Is there an AI tool for documentary rough cuts?

The part of a documentary that is text

An interview-led documentary has two layers. One is what people say: the argument, the chronology, the admissions. The other is everything else: faces, places, archive, silence, music. The first layer can be read. The second has to be watched.

Most of the early structural work happens in the first layer, which is why editors have always made paper edits from transcripts.

What a language model is good at here

• Reading all of the transcript at once and holding it in view.
• Finding where a subject makes the same point twice, and where the clearest version is.
• Proposing an order: which statement sets up which.
• Returning the proposal as exact spans of the source, so nothing has to be found again.

What it misses

• Silence. A long pause before an answer can be the best moment in the film, and in a transcript it is nothing.
• Faces and delivery. Two takes with the same words are identical on paper.
• Picture. It does not know what is in frame, so it cannot choose B-roll or cutaways.
• Taste and ethics. What is fair to a subject is the editor's call.

Keeping the editor in charge

The useful question about any AI tool is what you can undo. In FirstCut every choice the model made is a clip you can move, shorten, delete or restore. Deleting is reversible. Every version is kept as a revision. A note can be attached to one beat or one clip, so a correction lands where you meant it.

The model selects and orders. It does not write lines, and it cannot add footage that is not in your files.

Privacy for unreleased material

With FirstCut the video file never leaves your computer. The browser extracts the audio track and uploads that for transcription. The speech-to-text provider receives the audio, and the language model receives transcript text only. The details are in how FirstCut edits your video without uploading it.

A realistic workflow

1. Import one interview, or the interviews for one strand of the film.
2. Generate a first structure and watch it.
3. Try a second structure as a new revision and compare.
4. Correct what a transcript reader could not know: restore a pause, swap a take, drop a line that plays badly.
5. Export a timeline and continue in your editor, where picture and sound are decided.