I Watched an AI Cut a Scene Better Than I Did. Then I Watched It Miss the Point.

There's a moment in the editing bay, around hour eleven, when the footage stops being footage and becomes a series of decisions you're afraid to make, and it was in exactly that state that a colleague sat down next to me, opened a browser tab, and said watch this.

He fed forty minutes of dailies into a generative tool and asked it to assemble a ninety-second cut. It did it in about four minutes. And the cut was, I have to be honest, competent. Clean. It found the reaction shot. It found the beat where the actor's hand trembles. It understood, in some statistical way, that a scene about a man losing an argument should breathe at the moment he stops talking.

I have been cutting for nineteen years. I have never once assembled a scene in four minutes that was that clean.

And then I watched it again, and I understood what was missing, and the missing thing is the entire job.

The machine edits the scene you described, not the one you shot

Here's what the tool did: it made the scene legible. Every cut served comprehension. Every reaction shot landed exactly where the audience would expect, and the whole thing moved at the pace of a well-made advertisement.

Here's what it couldn't do: it couldn't decide that the scene should be worse on purpose. It couldn't hold on a face two beats too long because the character is lying and the audience should start to suspect it before the other character does. It couldn't cut away to a detail that means nothing yet and everything forty minutes later.

Because a cut isn't just a way of conveying information. A cut is an argument about what matters. And an argument requires someone who thinks something matters.

There's a moment in Aguirre, the Wrath of God where the camera just keeps going, past the point of comfort, past the point where a reasonable editor would have cut, and the whole film's thesis about madness lives in that refusal. No model produces that refusal. Not because it lacks the capability, but because refusing is not a thing you do when your objective function is "satisfy the prompt."

What "AI video" actually is right now

Let me be precise, because the discourse is mush.

There are three different things being sold under one name.

Generation is making pixels that didn't exist. Sora-class tools. Genuinely astonishing, genuinely not filmmaking. A generated shot has no coverage, because there's no set to return to. You can't get the reverse angle by walking the camera around, because there is no camera and no room. You get one shot. If it doesn't cut with the next one, you regenerate, and now nothing matches.

Assembly is what my colleague's tool did. It's a sort. It's very good at sorting. Sorting is maybe fifteen percent of editing.

Prediction is the streaming algorithm deciding what gets made in the first place. This is the one that should scare you, and it's the one nobody puts in the demo reel.

The part that actually changed my work

I'm not going to pretend I don't use these tools. I use them constantly, and here's where.

Transcription and search, obviously. But deeper than that: I can search forty hours of footage for "the moment she realizes," and it finds candidates. Not because it understands realization, but because it's watched ten thousand moments where a face changes and it knows what those look like. That used to be three days of screening. Now it's an afternoon, and the afternoon is spent on choices instead of logistics.

The other one is versions. I can generate nine rough assemblies of a sequence overnight, and the morning is no longer about building from zero. It's about rejecting eight things, and rejecting is fast. Rejection is where taste actually lives anyway.

What I don't let it do is the last pass. The last pass is where you take a scene that works and you make it say the thing you actually believe, and that process is not efficient. It is, in fact, deliberately inefficient. It requires sitting with a frame long enough to get sick of it and then long enough after that to find what you were actually trying to say.

The economics nobody wants to talk about

Streaming algorithms have already restructured this industry in ways that had nothing to do with generation. They decide what gets greenlit based on the shape of what already worked. They ask for the second season before the first one has a chance to be misunderstood. They want the scene that keeps people from pressing pause, and the scene that keeps people from pressing pause is almost never the scene that stays with them.

That's not an AI problem. That's a business model problem that AI made cheaper to execute. But the two are entangled now, and anyone who tells you they can be separated is selling something.

Here's the version I believe: the tools will get good enough that anyone can make a competent-looking thing. Competent-looking things will become worthless, because there'll be infinite supply. What will be scarce is the thing that only one person would have made, the one with the weird refusal in the middle, the one that holds too long.

So what do you do

If you're a young editor, learn the craft anyway. Not because the tool won't do it, but because you cannot break a rule you don't understand. The reason I could see what the machine missed is that I've spent nineteen years learning why cuts work, which means I can see the moment a cut is technically correct and emotionally false.

And if you're building these tools, please stop measuring success by how much your output looks like the films we already made. We have those. We don't need them again, smoother.

The frame is always lying.