# How to Fact-Check an AI Documentary Before Publishing

> A claim-by-claim fact-checking workflow for AI-assisted documentaries covering sources, quotations, dates, causality, visuals, captions, and correction records.

Updated: 2026-07-11
Audience: Documentary creators, editors, researchers, and channel operators
Canonical: https://onira.studio/guides/how-to-fact-check-ai-documentaries

## Key takeaways

- Extract checkable claims instead of reviewing the script by impression.
- Verify visual implications separately from narration.
- Preserve source, reviewer, decision, and correction records.

## Create a claim inventory

Read the final narration and list every checkable statement: names, dates, quantities, quotations, firsts, causes, consequences, locations, titles, and claims about what a person believed or intended. Break compound sentences into separate claims so one supported clause does not hide another unsupported one.

Classify each entry as documented fact, attributed interpretation, creator inference, or visual reconstruction. The classification determines how the line should be sourced and worded.

- Atomic claim text and timestamp.
- Claim type and required evidence level.
- Source, reviewer, status, and correction note.

## Trace claims to the strongest available source

Prefer authoritative primary material for direct records and reliable secondary sources for context and interpretation. Check the actual source rather than relying on a search snippet, generated summary, or citation copied from another article.

Confirm that the source supports the precise wording. A page mentioning two events does not necessarily establish causality between them. A quotation needs the original speaker, wording, context, and translation checked when possible.

- Open and read the supporting passage.
- Match certainty in the script to certainty in the evidence.
- Use attribution where interpretation or dispute matters.

## Audit what the visuals assert

A scene can make a factual claim without words. Clothing may place a person in the wrong century. Architecture may imply the wrong region. A generated newspaper or inscription may appear to document a statement that never existed.

Review each realistic scene against period, geography, identity, and event constraints. Replace exact evidentiary visuals with verified assets or clearly illustrative treatment when the generation cannot be trusted to carry the fact.

- People, likeness, clothing, objects, and architecture.
- Geography, weather, scale, and chronology.
- Text, maps, documents, insignia, and symbols.

## Close the review and keep a correction trail

No unresolved high-risk claim should reach publication. Rewrite overconfident statements, remove unsupported detail, replace misleading scenes, and recheck captions after every narration edit. Record the final source set and the person responsible for sign-off.

After publication, provide a correction path and update the description or film when a material error is found. A transparent correction is a stronger trust signal than pretending the production process is infallible.

- Zero unresolved high-risk claims.
- Captions and chapter text match the corrected narration.
- Source pack, sign-off, and correction log are retained.

## Publication checklist

- Every checkable statement appears in a claim inventory.
- Each material claim has an opened and reviewed source.
- Certainty and attribution match the evidence.
- Quotations, translations, names, and numbers are checked.
- Realistic visuals are reviewed for implied factual claims.
- The final source pack and correction log are preserved.

## Sources

- [US National Archives primary-source guidance](https://www.archives.gov/education/research/primary-sources)
- [Library of Congress primary-source guide](https://www.loc.gov/programs/teachers/getting-started-with-primary-sources/)
- [Google guidance on helpful, reliable content](https://developers.google.com/search/docs/fundamentals/creating-helpful-content?hl=en)

## Questions

### Can an AI fact-check its own documentary script?

AI can help extract claims and locate possible sources, but a reviewer should open the sources, assess their authority, and verify that they support the exact wording and visual implication.

### What claims need the most scrutiny?

Quotations, accusations, medical or financial claims, disputed history, precise statistics, causal claims, living people, and realistic scenes presented near factual narration deserve heightened review.

### Should sources appear in the YouTube description?

Publishing useful source notes can strengthen transparency, but the private production record should be more detailed and map important claims to their supporting material.

## Product boundary

- Onira delivers a final MP4; it does not upload or schedule posts on YouTube or social platforms.
- Onira provides a reviewable production workflow; creators remain responsible for approving the story, facts, rights, disclosure, and final publication.
- Director chat is limited to regenerating one selected PREVIEW timeline video clip; other available Studio controls are separate direct actions.
- Creators must review facts, sources, rights, realistic-synthetic-media disclosure, and platform policy before publishing.
- Onira does not guarantee YouTube monetization, reach, factual accuracy, or legal clearance.