When AI enters the edit bay: the Soderbergh question no one’s asking

Steven Soderbergh fed a rough cut of his new documentary to Meta’s MovieGen AI and let it suggest edits. The trades covered the workflow innovation. The tech blogs praised the collaboration. But I keep thinking about the woman in rural Kentucky whose face appears in that footage—did anyone ask her if she consented to having an algorithm analyze her expressions and suggest which moments of her life make better cinema?

This isn’t a Luddite panic about AI tools. I’ve watched editors use algorithmic color matching for years without ethical alarm bells. The difference here cuts deeper. When you point AI at documentary footage, you’re asking a black-box system to make aesthetic judgments about real people’s lives—people who never agreed to have their vulnerability processed through a statistical model trained on God knows what. Soderbergh’s experiment demonstrates a powerful editing tool. It also demonstrates how quickly we skip past consent when the technology gets exciting.

Documentary subjects operate in a strange legal twilight. They sign releases, yes, but those releases were drafted in an era when “editing” meant a human watching footage and making cuts. The implied contract was between subject and filmmaker—a relationship, however asymmetrical, between people. Now we’ve added a third party who never shakes hands, never builds rapport, never sits with the discomfort of choosing to include someone’s tears in the final cut. The algorithm doesn’t feel anything when it flags an “emotionally resonant moment.” That’s precisely the problem.

I’ve sat in enough post-production suites to know editors carry the weight of their subjects. They debate whether showing someone’s breakdown serves the story or just serves spectacle. They consider context the camera didn’t capture—what happened after they stopped rolling, what the subject said off-record, whether including this scene might damage a fragile recovery. These deliberations happen through human judgment informed by human conscience. When you delegate even part of that process to AI, you’re outsourcing ethics to efficiency.

The Soderbergh case becomes more complicated because his subjects include people who’ve died. The documentary covers his parents’ final years. Whatever consent conversations happened, they happened with people who can’t now respond to how their images are being processed. Film ethics have always struggled with posthumous representation, but at least human editors could hold themselves accountable to memory, to stated wishes, to the weight of depicting the dead. An AI system has no such tether. It processes deceased subjects with the same cold objectivity it applies to everything else.

This isn’t hypothetical hand-wringing. Productions are already using AI tools in assembly edits. The cost savings are real—rough cuts that might take a junior editor two weeks now take two days. But I’ve never seen a production budget line item for “additional subject consent for algorithmic processing.” I’ve never seen a release form that mentions machine learning. We’re rushing forward on the assumption that consent given for traditional documentary filmmaking transfers automatically to whatever new tools we invent. That assumption is legally shaky and ethically bankrupt.

The European productions I work with are starting to wake up to this. GDPR creates at least some friction—you can’t just feed footage of EU citizens into a US-based AI without thinking hard about data sovereignty. But most of the documentary world isn’t asking these questions yet. We’re still in the “wow, look what it can do” phase. I’ve been in that phase before, with other technologies that seemed purely beneficial until someone stopped to ask who was bearing the unexamined costs.

What makes this particularly urgent is that documentary subjects are often vulnerable by definition. We film people in crisis, in poverty, in grief—people who agree to participate because they trust the filmmaker, not because they trust the entire technological supply chain. When a person shares their story with a director they’ve spent months building relationship with, they’re not imagining that footage being parsed by the same class of tools that power TikTok recommendations. The betrayal isn’t intentional, but it’s real.

Some will argue I’m creating problems where none exist—that AI editing tools are just faster scissors, that subjects consent to “the film” regardless of process. But process shapes outcome. An algorithm trained on commercial cinema will flag different moments than a human editor working from documentary ethics. It doesn’t understand the difference between a useful revelation and an exploitative one. It can’t weigh whether a particular cut violates the spirit of what someone agreed to. It optimizes for engagement metrics we never discussed with subjects during the release conversation.

1. Documentary consent forms must explicitly address AI processing of footage and give subjects the option to exclude their material from algorithmic editing tools.

2. Productions using AI editing assistance need dedicated review by someone who knows the subjects and can advocate for their dignity—not just efficiency.

3. Posthumous subjects require even stricter protocols, with family consultation before any AI-assisted editing of deceased persons’ footage.

Soderbergh’s experiment will likely produce a good film. He’s a thoughtful filmmaker with decades of craft. But the trades won’t cover the consent conversations that didn’t happen—the questions no one thought to ask because the technology moved faster than our ethics. That gap is where documentary subjects always pay the price. We can close it, or we can keep optimizing our workflows and pretending process doesn’t matter. I know which choice lets me sleep better.

Todd Max Carey — Seven Journeys