AI-assisted research introduces practical questions about permission, privacy, bias, representation and disclosure. Those questions do not belong solely to a technology policy or a final legal review. They arise when a team chooses what to digitize, where to upload it, how to index a person, whether to reconstruct a scene and how to frame the result for an audience. A responsible process makes those choices visible, assigns someone to review them and keeps the evidence chain intact.
Permission is more than access
An item visible in an online catalogue is not necessarily cleared for reproduction. A repository’s access terms, a donor agreement, copyright ownership, privacy obligations and third-party rights may all apply. A documentary team should record the applicable terms for each intended use, including editing, excerpting, translation, online publication, broadcast, educational use and promotional materials. If the production wants to transform an image, isolate a voice or use a record as training input, confirm that the relevant permission covers that activity.
Rights can attach to more than the underlying record. A recently created scan, restoration, soundtrack, photograph of an artifact or published transcription may have separate restrictions. Keep the contact, correspondence, granted scope and expiry information with the item’s production record. Where rights are uncertain, pause and seek guidance instead of assuming that a citation or credit line substitutes for permission.
Protect sensitive people and records
Historical distance does not remove the possibility of harm. Records may expose survivors, relatives, informants, people accused of wrongdoing or communities whose knowledge is subject to specific protocols. A team should consider whether an item contains personal information, whether its original collection context included restrictions, whether republication could identify a living person, and whether a new AI search makes sensitive material easier to discover than before.
Minimize data. Do not upload confidential interviews, restricted files or identifying records to a third-party service unless the production has authority and has assessed processing, retention, access and reuse terms. Limit internal access to what the project needs; use secure storage; and document deletion or retention decisions. Face recognition, voice matching and identity inference deserve particular caution because a mistaken label may spread across a catalogue and become difficult to correct.
Examine bias in the archive and the tool
Collections reflect the choices of people and institutions that created, preserved and described them. They can overrepresent official voices, people with status, written records and events that institutions considered worth saving. Search systems may add another layer of unevenness by retrieving well-described and digitized records more readily than poorly catalogued material. A neat cluster of results is not a representative sample.
Ask whose language is indexed, which names or spellings are excluded, whether the training or retrieval method performs unevenly across accents or scripts, and who is missing from the source base. Record search limitations and test alternative terms. Do not ask an AI system to impersonate a missing historical voice as if it were testimony. If a production uses a contemporary performance or reconstruction, identify it as such and explain the basis for the interpretation.
Keep synthetic material distinct
Generated images, voices, text and reconstructed scenes can be mistaken for direct records, particularly when designed to look or sound plausible. Do not present generated material as an archival photograph, authentic recording, original quotation or unedited historical document. When synthetic or substantially altered material is used, document how it was made, what source material informed it, what was changed, who approved it and why it was necessary.
Disclosure should be clear, timely and comprehensible. Put it near the material or at the moment a viewer needs it to understand what they are seeing. A credit buried at the end cannot correct a misleading impression created throughout a sequence. Consider whether the project can tell the story with source material, diagrams, explicit contemporary illustration or narration about uncertainty instead. Technical possibility is not an editorial justification.
Give consent and consultation meaningful weight
Where interviews are involved, explain the project’s purpose, intended distribution, editing process and any AI-assisted processing in language participants can understand. Existing consent may not cover a new use. Give participants a real route to ask questions or raise concerns, and handle requests consistently with the production’s obligations and commitments. Consultation with affected communities should not be reduced to a last-minute endorsement request after the narrative is fixed.
Consultation cannot transfer editorial accountability away from the production. It can identify overlooked context, foreseeable harm and more respectful ways to represent people. Document who was consulted, what concerns were raised, what changed and why a suggestion was not adopted. Avoid presenting a single consultant as speaking for an entire community.
Make editorial responsibility operational
Assign clear roles for research verification, rights clearance, privacy review, AI tool approval and final disclosure. A risk register can note the material, possible harm, likelihood, mitigation, reviewer and decision date. Revisit it when a project changes from internal research to a public cut, adds new distribution or uses a tool for a different purpose. Keep a versioned record of generated outputs that materially affect the script or edit.
Set a useful threshold for escalation: disputed identity, sensitive personal information, uncertain rights, synthetic likeness, high-impact allegations or a source whose context is unclear should receive senior editorial review. A checklist cannot anticipate every history, but a clear route for questions prevents a production deadline from silently deciding an ethical issue.
Before release
- Check rights for the specific planned use and transformations.
- Review privacy, dignity, cultural context and potential re-identification.
- Label and document generated or substantially altered material.
- Confirm that public claims match the sources and convey uncertainty.
- Retain an accountable decision record and a correction pathway.
Explore source verification, visual sequencing, or contact the editorial team through Compose a Research Note.