Documentary already has a process for handling materials it didn't shoot and can't fully trust. 

The public argument over generative AI in filmmaking runs between two extremes: this technology will hollow out the craft, or resistance is nostalgia and the future belongs to whoever prompts fastest. Productions live in the space between, one approval at a time.

I produce and shoot nonfiction. I've also run production departments, which means years of turning creative intentions into schedules, budgets, and decisions someone can defend later.

From that seat the useful questions are narrow: what task is the tool doing, who approved it, and would the audience feel misled if they knew?

The legal ground is moving

At the Produced By conference in June, a panel on AI's legal terms and conditions included a slide I keep thinking about. As AI gets integrated into ordinary tools, the question that matters is moving from "was AI used?" to "how much human creativity was involved?"

"Was AI used" stopped being a useful question the moment these features landed inside edit systems; on a 2026 production the answer is always yes. What a producer has to demonstrate is substantial human authorship over the work.

The same panel walked through a catch-22 most coverage misses: use a raw AI output untouched and you may keep the vendor's indemnity, but you likely lose copyright, because there's no human authorship to protect. Shape the output and you may gain copyright while the indemnity falls away.

Neither path is without risk. What you can do is know which risk you're carrying, on which elements, in writing, before delivery. That's chain of title work, and it belongs in prep.

Nonfiction carries a different contract

In scripted work the audience expects illusion. Nobody feels betrayed learning the set was a volume stage.

A documentary is a claim: this happened. The Archival Producers Alliance's best practices quote G. Roy Levin's definition of the genre, an "inherent obligation to reality." Everything we make trades on the audience's willingness to believe what we show them.

So the disclosure question in nonfiction runs past what's legally required. The test I'd put first is the one the Produced By panel listed last: would the public feel misled by this use? If the answer is yes, the tool is wrong for that task.

Where the tools help

I'm optimistic about these tools in documentary hands. But optimism about "AI" in the abstract is worthless to a production; so is dread. The evaluation has to come task by task, because the same technology that saves a finishing budget on one project can poison the whole film on another. Here's a breakdown of the key areas where generative AI is showing up, and how to handle each.

Restoration and cleanup of degraded archival is a clear win. The APA's guidelines deliberately set aside retouching, restoration, and up-rezing as minor alterations; their concern is generation of new material, and alterations that change meaning such as from photo to video. This kind of use will require the same discernment and sensitivity to factual basis we've always practiced with "re-enactments," now heightened to include consideration of the format as well as the content.

Post tools for relighting, mix and color assistance, and object removal extend what a lean finishing budget can hold. Some of that used to require a facility, now it may be only a talented artist with AI-assisted software. AI transcription, logging, and translation already changed doc post, and nobody mourned, though we did learn those outputs need human QC and cleanup; quality varies between vendors and inputs. The lesson from these cases is that it takes skilled humans to drive and review and iterate; the tools just make it faster and cheaper.

For independent producers the highest-return use might be development: creating visuals for a pitch before you can afford to shoot it. But be careful to label and track these elements if they end up in the same edit as the project itself.

Archival gaps are the most tempting case, but there is a dangerous trap: the fix can reinforce the bias that created the gap. Historically, who got recorded tracked money and race, and who got preserved tracked the same, and most physical archives have never been digitized. The models are trained on that skewed record, so a generated image of the past may inherit the biases that caused the lack of original archival material.

The APA's case study on Zac Manuel's The Instrument makes this concrete. Manuel set out to recreate the singing voice of his grandfather, a Black tenor in New Orleans whom nobody ever recorded. The voice companies couldn't rebuild it from six minutes of deposition audio, and an early sample came back with a British accent instead of his grandfather's Southern Creole. Manuel's answer was to put the attempt in the film itself: the process, its limits, and its biases play out on screen, where the audience can see exactly what the machine is doing. He made the disclosure part of the story itself.

My recommendations: no synthetic material presented as documentary record. No synthetic subjects. No voice or likeness without documented consent and disclosure. That's where I draw a line, whatever the tools learn to do next.

Treat it like archival

The APA was founded on these challenges of use and disclosure. Archival producers came together in 2023, worried about "fake archival" creeping into films, and wrote the field's first real standards: cue sheets for every GenAI element, modeled on the deliverables we already produce for music and archival. Cue sheets can track prompts, software versions, reference materials and their copyright status, and of course, timecode stamps. The APA toolkit includes tracker templates, crediting suggestions, and workflow tips by production phase. I intend to run with all this and help make it standard.

What I'd add from the production-management side: the system only works if it's staffed and scheduled.

The APA suggests one person own the tracker, and I agree. On most shows that's the archival producer, an assistant editor, or a post supervisor brought on early. But owning the log and owning the decisions are different jobs. A generative choice can surface anywhere on a production: an editor reaching for generative extend to save a shot, the art department mocking up a period image, a director wanting a scene visualized that nobody filmed. Each of those calls already has a natural home on the org chart. Creative weight sits with the director, the editor, the art department. Feasibility, cost, and risk sit with the producer, line producer, and post supervisor. The archival producer bridges the two, because provenance is already their whole job.

Existing roles can manage all of it. The work, in prep, is defining which role holds which level of responsibility, and who is ultimately accountable for AI use and its disclosure.

And legal review needs a real budget and a slot on the calendar. That way, approvals and disclosure get settled well before picture lock, and finishing stays on schedule.

Ethical choices that never make it onto the calendar stay theoretical, and the default may become non-disclosure by neglect.

We've written playbooks under uncertainty before

Producers went through a forced shift with no playbook recently. When COVID hit, the company where I ran production stopped shooting, as everyone else did, due to the stay-at-home orders. We were mid-production on two Netflix documentary series, and with their support, we soon resumed working, first in post: building workflows for remote editorial, then remote color, without dropping finishing standards.

Getting back into physical production was harder. Our first project to shoot during COVID was Two Distant Strangers, built around close-contact fighting and intimate scenes, activity SAG-AFTRA generally prohibited under the early return-to-work rules. We got permission by developing a safety plan built on 6-hour turnaround PCR tests: talent tested the night before those scenes and cleared before heading to set in the morning. The plan worked, no one got sick, and the film went on to win the Academy Award for Best Live Action Short.

Through the first year of COVID the guidance changed frequently, and we adjusted our plans to reflect every iteration.

That's the skill using AI in production will demand: living policy, revised as the world evolves, specific enough that a crew can follow it and a studio partner can sign off on it.

The norms are being written now

Standards for generative work in nonfiction are being set right now, by people who care to shape them: the APA working groups, the guild panels, the productions willing to document their process so the next crew doesn't start from zero. Opinions, for and against, won't settle it. Producers comparing notes on specifics will.

The most useful thing a working producer can do this year is boring: track your uses and write down your reasoning. While we can't broadly share our cue sheets due to NDAs, the template travels fine, and so does a summary of the big decisions once a show is released.

Done right, this technology puts visual storytelling within reach of projects that could never afford finishing or didn't have natural archives. That's a future with more diverse documentary stories and an audience that shows up because of trust in the storytelling. Worth the paperwork.

Notes and sources

A note on process, because this piece argues for disclosure: I developed this essay in conversation with Claude, Anthropic's AI assistant, which helped with research, structure, and drafting. The positions and the experiences are mine, and so is the final language, after several rounds of my own rewrites. 

The legal points draw on "Artificial Intelligence: The Terms & Conditions," a panel at the Producers Guild of America's Produced By Conference, June 2026, with Lori McCreary (CEO, Revelations Entertainment; co-founder, FoodFight USA) and Will Schildknecht (partner, Latham & Watkins). Nothing here is legal advice; talk to counsel who works in this area.

The Archival Producers Alliance's Best Practices for Use of Generative AI in Documentaries (Sept 2024), GenAI Tool Kit (April 2025), and case studies, including "The Instrument." Guidelines copyrighted by the Archival Producers Alliance, and used by permission.

Featured image: a mock cue sheet designed by me with Claude, based on the APA AI Tracker & Cue Sheet template.