Not every AI film clip is the same kind of evidence

The shot

Three sources published within a few years of each other present different kinds of evidence about AI in filmmaking. Runway's AI Film Festival names Grand Prix winner Riccardo Fusetti's 'Generation' and eight other prize-winning shorts, each credited to an outside filmmaker who submitted a finished work. A New York Magazine feature published in June 2025 quotes Asteria co-founder Natasha Lyonne describing, in her own words, that she 'had not figured out at all how to use it in writing that I actually feel good about.' And OpenAI's own announcement of Sora 2, published 30 September 2025, shows curated clips captioned only with the prompt that produced them, such as 'a guy does a backflip.' These are not interchangeable.

What the documents show

The festival's winners page is a named, juried, credited record: a stated jury selected specific finished films from outside submissions, which is verified production use in the sense that a named person made and finished the work. Lyonne's quoted account is a public experiment: an individual's own report of her experience with a tool, offered without a finished, reviewed production behind it. OpenAI's launch post is a promotional demonstration: the company selected every clip shown and wrote the captions, and states plainly that 'the model is far from perfect and makes plenty of mistakes,' which is itself a vendor's own disclosure, not an independent measurement of how often those mistakes occur.

The workflow

Applying this framework to a new claim is a three-question exercise. First, ask who selected the example: a jury choosing from many submissions, an individual describing their own use, or the tool's maker choosing its own best output. Second, ask what documentation stands behind the claim: credited authorship and a public screening, a first-person account, or marketing copy alone. Third, ask whether a quoted benchmark number was run by the model's own maker or by a separate party under stated, repeatable conditions. Each step is answerable from what the source itself states, without special access.

What the tool does not change

None of the three source types substitutes for the others. A juried festival still depends on a jury's judgment about what counts as finished and good; an artist's own account still depends on that person's memory of their own experience; a vendor's demonstration still depends on the maker's choice of which output to publish. Treating a demo reel as a controlled test, or a single artist's account as industry-wide, is the error this framework is built to catch.

• Is the example in front of you credited to an outside author, a company, or the tool's own maker?
• Was the output selected from many attempts, and if so, by whom?
• Does a capability claim rest on the maker's own words, or on a separate, repeatable measurement?

This is an editorial classification, not a technical standard, and it will not resolve every ambiguous case. But naming which category a piece of evidence belongs to, before drawing a conclusion from it, is a discipline this site applies to every claim it reports.