To make a documentary with AI, start with a focused question and verified research. Build a script from real sources, collect footage and interviews, and use generated visuals only where their illustrative role is clear. Add narration, label reconstructions, check every factual claim and assemble the film in an editor.
- Build the account from attributable evidence, not plausible generated scenes.
- Separate real footage, archival material and AI illustration in the production record.
- Check names, dates, quotations, narration and visual labels before release.
The community-garden example below is a hypothetical planning exercise. It contains no reported interview, historical claim or completed AKOOL generation. Its purpose is to show how a source-led workflow can be organized without inventing missing evidence.
What can AI help with in a documentary?
AI can support planning and selected production tasks, but a realistic output does not establish a fact. Use it to organize questions, explore an illustrative composition or create a clearly labeled visual when that is appropriate. Keep responsibility for reporting, verification and editorial decisions with the filmmaker.
Separate three tasks that can look similar on a timeline. Evidence shows something about the real subject. Interpretation explains what the filmmaker thinks the evidence means. Illustration helps an audience understand an idea without claiming to record the event. Confusion between those roles is a larger problem than an imperfect visual effect.
For example, real footage of a present-day garden can establish its current appearance. An authorized photograph may show an earlier state, if its date and location are verified. A generated garden scene shows neither. It can illustrate an abstract idea, but it cannot replace an unavailable historical record.
How do you choose a focused documentary question?
Choose a question that can be answered with the reporting and permissions available. “How did this unused space become a shared garden?” is a manageable starting question. “How community gardens transformed an entire city” demands a broader evidence base and should not be inferred from one location.
Define the audience, geographic scope and time period before writing the narration. Decide whether the film explains a process, follows a person or investigates a disputed claim. Those choices determine which interviews and records you need.
For the hypothetical garden, the reporting plan might include present-day establishing footage, an interview with an authorized participant, dated archival images and observation of current activity. These are assignments to complete, not facts already gathered. Leave a clear gap if access to a source is not yet available.
Write down what would change your initial understanding. A useful documentary question allows the reporting to revise the story. If the conclusion is fixed before interviews and records are checked, the production can become an advertisement for an assumption.
What belongs in a documentary source log?
Keep a source log before drafting a polished script. Assign each proposed factual claim an identifier, then connect it to a source and the visual that will accompany it. This makes it possible to revisit a correction without searching every clip.
- Claim ID and wording: the exact statement you may include, with uncertainty retained.
- Source: a person, document, dataset or recording, plus the relevant page or timecode.
- Date and context: when the source was created and what period the statement concerns.
- Verification: what corroborates the claim and what remains unresolved.
- Permission and attribution: who controls the material, how you may use it and the required credit.
- Visual role: evidence, archive, reconstruction or general illustration.
- Editorial status: confirmed, qualified, disputed, awaiting permission or excluded.
A source-log template entry could read: “G-01 | [proposed opening date] | [dated record and page] | [corroborating source] | [credit] | [visual role] | awaiting verification.” Brackets are fields to complete. They must not become confident narration simply because the timeline needs an opening line.
An interview is evidence of what a person said, not automatic confirmation that every statement is correct. Check consequential dates, quantities and allegations against appropriate independent material. Preserve the interview recording and enough surrounding context to assess the quotation fairly.
How do you write a two-column documentary script?
Draft sound and picture together. In one column, write the intended narration, interview excerpt or natural sound. In the other, identify the exact visual and its source. Add a claim ID and verification status beside factual lines.
- Opening: audio, “[Verified description of the present location].” Picture, real establishing footage with recording date and location logged.
- History: audio, “[Confirmed account of how the project began].” Picture, verified archival photograph with source credit and date; otherwise mark the visual gap for reporting.
- Participant: audio, an authorized interview excerpt at a logged timecode. Picture, the actual interview or relevant real activity, preserving context.
- Explanation: audio, a clearly framed general observation. Picture, a labeled AI illustration only if it helps explain the idea without pretending to document the location.
- Closing: audio, a supported answer to the opening question, including unresolved limits. Picture, current real footage that does not imply a fabricated outcome.
This is a script template, not a finished account of a real garden. Do not invent a quotation to make the excerpt feel complete. A quiet moment of genuine location sound can be more informative than unsupported narration.
Read the script without the pictures. Does each factual sentence still have support? Then examine the pictures without narration. Could an audience reasonably infer something the sources do not show? A misleading edit can arise from the relationship between accurate elements.
When should you use real footage, archive or AI illustration?
- Real present-day footage: use it to show observed conditions or recorded activity. Log the time, place and context; avoid implying that a staged action was spontaneous.
- Verified archival material: use it for the period and subject it actually documents. Confirm provenance, date and permission. Do not treat a stock image as evidence of the specific event.
- Reconstruction: use only with an explicit editorial purpose and clear labeling. Do not fabricate a source's exact behavior or words as if recorded.
- AI illustration: use for a general concept or clearly identified hypothetical scene. Label it near the relevant shot and retain its generation record.
- Missing evidence: continue reporting, qualify the statement or omit it. A cinematic substitute does not resolve an evidentiary gap.
Avoid relying on an end-credit disclosure alone when a realistic generated shot could be mistaken for archive. The viewer needs the distinction at the moment it affects understanding. A plain label such as “AI-generated illustration; not footage of this location” is more informative than a vague production credit.
How can you plan an illustrative clip with AKOOL?
Choose a generic subject whose appearance will not be mistaken for a specific reported event. A seedling in a simple garden bed could illustrate growth between factual sections, provided its role is clearly disclosed. It should not be presented as a record of the documentary's actual garden.
An illustrative, untested prompt is: “Close-up of a generic young seedling moving gently in a breeze in a simple garden bed. Soft daylight, fixed camera, no people, no identifying location and no text.” This is a proposed shot, not an output claim.
For an image-based workflow, follow AKOOL's Image to Video help guide. Use an authorized reference, select an available model and describe a limited motion. Check the controls and credit estimate in your account. Save the input, prompt, model version, settings and output selected for review.
Inspect the clip for implausible growth, changing leaf shapes or an unintended identifying background. Keep the label attached when the clip is copied into the edit. If the illustration adds little information, omit it rather than using generated footage to fill every pause.
How do you handle narration and translations?
Finalize the factual wording before recording narration. A human narrator can read the approved script directly. If using a synthetic voice workflow, first confirm its current controls and permissions; this article does not report a tested narration generation or provide unobserved button-by-button instructions.
Check pronunciation of names, dates, organizations and place names against reliable references or the person concerned. Listen to the exported audio, not only the written script. A correct text input does not establish that the spoken output is correct.
Keep quotations distinct from paraphrases. Do not synthesize an interviewee appearing to say words they did not record without an explicit, carefully considered disclosure and permission. For translated versions, have a fluent reviewer compare meaning, names, numbers, captions and timing against the original.
A translated caption may be accurate while a generated voice mispronounces a key name. Review each representation separately. Maintain the source-language recording and a documented translation so later corrections remain possible.
What does the Ossa Film example actually demonstrate?
AKOOL's Ossa Film case study describes face replacement used to help protect participant identities in JOHATSU – INTO THIN AIR. It is a specific post-production application, not evidence that AKOOL researched, narrated or assembled the entire documentary.
Face replacement alone should not be treated as a guarantee of anonymity. Identity can also be suggested by voice, location, background objects, biography and other context. A production handling vulnerable participants needs a complete editorial and participant-protection review rather than relying on one visual alteration.
How do you assemble and verify the final film?
Use an editor suited to your source formats, such as DaVinci Resolve, and keep the source-to-shot map beside the timeline. This is an assembly plan, not a claim that this hypothetical example was completed in that editor. Import real footage, archive and generated illustrations into clearly named groups.
Build the factual sequence first. Place approved narration and interview extracts, then add the matching source material. Add labels and credits before polishing transitions so they are not forgotten at the end. Keep exact captions as editable text.
- Recheck every factual narration line against its logged source.
- Watch interview edits in context and confirm that the sequence preserves meaning.
- Match historical images to their verified dates and locations.
- Check that AI labels remain readable throughout the relevant shots.
- Review permissions, attribution, music and participant agreements.
- Listen for pronunciation errors and inspect caption spelling and timing.
- Have someone unfamiliar with the edit identify what they believe is real footage, archive and illustration. Investigate any confusion.
- Watch the complete export at normal speed and retain a versioned source record.
Begin with a short, well-supported sequence rather than an ambitious account with missing evidence. Use AKOOL's video generator when a clearly labeled illustration has a defined purpose. For broader shot planning, the AI directing guide supplies production context; the documentary's factual foundation still comes from reporting.
