Substantive errors are corrected in place
Where a published claim is wrong, it is fixed on the page it appeared on, not quietly dropped. The changelog records the date and what changed, so a reader who cited an earlier version can see what moved.
Method
How we decide what
is true enough to publish.
This page exists because a site that tells filmmakers what AI cannot do should be willing to say how it knows. Everything below is a standard we hold ourselves to, and you are entitled to hold us to it as well.
The standard
Claims come from practice, not benchmarks
Every claim published here originates in a real production or a real teaching context. Nothing is derived from vendor documentation, launch demos, or benchmark scores, because none of those predict how a tool behaves in the middle of an actual edit with an actual deadline.
A limitation is documented when it is encountered, then checked against repeat attempts before it is published. If it happened once and never again, it was a fluke and does not go up. If it recurs under the same conditions, it is structural and it does.
The bar for publishing
Provisional, dated, and open to being wrong
Findings are published with the date they were established and the conditions they were established under. Tools change monthly; a limitation documented in June may not hold in December, and a claim without a date is not a claim you can check.
Where a question has been raised but not answered, it is published as an open question rather than resolved with a guess. That is slower and less quotable. It is also the difference between documentation and opinion.
When we get it wrong
Corrections policy
We will get things wrong. The question is what happens next.
Withdrawn limitations stay visible
When a tool improves and a documented limitation stops being true, it is withdrawn and the withdrawal is logged. A limitations list that only ever grows is a list nobody is checking.
Corrections are not negotiable
A vendor is welcome to tell us a claim about their tool is wrong, and we will test it again. What decides the outcome is the retest, not who asked.
Tell us
If something here gave you a wrong answer, that is the most useful message we can receive. Report it — there is a dedicated option on the contact form.
Disclosure
Who pays for this
No affiliate links. There are none anywhere on this site. No tool listed in the market map or the tools index carries a referral code, and no click here earns us anything.
No sponsored placement. No vendor has paid, or can pay, to appear in the tools index, the market map, or any article. Inclusion follows from use in real work.
What does pay for it. Training and advisory work. The tools and the writing are free and stay free; they are how the thinking gets tested in public, and they are the reason anyone would hire us. That is the whole business model, stated plainly.
In a field where most tool coverage is affiliate-driven, saying this is worth more than it should be.
Disclosure
How AI is used to make this site
It would be incoherent to publish a site about honest AI use without saying how AI was used to build it. So:
AI assistance is used in drafting and in building this site — copy, code, and the design tooling. That is disclosed here rather than left for you to guess.
AI is not used to generate claims, findings, or evidence. Every substantive assertion on this site — every documented limitation, every statement about what a tool does in practice — originates in work a human did and verified. A model can help say a thing more clearly. It cannot be the reason the thing is believed.
Nothing on this site is a synthetic account of something that did not happen. No fabricated case studies, no invented testimonials, no illustrative examples presented as real ones.
Position
Ethics in nonfiction work
These four lines were drawn on a specific documentary and have since become the standing position. They apply to our own production work and they are what we teach.
No generative reconstruction of testimony
Nothing a contributor says, and no image of a contributor, is generated, extended, or synthetically altered. In nonfiction a synthetic voice is not a production shortcut; it is a betrayal of the subject.
Analysis, never synthesis
AI tools may read the record; they do not add to it. Transcription and structural review analyse material that already exists. The moment a tool would put something new in front of an audience, it is out of scope.
Consent as a process
Consent is revisited as a film evolves, because agreeing to an interview is not the same as agreeing to the film that interview ends up inside. This costs time in post. It is not optional.
Disclosure on the record
Where AI tools are used, that use is documented at the level of specificity a festival, a broadcaster, or an ethics committee would need. Nonfiction earns trust by showing its working.
Found something wrong?
A correction is more useful to us than a compliment. If a claim here does not match what you found in practice, we want to know.