Is AI taking over filmmaking, or just the parts nobody enjoyed

AI is taking the boring half, not the whole job

No, AI is not taking over filmmaking. It is taking over the parts nobody enjoyed. A comparison rather than a prediction, the argument for that means this piece compares three ways of getting a frame made and says where each one genuinely wins. Strip away the boring half, the rotoscoping, the versioning, the transcription, the clean-up, and what is left is a performance in front of a lens and a decision about where to put it. That is the part the rest of this piece defends.

Search engines get the question as a threat. Is AI taking over filmmaking, yes or no. The honest answer is that it depends which stage of the job you mean, and most people asking have never had to separate the stages because they have only ever seen the finished film.

I do not think a model replaces a set. I do think it quietly deletes whole line items from a budget, and that is a different claim. One is a story about robots. The other is a story about arithmetic.

What a model actually does on a set

Three real options exist, and most arguments confuse them. The first is a generative tool doing one discrete task: a shot extension, a clean plate, a translated voice, a version. It sits inside a normal production and touches one stage. The second is a hybrid, where a model handles a defined block of the work, usually pre-visualisation or the social cut-downs, and a crew does everything else. The third is the conventional crew-led shoot.

On paper the three look closer than they are on screen. A generative tool produces a file. A hybrid produces a production with a file inside it. A conventional shoot produces rushes, and rushes are the only one of the three that carries a performance nobody planned for.

Keep the unit of work straight. A model is good at a task. A shoot is good at a day. Do not compare a task to a day and call it a contest.

The three options side by side

The dimensions that decide a job are not the ones the demos show. Turnaround, the cost of changing your mind, control over continuity, how it survives a client review, and who carries the liability when something is wrong. Everything else is taste.

Dimension	Generative tool, one task	Hybrid, model inside a normal production	Conventional crew-led shoot
Turnaround	Fast for a single asset. Slow if you need it to match anything else.	Fast on the stages it touches, normal on the rest.	Fixed by logistics. Locations, light and people take the calendar.
Cost to iterate	Cheap per attempt, expensive per usable result, because you sift.	Cheap on the model stage, unchanged on the shoot day.	Expensive. A redo means a call sheet, not a prompt.
Control over continuity	Weak over a sequence. A continuity sheet remembers the last shot in a way the model does not.	Strong, because a person holds the continuity and the model fills a gap.	Strong. That is what a script supervisor and a marked-up script are for.
Survives a client review	Variable. Looks fine on a phone, gets questioned on a big screen.	Usually yes, if the model work is hidden inside a real frame.	Yes, because the client can stand next to the monitor and see the take.
Who carries the liability	Unclear. The vendor sells capability, not indemnity.	Shared, and needs writing down before the first call sheet.	The production. Which is why insurance exists.

Read the table down, not across. The machine wins the top two rows. The crew wins the bottom three. That is the whole argument in one grid.

Where the machine genuinely wins

Some stages punish anyone who argues for a human out of stubbornness. Pre-visualisation is one. A director can block a sequence, test a lens idea and throw it away before anyone books a location. Versions and cut-downs are another: one master becomes 16:9, 9:16 and 1:1, and the reframing that used to eat an afternoon is now a pass you supervise rather than perform.

Clean-up is the quiet one. Removing a light stand, a cable, a logo on a shirt. Translation and sub-titling, especially across Indian languages, where a first pass that is roughly right saves a translator a day. A scratch track for timing an edit before the composer is anywhere near it.

None of that is the film. All of it is work that used to sit on somebody's Friday.

The warning is on the record. Filmmakers speaking on a Kling AI panel, as reported by Deadline in October, said that AI productions still require traditional skills to be cinematic, and that you need experts to prevent everything looking the same. That is the caveat attached to every win above. Speed comes to the tool because it averages, and averaging is the enemy of a look.

Where it loses, and loses badly

Continuity across a long shoot is where a model falls apart first. Not a single shot. A sequence. The same jacket, the same light on the same face, the same coffee cup in the same place across forty setups and three weeks. A crew holds that in a continuity sheet and a tired pair of eyes. A model holds it in a prompt you have to rewrite, and it drifts.

Performance is the second loss. A face that has to carry a story is doing something a model can imitate and not inhabit. The hard part is small. It is the pause before a line, the blink that arrives a beat late. That is the actor deciding, and you cannot average a decision.

Then the client. A client who needs to look at the frame and trust it. A frame that came out of a model can look right and still feel unearned, and the person paying can tell even when they cannot say why.

Think about how a colour grade works. Scopes can be perfect and the shot can still be dead, because the grade is not a measurement, it is a judgement made by someone looking at a face. A model can match the scopes. It cannot make the judgement. Same problem, different stage.

And when the deliverable is a likeness, a clearance or a contract, the model is not the authority. A lawyer is. I am not, and I say so before anyone signs anything.

The step people get wrong

Buying the tool before scoping the deliverable is the mistake. The tool is the easy purchase. The scoping is the work, and it happens before a single credit is spent.

I read the brief first. What is the film for, who watches it, and what does it have to make them do. Then I count the formats. A launch film with six cut-downs, three languages and a vertical set is a different job from one master. The count decides how much of the model is useful and how much is a distraction.

Then I decide which stage the model touches. Pre-vis, clean-up, sub-titles, a version. One stage, named out loud, written into the plan. Then I keep it away from the part the client is actually paying for, which is almost always the human moment the film exists to show.

The step people skip is naming the stage before they buy. Buy first and the tool starts looking for a job, and every problem starts looking like a nail. Scope first and half the time you find you did not need the tool at all, and the other half you find exactly where it saves a day.

What it does to a small production house

Line items that shrink are the ones nobody was proud of. Clean-up, versioning, transcription, the first assembly of sub-titles. Those hours come down, and they should.

The shoot day does not shrink. Locations, light, people, the travel to get to them. A house of over six years still budgets for the shoot day because the shoot day is where the film is actually made, and a model has never stood on a location waiting for the light to do the thing it does for eleven minutes.

What does change is the reflex. When a client asks for one more version, the old reflex was to quote a day of edit. The new reflex is to see if the model can do it in an hour. Sometimes it can, and charging for a day would be dishonest. Sometimes it cannot, and quoting an hour would be worse. I cannot always tell you which reflex is right, and I am not sure I trust either one completely. Good capital discipline says price the work that is actually done. The other reflex, the one that wants to fix everything with the new tool, is the one to watch.

Who you should actually ask

Route the question to the person who actually owns the answer. For what a model can and cannot do on the current version, ask the vendor. They will oversell it, but they know the spec sheet better than anyone in the room.

For the real cost of a redo, ask the client's finance team. They know what a missed date costs them, and that number is the one that decides whether a model is worth the risk on their job.

For likeness, rights and clearance, ask a lawyer. Not me, not a vendor, not a forum. That is a legal question and it has a legal answer, and getting it wrong is expensive in a way a bad frame is not.

For anything physical in India, ask a local operator. A local crew starts the day already knowing what a flown-in crew spends its first day discovering. Airport filming and aerial work need the longest lead time of anything on an India shoot, and no model shortens a permission.

Ask me what a frame needs to hold up on a screen, and which stage of the work a model should and should not touch on a job I am running. Outside that, ask the people above. I mean it.

The parts nobody enjoyed were never the film

Scaffolding was always the boring half. Clean-up, versioning, the continuity sheet, the sub-title pass. It held the film up and it was never the film.

So the machine can have the scaffolding. That is not the loss. The loss, if there is one, is further out. A generation that never learns the scaffolding may never learn why the frame is built the way it is, because the scaffolding is where you learn it. You learn continuity by holding it. You learn a cut by making a bad one and watching it fail. Strip that away and you get people who can generate a shot and cannot tell you why it should be the next one.

Is AI taking over filmmaking. No. It is taking the parts nobody enjoyed, and the parts nobody enjoyed were the school. The open question is whether you can skip the school and still know what to do with the frame when it arrives.