Published numbers, all real, none directly comparable:

  • Resemble AI benchmarks (August 2026, public leaderboards): 99.5% pooled audio accuracy, 95.8% on images, 98.2% on video.
  • Pindrop: 99% detection rate on known synthetic voice engines.
  • Resemble Detect-2B: about 94% on unknown voice generators.
  • Reality Defender: enterprise claims validated against customer traffic, not public sets.
Deepfake Detection Accuracy Comparison: Why Vendor Claims Don't Line Up
Image via Wikimedia Commons

Why they do not line up:

- Test sets differ. Vendor-curated sets flatter; public benchmarks like FaceForensics++ punish. - Known vs unknown generators. Every detector aces generators it trained on. Unknown ones are the real test. - Compression. Lab accuracy drops hard on recompressed, cropped social media video. - False positives are hidden. A 99% detector that flags 4% of real content (Resemble reports 4% image false positives) creates its own crisis. Sources: CISA Demand the false positive rate and the test conditions, not just the accuracy headline.