A clearer view of how the video analysis works

The report is built by reviewing multiple signals from the uploaded file, then combining them into an AI-likelihood estimate with confidence.

Analysis modules

What we analyze

4 active modules

Frame scan

Visual details

We inspect sampled frames for unusual textures, edge patterns, lighting mismatches, scene repetition, and compression behavior that can appear in manipulated or AI-generated media.

Motion trace

Temporal consistency

We compare nearby frames to look for flicker, shape drift, unstable motion, or details that change in ways that do not match normal video continuity.

Voice signal

Audio cues

When usable audio is present, we review speech rhythm, spectral patterns, and repeated artifacts that can suggest voice cloning or synthetic processing.

Origin trace

Metadata and provenance

We read available container metadata, codec information, timing fields, and provenance clues to understand how the file was packaged and whether any origin evidence is present.

Workflow

Step-by-step process

From upload to report, the pipeline breaks one review into a sequence of modules so each signal can be inspected clearly.

5 stages

Upload video

You provide a video file for analysis. The system works from that file only and prepares it for processing.

Extract frames, audio, and metadata

The video is broken into sampled frames, audio data when available, and file-level metadata so each signal type can be reviewed separately.

Analyze visual and temporal inconsistencies

We check frame quality, scene realism, motion continuity, facial stability, lip-sync alignment, and other patterns that can change when content has been synthesized or heavily manipulated.

Combine signals

The system weighs findings across visual, scene, temporal, audio, and metadata signals instead of relying on a single clue.

Generate an AI-likelihood result with confidence

The final report summarizes the strongest signals, estimates AI-likelihood, and shows how confident the system is based on signal quality and consistency.

Reading guide

How to read the result

Every scan is meant to be readable by normal users. Instead of a single verdict, the report shows the estimated AI-likelihood, how confident the system is, and which signals influenced the result.

AI-likelihood result

Primary

This is an estimate of how strongly the observed signals resemble AI-generated or manipulated video. It is not a legal finding or absolute proof on its own.

Confidence

Confidence reflects how clear and consistent the available signals were. Lower confidence can mean the file is short, noisy, compressed, or missing enough evidence for a strong conclusion.

Signal breakdown

Reasoning

The report shows which categories contributed to the result so you can review the reasoning instead of relying only on a single score.

Sample Report Output

Example of how a mixed-signal review can be presented.

Mixed signals

AI Likelihood76%

Confidence Score84%

Detection Insights

3 findings

Subtle face-detail instability appears across several frame transitions.
Speech timing and mouth motion diverge briefly in voiced segments.
Container metadata is readable, but provenance evidence is limited.

Review boundaries

Limitations

Very short clips, heavy compression, screen recordings, and edited re-exports can reduce reliability.

A low AI-likelihood result does not guarantee that a video is authentic.

A high AI-likelihood result should be reviewed alongside source context, provenance, and human judgment.

Review alongside source context and human judgment

Handling notes

Privacy

Analysis runs in a backend job after upload, where the service extracts the frames, audio, and metadata needed by the available modules.

Results are cached in browser session storage so the results page can be revisited during the same session.

Authentication, billing, and daily usage enforcement are handled by the backend. Uploaded media should still be treated as sensitive content.

Treat uploaded video as sensitive content