AI explainers
How accurate are deepfake detectors?
Deepfake detection is the highest-profile category in AI detection. It is also the one with the largest gap between benchmark numbers and field performance. Here is what the figures actually mean.
Three failure modes that drag accuracy down
Compression: TikTok, Instagram, and messaging apps recompress video aggressively. The temporal artifacts deepfake detectors rely on degrade fast under recompression.
Adversarial tuning: deepfake creators specifically test against common public detectors. New deepfake variants often pass detectors that worked on the previous generation.
Short clips: under 3 seconds of video gives the engine little to work with. Most viral clips fall in this range.
What the benchmarks actually measure
Public deepfake-detection benchmarks (FaceForensics++, DFDC, and similar) report numbers on curated datasets. The 95% accuracy you sometimes see is on those datasets, against deepfake methods known at training time. Real platform content is messier.
Practical accuracy ranges
On clean, well-known deepfake methods: 85–95% true-positive rates are reasonable. On adversarially-tuned recent deepfakes: 60–80%. On heavily recompressed reposts: lower still. False-positive rates on real video tend to spike with heavy filters and professional cinematography.
How to compensate
Run more than one signal: face-coherence + lip-sync + audio synthesis markers, separately, give a more robust read than any single number. Combine with provenance (does the source account own the footage?) and context. The combination is far stronger than the score alone.
Try the tool
Deepfake Detector
The deepfake detector reports each signal separately so you can see which ones drove the score.