The Most Useful AI Rule in Documentary Is a Piece of Paperwork

In April 2024 viewers of Netflix's true-crime documentary What Jennifer Did started freezing the frame on two photographs of its subject, presented as snapshots from her younger years. Futurism pointed to mangled hands, objects on a shelf that seemed to defy gravity and teeth that did not match from one picture to the next, and said the images had "all the hallmarks of an AI-generated photo." The film's executive producer, Jeremy Grimaldi, told the Toronto Star the photos were real: "The foreground is exactly her. The background has been anonymized to protect the source." I do not know which account is closer to the truth, and that is the point of this post. Nobody watching could know either. When the question came, the only thing the production had to offer was a sentence in a newspaper.

Five months later the Archival Producers Alliance — founded by the archival producers Rachel Antell, Stephanie Jenkins and Jennifer Petrucelli — published its best practices for generative AI in documentaries. Most of the attention has gone to the parts about telling audiences and simulating real people. The part I would put first is the least glamorous one. The guidelines ask productions to keep "cue sheets for all GenAI elements": the prompts, the software version, the date created, any reference material fed in and its copyright status, and "descriptions and timecodes for where GenAI materials appear in the production." It is the music cue sheet, applied to pictures. The same document draws a useful line — ordinary retouching, restoration and up-resing are out of scope; the concern is alterations to primary sources "that change their meaning in ways that could mislead the audience." A rebuilt background behind a real face, in a photo shown as an archival snapshot, sits somewhere near that line. A log is how a production would know which side, and how it would show anyone else.

Archive-driven films already live on paperwork. Every clip and still in a cut has to trace back to a source and a licence before the film can be delivered, which is why the log ends up being a second version of the edit. Generated material is the one element that arrives with no paper trail at all, and the APA names that problem precisely: such material carries "perceived authority, yet no accountability of authorship." A cue sheet puts the authorship back. It also has a practical payoff beyond ethics. The Producers Guild's magazine, writing up the APA's follow-on tool kit, notes that this kind of record helps when a production goes for errors-and-omissions insurance and for distribution. The people who will ask about your AI use first are probably not critics. They are the insurer and the buyer.

There is a newer example of a production trying to answer the question in advance. 1971: Billy Joel's Lost Interview, a 20-minute short directed by Adam Ripp, premiered on September 25 at the Long Island Music & Entertainment Hall of Fame's documentary festival. Its sound is an authentic quarter-inch reel-to-reel recording made on December 14, 1971; its pictures are AI-generated animation of Joel at 22. Ripp committed to putting the original master tapes on display at the festival. That is the right instinct — bring the primary source into the room — but notice what it covers. The tape proves the sound. The picture is the new part, and a tape cannot speak for it. What can is the record of what was generated, with which tools, from which reference images. The announcement does not say whether the production kept one, and it may well have. The point is that the reel on the table answers a different question from the one viewers are most likely to ask.

How a film discloses AI to its audience is a judgment call, and the APA says plainly that "different types of transparency may be required, depending on context." Keeping the record is not a judgment call. It costs almost nothing on the day a shot is generated and is close to impossible to rebuild a year later, after the prompts are gone, the software has updated and the assistant editor has moved on. Camera signatures and content credentials will increasingly vouch for new footage, but they cannot vouch for a 1971 tape, a family snapshot or a shot a model invented. For that material, the provenance is whatever the production wrote down. The day someone freezes your frame, you want to hand them a log, not an adjective.