Post-production has a different AI problem from every other part of the pipeline, and the difference is granularity. Nobody in post generates a standalone AI asset; they apply generative tools inside shots, as a paint-out here, an extended background there, an upscale on an archive clip, a cleaned-up line of dialogue. The AI use is real and it accumulates, but it is distributed across a timeline instead of sitting in a folder, so it leaves no natural trail of its own.
The trail matters because delivery paperwork asks for it. Broadcaster and platform deliveries come with QC processes and declaration forms, and when a form asks whether content was manipulated with AI, post is the department that actually knows the answer, shot by shot, or should. The person filling in the form is often a delivery or production manager who was never in the suite, and without a record they can only guess in one direction or the other.
Versioning multiplies the problem. Shots are revised across conforms, and the AI work belongs to a specific version of a specific shot, not to the shot in general. Six months later, the question 'did the delivered conform include the AI fill in shot 47' is either a lookup in a log or an archaeology project across render folders and departed freelancers.
A record keyed to shot IDs and versions solves this at the cost of one habit. It shows what was done to which shot, with which tool, by whom, and when, and it lets the delivery declaration be generated from evidence instead of memory. It should also outlive the project, because bounced shots, re-versioning requests, and repurposed masters all arrive after the wrap party.
Where documentation breaks down
- AI features sit inside compositing and editing packages as one tool among many, so their use leaves no separate trace unless someone writes it down.
- The heaviest fixing happens in the final days before delivery, under deadline, which is precisely when nobody logs anything.
- Shot versions multiply across conforms, and the AI work is attached to a version, not to the shot, so the delivered state is easy to misreport.
- Declaration forms are filled in by people who were not in the suite, and without a record they systematically over-declare or under-declare.
- When QC bounces a shot months later, the artist who fixed it has often left, taking the only knowledge of what was done with them.
A documentation routine that holds up
- Keep one AI-touch log keyed to shot ID and version: what was done, with which tool, by whom, and when.
- Fold it into the shot's finaling checklist, so no shot is marked final without a log entry, including an explicit 'no AI used' where that is the truth.
- Carry the log forward on every conform and mark which shot versions actually shipped.
- Before delivery, generate the declaration from the log rather than from memory, so the form and the record cannot disagree.
- Archive the log with the master files, because bounce-backs, re-licensing, and repurposing requests will reference it long after the project closes.
Frequently asked questions
Does an AI upscale count as manipulation we have to declare?
Declaration categories depend on the recipient's form and on the rules that apply in the delivery market, and they draw lines in different places. The safe position is a record precise enough to answer either way: which shots were touched, with what, and how heavily. The Article 50 guide covers what the EU rules actually ask.
We use AI features built into our editing and compositing software. Do those need records too?
Yes. A declaration form asks what happened to the picture and sound, not whether the tool was a plug-in, a menu item, or a website. If a generative feature changed the content, it belongs in the log regardless of where it lives.
What do we do about shots fixed at two in the morning before delivery, with no log entry?
Reconstruct the entries immediately, while the people involved still remember the session, and mark them as reconstructed rather than contemporaneous. A dated late record is far better than none, but the real fix is making the log part of finaling, so the two-in-the-morning shot cannot be marked done without it.
How long should we keep the records?
At least as long as the delivered master can come back to you, which in practice means years. Repurposing, re-licensing, complaints, and re-deliveries all arrive after the project has closed, and each one starts with a question about what was done to the material.
Who in post should own the log?
Whoever owns shot status. AI logging fails when it is a separate system living next to shot tracking, and it works when it is one more field on the pipeline the team already runs every day.