Where AI video tools get their pictures

Twenty-two tools were checked. Thirteen of them choose the pictures for you, and twelve of those thirteen tell you nothing about where any of them came from. One names some. None names the source of every clip.

Why this is the question

A finished video makes a claim, and a claim you can check is worth more than a picture that merely looks right. If a clip is presented as footage of a real event, the only thing that separates it from an illustration is a record of where it came from. That record also decides whether the video can be reused, whether a correction is possible, and whether a platform reviewing a copyright complaint has anything to look at.

So this audit asks one question of each tool, and nothing else: when it hands you a finished render, can you find out where a clip came from.

Not every tool is asked

Nine of the twenty-two do not choose pictures at all. They edit, caption or clip footage you already have, or they put a synthetic presenter on screen instead of a subject. Counting those against a question they are not asked would make the headline larger and the finding worthless, so they sit out of it and are listed separately below.

Where the pictures come from

One entry reads Not measured, and it stays that way on purpose. Fluxnote says it supplies footage and never says whether that footage is generated or drawn from a library. Choosing one for it would be exactly the failure this audit is about.

What a render tells you

Entry price is often not published either

A second thing worth recording while looking at 22 products: 13 of them publish no entry price on a page you can read without an account. Of the 9 that do, the cheapest starts at $10 a month and the most expensive at $149, with a median of $29.

The full table

Who published this

ViewMade is not in the audited set, because ViewMade published the audit and a product marking its own homework is not a measurement. For the record, and so the bias is visible rather than hidden: ViewMade cuts real archive footage and writes down where each clip came from, which is the column above it would score well on. Every other column is one a reader should check against the products themselves.

Limits

These readings were taken in August 2026 from what each product publishes and from using it, and products change. Source naming was read as what a finished render or its export actually carries, not what a marketing page promises. Where a product did not make an answer available, the table says so rather than guessing, and one row does exactly that.