Every few weeks a tool arrives that does something the last one could not, and the conversation resets. Can it do lip sync now? Can it hold a character across shots? Can it match a grade? These are real questions and they have real answers, and the answers are worthless within about six months.
The question that does not expire is a different one: what are you willing to hand over, and how would you know if you were wrong to? That question is the same in 2026 as it was in 2023, and it will be the same when the current generation of tools is embarrassing to look at.
The 4Ds are how we teach it. They are not a maturity model or a checklist to be completed. They are four tests you run against a specific task, and any one of them failing is a reason to keep the task.
Delegate — what can be handed over without loss?
The safest candidates share two properties: the work is mechanical, and the result is reversible. A first-pass transcript. A rough assembly you will restructure anyway. Forty variants of a location concept you will throw thirty-nine of.
What makes these safe is not that they are unimportant. It is that a bad result is visibly bad and costs you an hour. The danger zone is work where a mediocre result looks like a fine one — and that is most of the interesting work in a film.
Describe — can you specify what you want?
This is the test that catches people out, because it looks like a prompting problem and it almost never is.
If you cannot describe the thing you want with enough precision for another person to attempt it, the tool is not your bottleneck. Most disappointing output is an unclear intention rendered faithfully. The model did exactly what was asked; the asking was the vague part.
The useful side effect is that this test improves the work whether or not you end up using a tool at all. A director who can specify what they want gets better results from a DoP too.
Discern — can you tell whether it is any good?
This is the load-bearing one, and it is where the real risk lives.
A tool that produces plausible work for someone who cannot evaluate it is worse than no tool at all, because it removes the friction that would otherwise have told them they were out of their depth. A student who cannot yet hear a bad sound mix will accept an AI-cleaned track that has quietly destroyed the room tone. They have not saved time. They have shipped a worse film faster, and learned nothing.
This is the argument against the common claim that these tools democratise filmmaking. They lower the cost of producing output. They do not lower the cost of judgment, and judgment is the part that was scarce.
Practically: if you cannot yet evaluate the output of a task, that is not a task to delegate. It is a task to learn.
Diligence — what must be checked regardless?
The fourth test is not about quality. It is about the things that remain your responsibility no matter how good the output is: consent, provenance, disclosure, and anything a viewer is entitled to assume is real.
In fiction there is room to argue about where these lines sit. In documentary there is not. An audience watching nonfiction is extending a specific kind of trust, and a synthetic element that goes undisclosed does not just breach an ethical guideline — it makes the film a different kind of object than the one the audience thinks they are watching.
Our own position on this is on the method page, and worked through against a real film in the 40 Years of Silence case study.
Why four, and why in this order
The order matters. Delegate and Describe are about the task. Discern is about you. Diligence is about the audience. Working through them in sequence means you cannot arrive at the ethical question after the creative decision has already been made — which is how most of these decisions actually go wrong.
None of this tells you which tool to buy, and that is deliberate. The calibration instrument runs the framework against eighteen documented limitations, and the letter generator in Assistant Editor Essentials is built on it. Both are free.