AI in documentary post-production: where it helps, where it lies, and where it crosses a line

The conversation about AI in documentary has gotten stuck on the wrong question. The interesting question isn’t whether to use AI in post-production. The interesting question is what each specific use does to the documentary’s relationship with truth — and the answers differ dramatically depending on the tool.

I’ve been watching emerging-director friends adopt AI tools across the workflow over the past two years. Some uses are pure productivity gain with no ethical weight. Some are practical conveniences that introduce subtle distortions worth knowing about. And some — including ones that have already crept into festival films — cross a line documentary has historically taken seriously and should keep taking seriously.

Where it helps without complication

Transcription is the clearest win. Good AI transcription has cut the time and cost of paper edits substantially, and the output is searchable in ways that human transcription rarely matched. The remaining errors are real but discoverable. An editor with access to AI-transcribed, time-coded material can find moments in days of footage that would have taken weeks to surface manually.

Translation drafts are similarly useful, with one important caveat. The first-pass translation gets you 80% of the way there. The remaining 20% is where the meaning lives, especially in interviews where cultural specificity, idiom, and emotional register matter. AI drafts speed the work; they don’t replace the human translator who actually understands what was said. Productions that skip the second step ship subtitles that look correct and convey something subtly different from what the subject meant.

Logging and metadata work is another clean win. AI can identify faces, locations, objects, and rough subject matter across thousands of hours of footage. This is the kind of cataloging work that used to require an assistant editor’s full attention for weeks. Now it can be done in days, with human verification. The ethical weight is near zero. The productivity gain is large.

Where it introduces distortions worth knowing

Audio repair tools are powerful and getting more powerful. Background noise removal, room-tone reconstruction, dialogue isolation — these tools can rescue interviews that would have been unusable. They can also, if pushed too far, change what was actually said. A subject mumbling becomes clear. Hesitation gets cleaned up. The audio version becomes a smoother, more confident version of the moment that occurred.

Most viewers won’t notice. The subject might. And there’s a question worth asking, project by project: is the audio repair restoring intelligibility, or improving the subject beyond what they actually delivered? The line matters. Documentary subjects are not stage actors. Their hesitations, pauses, and verbal struggles are part of what they said.

Color grading driven by AI tools has its own quiet effects. Skin tone normalization, especially across footage shot in different conditions, can make a sequence visually coherent. It can also subtly shift how subjects appear in ways that change perception. Documentary subjects whose skin tones get adjusted across a film are being depicted differently than they were filmed. This is rarely deliberate manipulation. It’s a side effect of tools that prioritize visual consistency.

Generative restoration of damaged or missing material is the trickiest category. If a section of archival film is decayed and an AI tool fills in plausible-looking content based on surrounding frames, what’s on screen is not a recovered image. It’s a generated approximation. Audiences usually don’t know. Producers usually don’t disclose. The ethical weight depends on what’s being filled in, but the practice deserves more transparency than it currently gets.

Where it crosses a line

Voice cloning is the clearest case. The capacity exists to take a few minutes of a subject’s recorded speech and generate new speech in their voice saying things they did not say. There are legitimate edge cases — a subject who participated in production has died, and the family has authorized using their voice to read a written passage they wrote in life. Those cases require extraordinary care and explicit disclosure.

What’s happening more often, and more quietly, is voice cloning being used to smooth audio: replacing a botched word, completing a sentence the subject started but couldn’t finish, providing voiceover narration in a deceased subject’s voice without clear authorization. These uses produce a documentary in which the words the audience hears were not, in fact, said by the person they’re attributed to. That’s a category change. It’s no longer documentation.

Synthetic re-enactment of historical figures using their actual voices and likenesses is another line crossed. A documentary can speculate about what someone said. It can dramatize. It can use actors and label them as actors. What it cannot do, in a form that takes its relationship with truth seriously, is generate a synthetic version of a real person saying things they did not say and present it as if it might have happened.

Some current festival-circuit documentaries do exactly this. The defense is usually that the synthetic is based on documented evidence or represents what we know the person believed. This is the wrong frame. Documentary’s claim on the audience is that what’s on screen actually happened. Synthetic re-enactment of real people erodes that claim, regardless of how researched the underlying speculation is.

The disclosure standard

The simplest rule I can offer: if a tool changed the material in a way an attentive viewer would care about if they knew, disclose it. Transcription tools don’t require disclosure because they don’t change the material. Voice cloning of a deceased subject requires explicit disclosure because viewers would care.

This standard is conservative in some ways and liberal in others. It lets productions use a lot of AI without ethical anxiety. It also rules out a category of use that some current films treat as normal and that, in my view, the form should not.

Where this leaves us

The directors who use AI thoughtfully in post-production are working faster and finishing films they couldn’t have finished before. That’s a real gain. The directors who use it without distinguishing between the categories are slowly shifting what documentary means as a form. That’s a real loss, and it’s happening without much public debate inside the industry.

Key takeaways

1. AI in documentary post-production is not one question — transcription, translation, audio repair, color, voice cloning, and synthetic re-enactment have radically different ethical weights and should be evaluated separately.
2. The line worth holding is whether the tool changes what was actually said, shown, or done in ways the subject wouldn’t recognize; restoration is fine, fabrication is not.
3. Disclose any use that an attentive viewer would care about if they knew; the conservative standard preserves documentary’s claim on truth without prohibiting genuinely useful tools.

The form has metabolized new technology before. It will metabolize this one too. Whether it does so with intention or by drift is up to the practitioners deciding, in each edit room, what they’re willing to ship.

Todd Max Carey — Seven Journeys