Protecting the Master

This is the least exciting chapter in the course, and it's the one I'd make mandatory if I could. Because here's the thing about AI tools that nobody puts in the demo reel: they fail weird. A traditional filter fails loudly — wrong settings, obvious result, undo. AI tools fail quietly. An upscaler invents texture on a face that was never there. A dialogue isolator eats the room tone that made the scene feel real. A color pass drifts your skin tones two clicks warm and you don't notice until the festival screening. The failure shows up late, subtly, and sometimes permanently.

So before we run a single AI process in this course, we install the iron rule: AI processes copies. Never originals. Ever. Your camera masters are the negative. No tool — no matter how good, how expensive, how "non-destructive" the marketing claims — ever gets write access to the negative. Everything AI touches is a duplicate you can throw away without a second thought.

The rules, in order

The iron rule turns into a working system with a handful of habits. Here they are, in the order they save you:

• Originals are read-only. Camera masters go into an originals folder the day you ingest, and nothing writes to that folder again — you, your NLE, or any AI tool.
• Four folders, always. Originals / working / AI-processed / exports. Every file in the project lives in exactly one, and AI output never lands anywhere but its own folder.
• Name versions so the process is visible. Something like scene04_t03_denoise_v2 tells you what was done and how many times. "final_final_REAL" tells you nothing at 2 a.m.
• Keep the rollback path alive. Every AI-processed file must trace back to an untouched source in one step. If you can't roll back, you don't have a workflow — you have a gamble.
• Back up 3-2-1. Three copies of the originals, on two different types of storage, one of them off-site. AI didn't invent this rule; it just raised the stakes on breaking it.
• Think before you upload. Cloud AI tools mean your unreleased footage is leaving your machine. Read what the service does with uploads before your rough cut lives on someone else's server.

None of this is glamorous. All of it is the difference between "the tool ruined a copy" and "the tool ruined the film."

Why rollback is the whole game

Here's the deeper reason for the folder discipline. AI processing is generative at its core — even the cleanup tools work by predicting what "should" be there. Most of the time the prediction is right. Sometimes it's confidently wrong: hallucinated detail on a license plate, a smoothed face that reads uncanny on a big screen, a "cleaned" dialogue track that lost the breath before the line. You often can't judge these failures until you see the shot in context, days or weeks later. Rollback is what lets you change your mind at no cost. Version naming is what lets you find your way back. The folder structure is what makes both automatic instead of heroic.

And a word on privacy, because it's now part of workflow: an unreleased film is confidential material. Before you drag footage into any cloud tool, ask the boring questions — where is this stored, who can see it, is it used for training, can I delete it? For most projects the convenience wins and that's fine. But make it a decision, not an accident. A festival premiere with a "world premiere" requirement, a client NDA, an actor who hasn't approved a scene — all of these can be quietly violated by an upload you didn't think about. Local tools, including everything built into DaVinci Resolve and Premiere Pro, sidestep the question entirely.

Foundation set: you know what AI changes, what your stack looks like, and how to run it without risking the negative. Next chapter is the one that ties the whole course together — where automation ends and taste begins.