Documentary research often begins with an appealing sentence: a date, a quotation, a name, a dramatic event. AI tools can produce such sentences quickly, and they can also make a weak connection sound settled. A responsible workflow starts by asking a less exciting question: what specific record supports this wording, and can another person inspect that record? The answer should lead beyond a model’s summary, beyond a search snippet and beyond a citation copied from an unknown source.
Turn a sentence into a checkable claim
Break a broad line of narration into propositions. “The crowd gathered at dawn and demanded the mayor’s resignation” contains a time, a group, an action and a stated motive. Each may need different evidence. A photograph could support that people were present, but may not establish the hour or what they demanded. A newspaper may report a demand, but its author, publication date and political perspective matter. A later memoir could add detail while also reflecting memory and hindsight.
Write down the exact claim, its intended wording, the evidence currently available and the confidence or uncertainty. Do not let a generated summary merge several separate sources into an apparently unified witness. When the tool supplies a reference, inspect it independently. Fabricated citations and mismatched page references are well-known failure modes of language models; a plausible book title is not proof that the quotation appears in that book.
Return to the record, not just its description
Record a stable identifier, repository, collection, item title, creator, date, page or frame, and the access path. For an oral-history recording, note the interview identifier and timecode; for a moving-image reel, include the relevant sequence and any supplied shot list. Save a reference copy or link when permitted, but respect repository terms. A catalogue entry is useful context, not a replacement for the document, photograph, recording or film itself.
Read around the passage. Check the lines before and after a quotation, the page heading, the document’s intended audience and any editorial annotations. For a photograph, examine the full frame, caption history and surrounding collection. For a film clip, determine whether the footage has been edited, dated later or reused with a new caption. The meaning of an extract can shift when it is removed from its original sequence.
Separate evidence from inference
Evidence is what the record contains or shows, subject to its condition and context. Inference is the reasoning that connects records to a conclusion. A researcher may infer that two people met because separate records place them in the same room at the same time; that remains an inference unless a direct account establishes the meeting. A model can help suggest connections worth testing, but its suggestion is not an additional witness.
Build a claim ledger with distinct fields: claim wording; supporting source identifiers; exact extracts or timecodes; contradictory material; contextual scholarship; inference; confidence; reviewer; and final editorial decision. This makes disagreement manageable. If a source is ambiguous, write down the competing readings. If the evidence does not resolve a question, let the narration preserve that uncertainty instead of choosing the most cinematic option.
Use AI as a research assistant, with boundaries
AI can support transcription search, name variation discovery, translation drafts, topic clustering and comparisons across a large set of records. Ask the system to identify candidate passages, not to certify their meaning. Keep the query, output and model or tool information when it materially affected the research. Compare machine transcription against the audio or image, especially for names, numbers, politically charged language and quotations destined for narration.
When summarizing multiple documents, ask for item-level references and verify each one. A tool that cannot show its evidence trail may still help with brainstorming, but its output should not enter a script as fact. Protect personal, confidential or restricted records from unauthorized uploads, and do not assume that a vendor’s privacy or retention terms are suitable for sensitive research.
Check for disagreement and silence
Verification includes looking for what might undermine a preferred account. Search variant spellings, alternative terminology and sources created by people with different positions. Consider who made a record, who was absent from it, what its purpose was and which voices were excluded from the archive. Silence is not proof that an event did not happen, but neither is it permission to invent a missing account. A careful film can explain the limits of the surviving record and the method used to address them.
Prepare a reviewable handoff
Give the editor or fact-checker the exact wording, source path, quotation context, rights information and any uncertainty that affects the scene. Flag claims whose evidence is indirect, disputed or dependent on a generated search result. Keep a versioned record of changes so that a cut can be audited after narration is shortened. This practice is especially useful when a line has travelled through research notes, script drafts, subtitles, publicity and educational materials.
Useful reference points include the Oral History Association’s Principles and Best Practices, the Society of American Archivists’ ethical guidance, the International Council on Archives’ ISAD(G), and the catalogues of the British Library, Library of Congress and The National Archives. They offer different kinds of support; none replaces direct source reading or editorial judgment.
A practical claim test
- Can a reviewer locate the exact item and passage?
- Does the source support the wording actually used?
- Have context, counter-evidence and provenance been checked?
- Are inference and machine-generated suggestions labelled?
- Does the film communicate uncertainty honestly?
Continue to Assembling the Visual Argument or explore the ethical edit.