A Gaffer, Not a Director

A film set is one of the last places where hierarchy is honest. Everyone knows exactly what they do, and the good ones take fierce pride in the narrowness of it. The gaffer does not want the director's job. The gaffer wants the light to be perfect and the director to never once have to think about how. That relationship — total competence in service of someone else's vision — is the right way to think about artificial intelligence in creative work. Not as a rival author. As the best crew you have ever had.

We say this as people who build AI into production software for a living, which means we spend more time than most arguing about where it belongs. The argument almost always resolves the same way: the machine should hold the jobs that are demanding but not authorial, and it should be kept firmly away from the ones that are.

The line runs through authorship

Consider what a production actually asks of its people. A great deal of it is retrieval, bookkeeping, and propagation — necessary, unglamorous, and unforgiving of error. Which takes were circled. What the continuity was in the shot before. Whether moving the Tuesday scene breaks the actor's availability on Thursday. This is exactly the work a machine should carry, because it is laborious precisely where humans are weakest and it is checkable — there is a right answer, and getting it wrong is a mistake rather than a choice.

The machine should hold the jobs that are demanding but not authorial — and be kept firmly away from the ones that are.

Then there is the other kind of work: whether the scene should exist at all, where to put the camera, when to cut, which imperfect take is somehow the true one. None of these has a right answer. Each is a judgment, an assertion of taste, a small act of authorship. Hand these to a model and you have not saved labour — you have quietly replaced the artist with an average of everyone who came before them, which is the one thing art is supposed to resist.

Assistance is not authorship

The confusion in the current moment is that the same technology can do both, and does not announce which it is doing. A tool that drafts a shot list is assisting. A tool that decides the shot list is directing. They can share an interface and a underlying model and still be, morally, completely different machines. The whole discipline of building these systems well is keeping that distinction visible — to the user, and to yourself.

So we hold a rule, and we hold it in the product, not just in our talks:

• The machine fetches, drafts, checks, and propagates. It never decides.
• Every suggestion is reversible, legible, and clearly marked as a suggestion.
• The human's judgment is the input the whole system is organised around — not an obstacle to route past.
• If a feature only works by quietly making creative choices for someone, we do not ship it.

A gaffer who started overruling the director would be off the picture by lunch, no matter how good the light. We think software has earned exactly the same deal: extraordinary competence, real trust, and not one inch past the line where taste begins. AI assists. The creative call is always human.