How can AI support provenance research responsibly?

Why this question is difficult

Generative systems can produce fluent historical narratives before the evidence is stable. They can merge similarly named actors or artworks, invent citations, conceal contradictory records, repeat source errors, and present probability as fact. Speed without traceability increases research risk.

The Research Commons treats this as a method question rather than a shortcut to a historical conclusion. Work remains source-first, identity-aware, revision-bound, and open to correction. Every consequential result requires an attributable human reviewer, and publication remains a separate approval decision.

A reproducible method

Use deterministic processing before language-model calls: validate inputs, normalize dates without discarding source wording, preserve multi-valued fields, and detect exact identifiers. Give agents bounded evidence packets and explicit tasks. Require citations or refusal, expose uncertainty, limit passes and cost, and record the model, prompt revision, output, corrections, and accountable reviewer.

Create one machine-readable packet for the exact revision being reviewed. Preserve source URLs, visible queries, access dates, quotations or field locations, object-identity anchors, contradictions, rejected candidates, alternative hypotheses, gaps, and limitations. Hash the canonical packet so later corrections cannot silently alter the evidence reviewed.

Evidence and responsible automation

Appropriate agent tasks include visible query planning, candidate retrieval after a human trigger, evidence extraction, citation resolution, identity-collision warnings, contradiction detection, missing-evidence lists, and reviewer-packet assembly. Evaluate precision, recall, citation accuracy, refusal accuracy, latency, and cost on a frozen non-synthetic set.

AI agents may assist with bounded query planning, retrieval after a human trigger, evidence extraction, citation checking, collision warnings, contradiction detection, and packet assembly. They may not establish identity, authenticity, title, wrongdoing, novelty, publication readiness, or legal conclusions. Missing evidence should produce a refusal or next research action, not invented completion.

Challenge the result before using it

AI must not authenticate an artwork, establish title or legal rights, resolve consequential identities, assert wrongdoing, declare scholarly novelty, contact people, publish, or authorize payment. Human review is not a ceremonial final click: the reviewer must reproduce sources, test alternatives, record corrections, and own the decision.

A qualified reviewer should reproduce the most material citation, challenge identity, test the alternative hypothesis, identify source dependencies, and record corrections. The reviewer’s decision must name what is supported, disputed, unresolved, or outside scope. A polished narrative, repeated database statement, or high similarity score cannot substitute for this work.

Primary method references

• Linked Art data model — Event-centered cultural-heritage data semantics and multi-valued evidence structures.
• NIST AI Risk Management Framework — Risk, measurement, transparency, and accountable human-governance guidance for AI-assisted work.
• CIDOC Conceptual Reference Model — Reference model for events, actors, objects, places, timespans, and evidence-bearing relationships.

Questions and limits