Facial Movement Analysis Detects AI Deepfakes at 95.3% Accuracy in Lab Conditions

A facial movement analysis tool detected deepfakes with 95.3% accuracy in controlled laboratory settings. The remaining 4.7% of deepfakes were not detected. These undetected deepfakes passed through the system. The tool was validated before deployment. Deployment has begun.

Validation in laboratories does not predict performance in the world. Lighting varies. Compression happens. Attackers apply resources to the problem. The 4.7% that failed in ideal conditions will increase in actual conditions. The gap between validation and field performance is known. The tool is deployed anyway. This is the sequence.

The deepfakes that will evade detection are already being created. They are being created because the tool exists and because the tool's limitations are discoverable. The 95.3% accuracy is a previous number. It describes a moment when the tool was better than attackers were prepared. That moment is now complete.