Two answers get repeated about AI and filmmaking, and both of them are wrong. The first says AI will replace filmmakers, that the camera and the writer's room and the cutting room are all heading toward the same obsolescence as the carriage maker. The second says AI changes nothing, that it's a fad, a toy, a marketing word stapled onto ordinary software. Neither answer survives contact with what's actually happening on real productions right now. The honest question isn't whether AI will change filmmaking. It's already changing it. The honest question is narrower and harder: what, specifically, does it touch, and what, specifically, can it not touch — not yet, and maybe not ever. Answer that question precisely, department by department, and you'll know more than most people currently arguing about this on either side.

Start with Lumet, because he already drew the line you need, decades before anyone was asking this question. In Making Movies, Lumet says the most important decision he makes on any film isn't the cast, the schedule, or the visual style. It's answering one question: what is this movie about. Not the plot — he's explicit that plot alone can be enough for a good melodrama, and that's fine. He means something else: what does the story mean to him, personally, right now, in this period of his life. He calls the answer the riverbed every other decision on the production gets channeled into. And notice what kind of question that actually is. It isn't a calculation. It has no correct answer you could look up or average across a thousand other movies. It's a judgment a specific person makes out of a specific life, and the judgment only works because that life is actually behind it. A system trained on every movie ever made can tell you what answers other films have given to that question. It cannot tell you what this one means to you, because it has no you, and meaning isn't a pattern you retrieve — it's a stake you put down.

Murch makes the identical argument from the cutting room, and it's worth holding the two side by side because they're answering the same question from opposite ends of the production. His Rule of Six in In the Blink of an Eye ranks the priorities an editor weighs on every cut: emotion first, then story, then rhythm, then the more technical concerns of eye-trace, planarity, and spatial continuity, in that order, every time you have to sacrifice one for another. Emotion sits at the top because, as Murch puts it, what an audience finally remembers isn't the editing or the camerawork or even the story — it's how they felt. That's not a measurement. There's no instrument that tells you the correct emotional temperature of a scene. There's only a person who has decided how they want another person, a stranger in the dark, to feel at this exact moment, and who can tell, cut by cut, whether that's happening. Lumet's question and Murch's priority list are the same kind of thing wearing different clothes: a human being deciding what something means and how it should land, with nothing underneath that decision except their own taste and their own history. That's the whole argument of this post in two sentences. Judgment isn't a step in a process you can speed up. It's the thing the process exists to serve.

Murch, more than anyone else in this curriculum's source library, actually saw this exact debate coming, and it's worth reading what he wrote about it before generative tools existed in any usable form. Writing at the end of 1999 about where digital technology was heading, he posed himself a thought experiment: imagine a device that could convert a single person's thoughts directly into a finished film, no crew, no actors, no compromise. Would you take that deal. He answers by sorting filmmakers into two temperaments. One is Hitchcock's, who said the film was already made in his head before he ever started shooting — a director for whom collaboration is a necessary cost of getting a fully-formed private vision onto a screen. The other is Coppola's, who described his own role as the ringmaster of a circus that is inventing itself — a director for whom the vision only really exists once it's been run through other people's talents and surprised him along the way. Murch's own answer to his thought experiment wasn't fear. It was a specific caution: watch for anything that encourages what he called a hermetically personal vision while discouraging the friction of real collaboration, because the history of painting and classical music in the twentieth century shows what happens when a solitary vision stops being checked by anyone else. That's the actual risk in front of you right now, a quarter century later, with tools Murch never lived to see by name. Not that a machine will out-imagine you. That you'll start treating your own collaborators as a delay between you and a vision a machine can now produce faster than they can — and lose the very friction that used to make the work better than what was in your head to begin with.

Mackendrick comes at the same problem from the opposite direction, and his version matters because it corrects a misunderstanding baked into the phrase "AI will replace the director" before AI even enters the sentence. Mackendrick spent a career trying to stamp out what he called the cult of the director — the idea that one person's name on a film means one person generated everything in it. His own description of the job: the director is nothing but the channel of other people's talents, the figure who has to dissolve into the work rather than stand in front of it. If that's true — and anyone who has actually run a set knows it is — then "AI replaces the director" was already a confused sentence before any software existed, because the director's job was never to be a solitary generator of content. It was to hold the throughline across dozens of other people's judgment calls: the cinematographer's, the editor's, the actor's, the composer's. A tool that can generate images or dialogue on its own doesn't replace that job. It just adds one more collaborator to manage, one whose judgment, unlike everyone else's on the call sheet, doesn't actually exist yet.

Now to the practical part, because abstraction without specifics is exactly the kind of thing that makes this subject easy to dismiss or easy to panic about, and you deserve neither. Start with pre-visualization, since it's where AI tools have moved fastest and where the case for them is genuinely strong. Robert Rodriguez, in Rebel Without a Crew, describes training himself for a decade to see an entire film cut together in his head before he shot a single frame of it, a skill forced on him by editing on two consumer VCRs that gave him no second chances. That internal eye — knowing which version of a scene actually works before you've spent the money to find out — is exactly what modern AI previsualization tools are now selling at industrial scale. Industry estimates circulating in 2025 put the figure at roughly eighty percent of Indian films already using AI extensively somewhere in their previsualization process, generated through platforms built specifically because most working directors aren't fluent in prompting software and shouldn't have to be. That's a real, useful, unglamorous tool. It lets you generate and discard ten visual options for a sequence in the time it used to take to discard one. What it cannot do is replace the internal eye Rodriguez built through ten years of having no other option. The tool can show you options. It still needs you to recognize which one is actually right, and that recognition is the same trained judgment it always was, just exercised faster.

Editing assistance follows the same pattern. Murch's own description of an editor's work splits into a horizontal task — what comes next — and a vertical one — what's happening at the same time within the frame. AI tools built for editors right now are almost entirely horizontal: searching transcribed footage for every take where an actor says a particular line, assembling a rough sync of dialogue across dozens of takes, flagging continuity mismatches a human eye might miss on hour fourteen of a long day. That's genuinely useful, and it's also exactly the kind of mechanical retrieval task Murch's Rule of Six was never trying to solve. The rule exists to help a human decide which of those retrieved options actually serves the emotion of the scene. No tool currently sorts footage by which cut will make a stranger in a dark theater feel something specific, because that sorting requires knowing what you wanted them to feel in the first place — which returns you, again, to a decision nothing outside a person has access to.

You'll find the same honest middle ground in how working directors are actually choosing to use these tools, rather than how the trade press describes them. Shakun Batra, the writer-director of Kapoor & Sons and Gehraiyaan, was among the first mainstream Indian filmmakers to experiment seriously with generative tools, building a company called Jouska AI that uses systems like Midjourney, Veo, and ElevenLabs for mood boards, world-building, and early development work. His own framing of where this goes is worth taking seriously precisely because he isn't selling the tools: the future he describes is hybrid, where actors are still shot and directed the traditional way, and AI gets used for environments and sequences that would otherwise have required resources a smaller production simply doesn't have. Notice what's absent from that description. Nothing about AI deciding what a scene means. Nothing about AI replacing the work of getting a real performance out of a real actor on a real day. Just a working filmmaker drawing the same line this post has been drawing from Lumet and Murch, in his own words, from inside his own production company.

Scheduling and script-breakdown tools deserve their own honest word here too, even though they generate far less debate than anything involving an actor's face or voice. Software that reads a script and flags every location, every prop, every speaking part, and drafts a shooting order around them used to take an assistant director days of careful line-by-line work. AI versions of that same task now do a first pass in minutes, and nobody serious is arguing that this diminishes the craft, because nobody ever claimed a shooting schedule was where the art of filmmaking lived. It's the clearest case in this entire post of a tool doing pure process with no judgment attached, which is exactly why it's the least controversial use of AI on a production right now, and exactly why it's a useful baseline for everything else: the closer a task sits to logistics, the less anyone worries about a machine doing it, and the closer a task sits to meaning, the more that worry is actually earned.

Dialogue and VFX cleanup tools are where you'll find the clearest real-world test cases right now, because two recent productions show both how well this can work and how badly it can fail, and the difference between them is instructive. The Brutalist, in 2024, used a Ukrainian company's voice technology called Respeecher to refine specific vowels in Adrien Brody and Felicity Jones's Hungarian dialogue, after a dialect coach had already spent months training their performances. When the use of AI became public and provoked a backlash, director Brady Corbet's clarification was precise: the actors' performances were completely their own, and the technology touched only pronunciation accuracy in a language neither actor natively spoke. That's AI working exactly where it belongs — inside an existing department, sound editing, correcting a technical detail downstream of a performance choice the humans involved had already made and owned.

Yash Raj Films' War 2, released in 2025, gives you the same pattern at a larger scale. To release the Hindi film in Telugu without reshooting a single scene, the studio used a Bengaluru company called NeuralGarage and its tool VisualDub, which adjusts an actor's lip and facial movement to match dubbed audio closely enough that the Telugu version received its own certification rather than being classified as a dub. The company's own description of the goal is worth sitting with: preserving the performance, identity, and speaking style of the actor while changing only what's needed to make the new language look native. Again: a real, named, currently deployed tool, solving a real distribution problem — Indian cinema's deep multilingual market — without claiming to make any decision about what the scene means or how the line should land. A dubbing artist and a director still decided that. The tool just made the mouth match what those humans had already chosen.

Be honest about the cost side of this too, because a department built on critical commentary shouldn't flinch from it. These same dubbing tools are already displacing real, paid, technical work, not hypothetically — Ghazal Khanna, a veteran Indian voice artist who has dubbed Netflix titles for the Indian market, estimates that somewhere between seventy and eighty percent of the brand voices in major Indian television and video commercials have already shifted to AI. That's not the judgment layer of the craft disappearing. Nobody's claiming the AI decided what the advertisement should mean. But the technical-execution layer underneath that judgment — the working voice artist who used to get paid to perform it — is a real job, held by a real person, and it is currently being lost at a real rate. Refusing both extremes doesn't mean refusing to notice that one of them is currently true for a specific kind of worker, even while the other extreme stays false for the kind of judgment this whole post has been describing.

Now hold the dubbing story against what happened to Raanjhanaa. The 2013 film, directed by Aanand L. Rai, ends with its protagonist shot and dying in a hospital, a death the whole shape of the story has been building toward — the cost of an obsessive, one-sided love the film never lets you look away from. In 2025, the studio Eros re-released the film with an AI-generated alternate ending in which the character survives, made without the consent of either the director or the lead actor, Dhanush. The technical execution, by every account, was convincing — the character sits up, opens his eyes, the room fills with relief instead of grief. And it was, by any measure that matters to this curriculum, a complete failure, because changing whether a character lives or dies in a film like this isn't a technical detail. It's the answer to Lumet's question. It's what the movie is about. Dhanush's own public response named exactly what had gone wrong, in language a craft book couldn't improve on: the new ending had stripped the film of its very soul. He's not speaking as a technologist. He's speaking as the person who knew, from the inside, what that death had cost the story to earn — and the tool that generated his survival had no access to that knowledge, because that knowledge was never data. It was a judgment, made once, by people who had reasons for it the system was never given.

Which brings you to what actually stays irreducibly human in all of this, and it's worth being concrete rather than sentimental about it. Lumet opens Making Movies with a story about Kurosawa, who once explained a specific camera framing on Ran by saying that an inch to the left would have exposed a Sony factory in the background, and an inch to the right would have shown an airport — both wrong for a period film. Lumet's point is simple: only the person who made that choice actually knew what was behind it. Nobody could have generated that decision from outside Kurosawa's own knowledge of that exact location on that exact day. Aamir Khan's decision to shoot Lagaan in synchronous sound gives you the same lesson from the opposite direction. Everyone around him, by his own account, told him not to — it had never been done that way out of Bombay, and the entire existing convention of the industry argued against it. He did it anyway, because of what he believed it would do for the truth of the performances, and it worked. That decision couldn't have come from a system trained to predict what's worked before, because the entire point of the decision was to go against what had worked before. Taste like that, conviction like that, only comes from a person's specific accumulated experience of what they believe a story needs — and that's not a gap in the technology that better training data will eventually close. It's a different kind of thing than the technology produces at all.

So here's the practical version of all this, for right now, with no hype and no shortcuts. Use these tools exactly where they're already proving themselves: previsualization that lets you test ten ideas before committing to one, footage search that saves you the hour you'd spend hunting for a single take, dialect and dialogue cleanup that solves a real technical problem after the performance is already locked, scheduling and breakdown tools that do the same arithmetic a production manager has always done, just faster. Use them the way Corbet used Respeecher — to refine something humans had already decided, not to decide it for you. And hold onto one honest test for the line you should never cross: if you ever find yourself asking a tool what your film is about, or how a moment should make someone feel, stop. You've just handed away the only part of this job that was ever actually yours to begin with — the part Lumet built an entire book around, the part Murch put at the top of his list, the part no amount of training data will ever have lived through except you.

Everything else in this department picks up from here, department by department: where these tools currently sit inside cinematography and previsualization, what they're actually doing right now inside editing and VFX pipelines, what the dubbing and localization shift already underway in the Indian industry means for the people whose jobs it touches, and what questions you'll need real answers to before you let any of this near a project that matters to you. None of it will be useful to you, though, if you arrive thinking the tool's job is to tell you what to make. It isn't, and on the evidence so far, it can't be. Your job was never threatened by what these tools can do. It was only ever threatened by your own willingness to stop doing it.