and the EU AI Act's deepfake-disclosure rules sit outside §230. Tennessee's ELVIS Act (2024) made unauthorized AI vocal cloning actionable; California's AB 2655 targets election deepfakes. The legal pincer is forming, but unevenly.
THE VC CALCULUS
Sequoia and Benchmark backing Decart at $450M signals that frontier investors expect real-time generative video to be a category, not a feature. The same firms funded the underlying model labs whose outputs the moderation systems cannot keep up with — the capital stack funds both sides of the arms race.