AI Video Detector logo AI Video Detector

AI video evidence

Understand Your AI Video Evidence Report

Learn how to read AI likelihood, confidence, suspicious signals, unchecked signals, and limitations before you act on a result.

Important: Reports are likelihood-based screening aids. They are not proof that a video is real or fake.

Reproducible sample · same video, same analysis

See what a report unlock adds

Watch real camera footage of a duck swimming in a pond, then compare the free preview with the full report from the same analysis. This licensed public sample is not a customer upload or a detection accuracy benchmark.

Video: Duck swimming at Knollenteich by PantheraLeo1359531 · CC BY 4.0. Resized and transcoded; source color-corrected by the author. Source attribution does not guarantee authenticity or imply endorsement.

3 sampled frames at 1.549, 4.646, 7.743 seconds · modal-ffmpeg-cv-audio-v11 · basic audio check

Free preview

Screening index: 44/100 · Inconclusive

Confidence indicator: 78/100. Neither value is a calibrated probability of AI generation.

  • Metadata reliability

    ffprobe extracted container, stream, codec, duration, and bitrate metadata.

    Signal value: 0.75

  • Compression anomaly

    Frame sharpness and histogram changes are checked for unusual compression/artifact patterns.

    Signal value: 0.078

4 additional recorded explanations: Temporal consistency, Texture entropy, Edge consistency, Audio-video consistency.

All limitations · visible in both
  • Visual indicators are rule-based screening signals, not a calibrated probability that a video is AI-generated.
  • A few sampled frames cannot cover every moment; normal scene changes can affect visual indicators.
  • This is a signal-based screening report, not proof.
  • CPU ffmpeg/OpenCV signals are weaker than a dedicated forensic or GPU deepfake model.
  • Social page URLs that cannot be safely fetched are not bypassed; upload/direct video URLs produce stronger reports.

Full · same existing report + PDF

Screening index: 44/100 · Inconclusive

Confidence indicator: 78/100. Neither value is a calibrated probability of AI generation.

  • Metadata reliability

    ffprobe extracted container, stream, codec, duration, and bitrate metadata.

    Signal value: 0.75

  • Compression anomaly

    Frame sharpness and histogram changes are checked for unusual compression/artifact patterns.

    Signal value: 0.078

  • Temporal consistency

    Sampled frames are compared for frame-to-frame consistency.

    Signal value: 0.775

  • Texture entropy

    Sampled frame luminance entropy is used as a lightweight texture signal.

    Signal value: 0.688

  • Edge consistency

    Canny edge density variation is used as a simple artifact proxy.

    Signal value: 0.966

  • Audio-video consistency

    Audio presence, duration alignment, silence, clipping, and level stability were checked with a low-cost audio sample. Paid reports can add sampled audio-motion windows. Semantic speech/video matching is not performed.

    Signal value: 0.003

Unlocking reveals these existing explanations and PDF export. It does not rerun analysis, add Pro audio analysis, or extend expiry.

All limitations · visible in both
  • Visual indicators are rule-based screening signals, not a calibrated probability that a video is AI-generated.
  • A few sampled frames cannot cover every moment; normal scene changes can affect visual indicators.
  • This is a signal-based screening report, not proof.
  • CPU ffmpeg/OpenCV signals are weaker than a dedicated forensic or GPU deepfake model.
  • Social page URLs that cannot be safely fetched are not bypassed; upload/direct video URLs produce stronger reports.
Inspect the actual sample output and reproducibility hashes

A good report should explain itself

A simple “AI” or “not AI” label is not enough. Social videos are messy. They may be edited, compressed, reposted, cropped, or partly synthetic.

An AI Video Detector report is designed to show the evidence behind the result, not just a single score.

What the report includes

  • AI likelihood: how strongly checked signals suggest AI generation or manipulation.
  • Confidence level: how reliable this specific check is.
  • Result state: likely AI-generated or manipulated, uncertain, no strong AI signals detected, or unsupported.
  • Signal summary: suspicious face, hand, background, motion, audio, or source clues.
  • Unchecked signals: missing audio, low quality, short duration, or limited access.
  • Recommended next steps: check the original source, compare other uploads, avoid accusations, or request expert review.

Score vs confidence

AI likelihood and confidence are different. A video can have a high AI likelihood but low confidence if the clip is short, blurry, or heavily compressed. A video can have a low AI likelihood but still not be proven real.

Use both numbers together. If confidence is low, treat the result as inconclusive.

FAQ

Questions this page answers

Why does the report separate likelihood and confidence?

Likelihood describes the strength of checked signals; confidence describes how reliable this scan is based on media quality and available evidence.

What should I read first in a report?

Start with result state, confidence, and limitations before reading suspicious signals.

Why can a report show unchecked signals?

Some inputs lack enough frames, audio, duration, or source access. The report lists those gaps instead of hiding them.