Provenance research guide · how-to
How to use AI safely in provenance research
Use AI for bounded extraction, comparison, and query generation while keeping citations, uncertainty, and human decisions visible.
Give AI a bounded evidence packet
Provide identified sources and ask for extraction tied to page, field, or quotation. Do not ask a model to fill gaps from general knowledge or produce a seamless ownership history without source-level support.
Provide identified sources and ask for extraction tied to page, field, or quotation. Do not ask a model to fill gaps from general knowledge or produce a seamless ownership history without source-level support. Use a source ledger and claim identifiers so another researcher can reproduce the decision, challenge an assumption, and distinguish evidence from interpretation.
Require cite-or-refuse behavior
Every factual output should point to supplied evidence or state that the evidence is insufficient. Treat names, dates, translated titles, and object matches as candidates until a researcher verifies the source.
Every factual output should point to supplied evidence or state that the evidence is insufficient. Treat names, dates, translated titles, and object matches as candidates until a researcher verifies the source. Use a source ledger and claim identifiers so another researcher can reproduce the decision, challenge an assumption, and distinguish evidence from interpretation.
Keep decisions attributable to people
AI may organize leads and surface contradictions, but authentication, title, legal, rights, restitution, and publication decisions require qualified human judgment. Retain the prompt, model, output, corrections, and final reviewer.
AI may organize leads and surface contradictions, but authentication, title, legal, rights, restitution, and publication decisions require qualified human judgment. Retain the prompt, model, output, corrections, and final reviewer. Use a source ledger and claim identifiers so another researcher can reproduce the decision, challenge an assumption, and distinguish evidence from interpretation.
Record the boundary of the evidence
End the research record by separating documented facts, supported interpretations, unresolved questions, and disconfirmed leads. A gap is not proof of wrongdoing, and a repeated statement is not independent corroboration when every repetition traces back to the same source.
Retain the source URL or call number, access date, quoted passage, page or image reference, and a file hash when a stable local copy is permitted. This makes the conclusion reproducible and makes later correction possible without rewriting the history of the research.