AI-Assisted Family Research: Tasks, Checks, and Boundaries

Give the Tool a Bounded Task

AI can organise messy notes, suggest search terms, compare transcriptions, or explain unfamiliar terms. Choose one small task and specify permitted material, required output, and how unknowns should be represented. “Discover my grandfather’s whole life” is insufficient as a sole instruction. Broad scope makes it hard to distinguish sourced information from a coherent story added by the tool.

All examples here are fictional. With three family letters, ask the tool to list addresses by letter identifier, without inferring reasons for moves, and to retain unclear characters. Treat the result as a draft worksheet and check every item against the images. It saves organising time without machine-certifying a life story. Researchers remain responsible for identity, meaning, and connections between evidence.

Understand That Fluent Text Can Be Wrong

Generative tools can produce confident, well-structured answers containing nonexistent records, wrong dates, or invented citations. NIST’s generative AI risk profile discusses this confabulation risk. In family research, the danger is not limited to absurd stories: one plausible village name, personal name, or marriage date can quietly enter a family archive and be copied repeatedly.

“The tool is very certain” does not describe evidence quality. Asking for supporting references helps checking, but the references themselves may be generated. Open the actual source and verify its title, repository, location, and whether its wording supports the statement. If inaccessible, record the output as an unverified lead. Even an accurate general explanation cannot establish that a particular ancestor experienced the same event.

Distinguish Types of Output

Distinguish extraction, explanation, inference, and suggestion in your request. Extraction should map to supplied wording; explanation interprets that wording; inference needs reasons and alternatives; suggestions concern next actions. These categories support checking without guaranteeing accuracy. Important names, dates, and relationships must return to sources regardless of their output category.

A fictional letter says, “Younger brother has returned home.” The tool can extract it, explain the relationship term, and suggest examining other letters for identity. It cannot identify the brother as Lin An merely because the pedigree includes him. Require unsupported names to remain unknown. If output still adds people or motives, remove unsupported content, record checking, and narrow the task instead of demanding more certainty.

Use It to Prepare Searches, Not Decide Identity

AI can suggest spelling variants, organise place-name forms to verify, or break a question into searches. Treat them as candidates and check suitability for the period and region. Romanisation can conflate several names, and historical place names may refer to multiple locations. Do not automatically attach an attractive English spelling to a Chinese person or merge people because the tool finds similar names.

For fictional Chen Lan, retain suggested romanisations and record which queries you actually tried and their results. Identity comparison still requires addresses, relatives, occupations, and timelines. If the tool claims an archive holds a particular register, check its official catalogue. Suggestions about where to search assist planning; verifying that a holding exists is a separate task.

Check Material and Permission Before Uploading

Before giving letters, interviews, or identity documents to an external tool, check for living people’s addresses, contact details, private experiences, or material others do not want public. Understand the service’s handling of submissions and respect agreed sharing limits. If uncertain, use redacted passages or fictional text to test the method. Possessing a scan does not automatically permit submitting it to every platform.

Do not upload the whole family folder merely for convenience. An address-organising task may need a few transcribed passages rather than everyone’s names in photographs. Retain identifiers so redacted working copies lead back to complete originals. Record submitted files, task, and date so later checking establishes what the tool actually saw. Minimising unnecessary input also clarifies scope and makes output easier to verify.

Save Checks and Decide What to Use

Save date, tool name, instructions, input identifiers, and important output in the research log. Record a visible version if available; do not invent one if not. For adopted content, explain human checks such as line-by-line image comparison, opening an official catalogue, or comparing another record. Cite actual historical sources on final person pages. AI drafts document work but cannot replace evidence for births, moves, or relationships.

You need not adopt every output. When detailed errors accumulate, manual organising may be more effective. Judge value by time saved on checkable work rather than professional-sounding prose. Family history centres on people who once lived. Tools should support careful reading while uncertainty, revisions, and judgment remain visible, allowing the next relative to continue research rather than inherit a fluent misunderstanding.

Test the tool on a short passage already checked manually and compare output with known results, looking for omitted negatives, confused people, or expanded dates. Success shows only that this task was usable, not that the next difficult item will be accurate. Use a fixed checklist for important output—item identifier, original passage location, names, and numbers—and verify each rather than relying on an overall impression that it looks correct.