Independent resource on AI content authenticity — detection, standards & policy
Community Forum →
Glossary › Deepfake Detection

Deepfake Detection

Deepfake detection identifies AI-generated or AI-manipulated video, audio, and images using spectral analysis, liveness signals, and temporal consistency checks. Detection methods and their limitations explained.

Also searched as: deepfake detection, how deepfake detection works, how to detect deepfakes

What Is Deepfake Detection?

Deepfake detection is the process of identifying AI-generated or AI-manipulated media — video, audio, and images — in which a person's likeness or voice has been synthesized or substituted. The term "deepfake" comes from the combination of "deep learning" and "fake."

How Deepfake Detection Works

For Video

  • Temporal consistency: Real video maintains consistent head movement, micro-expressions, and physiological signals (subtle skin color changes from blood flow). AI-generated faces often show temporal inconsistencies — unnatural blinking, abnormal head pose transitions
  • Frequency-domain artifacts: GAN-generated faces leave characteristic spectral artifacts detectable via DCT (discrete cosine transform) or Fourier analysis
  • Semantic analysis: Inconsistent shadows, reflections in eyes, ear geometry mismatches, and hairline artifacts

For Audio

Voice deepfake detection uses spectral analysis to identify patterns characteristic of text-to-speech (TTS) systems:

  • Breath noise: Real speech includes natural breath sounds between utterances. TTS systems often miss or unnaturally simulate these
  • Spectral flatness: AI-generated speech often shows unnatural flatness in frequency bands associated with vocal tract resonance
  • Liveness signals: Room acoustics, microphone noise floor, and environmental sounds are inconsistent or absent in TTS output

Current Performance: Voice Detection Benchmark

From our January 2026 benchmark across 600 audio clips:

  • Hive Moderation: 88% accuracy, 9% FPR (best for voice cloning detection)
  • ElevenLabs Detect: 83% accuracy, 13% FPR
  • Resemble Detect: 79% accuracy, 16% FPR

Equal Error Rate (EER) — the point where false acceptance equals false rejection — ranged from 8.2% to 14.1%.

Regulatory Context

Deepfakes of real persons face specific disclosure requirements under the EU AI Act (enforceable August 2026) and are the target of several US state-level legislation. AI watermarking for video is an emerging technical requirement. See Standards for detail.

AI Watermarking →False Positive Rate →C2PA →Content Provenance →

Discuss this topic

Join practitioners, researchers, and publishers discussing AI detection methodology in the community forum.

Open Forum →