Delegate
What can be handed over without loss? Usually the mechanical and the reversible: a first pass, a rough transcript, a variant set you will discard most of.
Signature instrument
Eighteen documented limitations across six clusters, and a practice studio built on the 4Ds — Delegate, Describe, Discern, Diligence — for deciding what a tool should and should not touch.
Abstract
Most guidance on AI in filmmaking is organised around capability: what a tool can now do that it could not do last year. That framing is nearly useless to a practitioner making a decision, because capability changes monthly and the decision does not.
This instrument is organised the other way round. It documents where these tools fail, in what circumstances, and with what consequence for the work — then gives you a structure for deciding, per task, whether to hand something over.
Method
Every limitation here came out of a real production or teaching context, not from testing tools against benchmarks. Each was encountered, documented at the time, and then checked against repeat attempts to establish whether it was a one-off or structural.
The six clusters are not a taxonomy imposed in advance. They are where the failures actually grouped once there were enough of them to group.
Limitations that later stopped being true get removed, and the changelog records the removal. Nothing here is claimed as permanent.
The framework
The part that outlasts any particular tool. Interfaces change within the year; the question of what you are willing to hand over does not.
What can be handed over without loss? Usually the mechanical and the reversible: a first pass, a rough transcript, a variant set you will discard most of.
Can you specify what you want precisely enough to be understood? If you cannot describe it, the problem is not the tool. Most bad output is an unclear intention faithfully rendered.
Can you tell whether what came back is good? This is the load-bearing one. A tool that produces plausible work for someone who cannot evaluate it is worse than no tool at all.
What must be checked regardless? Consent, provenance, disclosure, and anything a viewer is entitled to assume is real. In documentary this is not negotiable.
Live on this page
The instrument runs here in full. Open it full-screen if you prefer more room.
Reuse
This instrument is free to use, quote, and teach from, including commercially, provided it is attributed. If you are citing it in academic work:
Nagy, M. (2026) What AI Cannot Do in Filmmaking: a calibration instrument. CutToAi. Available at: https://cuttoai.net/calibration.html
The permalink above is stable and will not be moved. If a limitation is revised or withdrawn, the changelog on the tools page records what changed and when, so a citation can be checked against the version you read.
If you use it with students, we would genuinely like to know how it went — particularly where it did not work. Tell us.
The instrument is free as it stands. If you want it taught alongside the reasoning behind it, or embedded in a module, that is what training is for.
No schedule, no filler. You will hear from us when a tool ships, a claim gets tested, or something turns out to be wrong.
One email at a time, unsubscribe in a click. What we collect.