AI Video Detector

Trust & Transparency

Test Examples

How we test the detection pipeline and what the results mean. All test data is created internally — these are not real user scans.

Test methodology

We maintain an internal test set of videos created specifically for testing detection accuracy. This set includes AI-generated videos (from multiple generation methods), face-swap deepfakes, real recordings from various devices, compressed/re-encoded videos, and edge cases like animation and gaming content.

Each test case is run through the full detection pipeline — including frame extraction, internal analysis, provider analysis, and result aggregation — to verify that signals are detected correctly and edge cases are handled appropriately.

Important: Our test set is small and does not represent the full range of AI-generated or real videos in the wild. Test results are directional, not statistical accuracy claims.

Test cases

Text-to-video AI output

Expected: Likely AI — High confidence

A 10-second clip generated entirely by an AI model from a text prompt. Contains typical AI artifacts: unnatural motion, inconsistent textures, and structural distortions.

Result

Score 82, Verdict: Likely AI, Confidence: High. Multiple signals detected across temporal consistency, texture, and structure.

Source: Created with an open-source text-to-video model for internal testing.

Limitations: This test uses one specific AI model. Results on other models (Sora, Veo, etc.) may differ. We cannot test against every AI generation tool.

Face-swap deepfake

Expected: Likely AI — Medium confidence

A 8-second clip where one person's face is swapped onto another's body. Contains boundary artifacts and blink irregularities.

Result

Score 71, Verdict: Likely AI, Confidence: Medium. Face-related signals detected at jawline boundary and eye region.

Source: Created with open-source face-swap software for internal testing.

Limitations: Face-swap quality varies widely. High-quality swaps with consistent lighting may produce lower scores.

Real smartphone recording

Expected: Likely Real — Medium confidence

A 15-second street scene recorded in natural lighting with a phone camera. No filters, no editing.

Result

Score 12, Verdict: Likely Real, Confidence: Medium. No significant AI-generation signals found.

Source: Recorded with a standard consumer smartphone, no post-processing.

Limitations: This confirms the detector does not flag all videos as AI. However, a low score on one real video does not guarantee low scores on all real videos.

Compressed social media reupload

Expected: Unclear — Low confidence

A real video that has been compressed multiple times, simulating what happens when content is downloaded and re-uploaded across platforms.

Result

Score 38, Verdict: Unclear, Confidence: Low. Platform compression warning triggered. Some signals detected but below threshold.

Source: Original real video re-encoded 3 times to simulate social platform compression.

Limitations: Compression is a major source of uncertainty. The same video at original quality would likely score lower (more clearly real).

Animated / motion graphics content

Expected: Likely AI — Medium confidence (potential false positive)

A 6-second animated clip using standard animation tools. Synthetic by nature but not AI-generated.

Result

Score 55, Verdict: Unclear, Confidence: Medium. Texture and motion signals detected consistent with synthetic content.

Source: Traditional 2D animation created with standard animation software.

Limitations: This is a known false positive scenario. The detector identifies synthetic visual patterns, which includes traditional animation. Users should consider content type when interpreting results.

What these tests do and do not prove

Tests do show

  • The pipeline runs end-to-end without errors
  • Known AI-generated content is detected with appropriate signals
  • Real content is not always falsely flagged
  • Compression reduces confidence as expected
  • Known edge cases (animation, filters) are handled with appropriate warnings

Tests do not prove

  • Statistical accuracy rates (e.g., "95% accuracy")
  • Detection capability against all current or future AI models
  • Consistent results across all video types, qualities, and lengths
  • That the detector works equally well in adversarial conditions
  • That any specific real-world scan result is correct