AI Review is a technical assistant that runs on a video before humans watch it. It catches the mistakes that are embarrassing to receive from a client — clipped audio, a black frame at the end of an export, a typo in a lower third — so your reviewers can spend their attention on the cut itself.
Measured, not guessed
Everything in this group is measured straight from the file, frame by frame. Each finding says in its description that it came from a measurement, and no AI decides it:
- True peak above safe limit — peaks above −0.5 dBTP, the real peak level of the audio. Platforms audibly clip anything higher when they compress the file again.
- Loudness above web target / Loudness well below web target — average loudness far from the ≈ −14 LUFS the web expects. LUFS is the standard measure of perceived volume.
- Extended silence mid-video — a long gap of silence in the middle of the cut, where the audio probably dropped out.
- Black segment detected — with the exact in and out points.
- Frozen image detected — the picture stops moving while the video keeps running.
- Persistent side bars (pillarbox) and Persistent top/bottom bars (letterbox) — dead bars down the sides, or across the top and bottom, for the whole video. The classic wrong-aspect-ratio export.
- Interlaced content detected — the frames are interlaced, the old broadcast structure of two half-images per frame. Web delivery should be progressive: export again as progressive.
The thresholds are deliberately wide. A rule fires only on values that would be wrong for any web or social delivery, not on values that are merely unusual.
Read by the AI
On top of the measurements, the AI reads the video for the things that need reading:
- On-screen text and captions — misspellings, broken or unreadable text, and text that contradicts what is said.
- The spoken content — a word-level transcript of the video.
- Script adherence — when the task has a script attached, what was actually said is compared against it. See Transcription and script check.
- Client preferences — standing rules distilled from a client’s past feedback, managed in the client’s IA tab. Violations arrive titled “Client preference: …”.
What every finding carries
- a timestamp, so one click puts the player at the moment;
- a severity — HIGH, MEDIUM or LOW;
- a category (audio, video, text, export);
- a description saying what was measured or seen, and what to do about it.
What it will not do
- It will not approve or reject a video. Nothing changes status because of an AI finding on its own.
- It will not judge creative decisions — framing, pacing, music, narrative.
- It will not decide intent. Black segments, freezes and letterbox bars are reported as “verify whether this is intentional”, because the measurement is certain and the intention is not.
What’s next
Read how to work through the report, and what to do when a finding is wrong in Accuracy and false positives.