The AI cue sheet: an old idea for a new problem

The film industry already solved a version of this problem decades ago, and it did it for music. Every scored production keeps a cue sheet (opens in a new tab), a running log of every piece of music in the finished work: the title, the composer and publisher, how the music was used and for how long. It exists so that rights and royalties can be sorted afterwards without anyone having to reconstruct the soundtrack from memory, and it has been a quiet, load-bearing habit of the business for as long as television has paid composers. The AI cue sheet is the same idea pointed at a newer problem, a running record of where AI was used on a production and what its use produced, and the reason the analogy is worth taking seriously is that the old solution already proves the method works.

Why the music cue sheet works

It is worth being clear about why the music version has lasted, because the same properties are what make it worth copying. A cue sheet works because it is kept as the work happens rather than assembled at the end, because it is structured into consistent fields rather than written as prose, and because it travels with the production as a normal deliverable rather than being chased down later. None of those properties is about music specifically; they are about how you keep a reliable record of many small events spread across many people, which is exactly the shape of the AI problem. The music industry did not build the cue sheet because music is special. It built it because memory is unreliable and rights questions arrive late, and both of those are true of AI use too.

Why AI use needs the same discipline

AI use on a production is scattered across departments and vendors, mostly mundane, and almost never written down in one place, and three separate pressures now converge on that blank space. The copyright question turns on what humans actually did with which tools; the Article 50 disclosure duty turns on knowing which content was AI-generated or manipulated; and broadcasters, funders and insurers are beginning to ask for an account before they take on the work. What is striking is that all three are versions of the same underlying question, which AI was used and where did its output go, and a single record answers all of them at once, in the same way one music cue sheet serves every performing-rights society rather than a different document for each.

What an AI cue sheet records

The useful record is granular without being heavy, and it works best kept one use at a time rather than as a vague project note. For each meaningful use it is worth having the tool and what it was used for, who used it and at what stage, and the two facts that do the most work downstream: whether the output actually entered the final production, and whether it depicted a real, identifiable person. A light risk marker helps too, so that a spell-check and a synthetic voice are not documented to the same depth, which keeps the effort proportionate and the signal clean. That is close to the whole of it, and the discipline is in doing it at the time rather than the depth of any single line.

Where the analogy holds, and where it stops

Honesty about the comparison is part of what makes it useful, because an analogy oversold is worse than none. The cue sheet method transfers cleanly in its logic, capture as you go, keep it structured, let it travel, but the surrounding machinery does not yet exist for AI the way it does for music. There is no performing-rights society to submit an AI cue sheet to, no settled schema everyone agrees on, and no clean, industry-wide definition of what counts as a reportable AI use, which is harder than it sounds because the line between AI-assisted and AI-generated and the boundary of what counts as a deepfake are themselves unsettled. So the AI cue sheet borrows the habit, not the whole infrastructure, and it is better understood as a discipline a production adopts than as a filing it submits to an external body. That is a real limit, and naming it is more useful than pretending the parallel is exact.

What the record is, and what it is not

It is worth saying plainly what such a record proves, because claiming too much for it would undo its value. An AI cue sheet does not prove what happened the way a forensic system would; it preserves an accurate, dated account of what the people on the production reported about their own AI use. That is a more modest thing than proof, and it is also the thing that a broadcaster, an insurer or a court is actually looking for, because a record that is plainly a contemporaneous account of what was reported is more credible than one that claims to be evidence it cannot be. Claiming less is what lets it do more.

Keep it living, and keep it early

The last lesson from music is about timing. A cue sheet is worthless if it is written after the fact, and the same is true here, which is why the strongest version of the habit begins before production, when a producer sets out which tools are approved and what will need reporting, and continues through the shoot and the edit as people log their use against those categories. Gathering the record from the people doing the work, at the point they do it, is what turns it from a reconstruction into a record, and a structured way to collect that reporting is the method this product is built around, though the principle holds whether you use a tool or a disciplined shared document. For the policy that sets the categories in the first place, our post on writing an AI policy is the place to go next.