Defensible AI in the Archive: Confidence Flags, the Review Gate, and Provenance That Holds Up

An archive is a place of evidence. The records it holds, and the descriptions attached to them, may one day support a legal claim, settle a question of provenance, answer a Freedom of Information request, or underpin a funder’s final report. That evidential weight is the whole reason archives are trusted. So when AI enters the cataloguing workflow, the test it has to pass is not “is it fast” or “is it clever”. The test is: will the records it helps produce stand up to scrutiny?

This is the question that should govern any use of AI in description, and it is the one that separates a tool built for archives from a tool that merely processes text. A description that cannot be defended — whose origin is unknown, whose accuracy was never checked, whose reasoning is opaque — is not an asset. In an archive, it is a liability. Defensible AI means being able to say, of every field, who or what produced it, how confident the system was, whether a human reviewed it, and when. Here is how that is built.

The review gate: AI proposes, the archivist disposes

The foundational principle is simple and non-negotiable: AI proposes, the archivist disposes. No AI-generated field becomes part of the record until a human has seen it and approved it. The model produces a draft; the archivist makes it a record.

This is more than a safety feature. It is what keeps the archivist as the author of the description, with the AI as an assistant rather than a decision-maker. Large language models can produce plausible but wrong output — a date that reads correctly but is not, a name attached to the wrong person, a subject that misreads the content. Without a review step, those errors enter the catalogue silently and propagate. With a review step, they are caught where they should be: at the point of human judgement, before anything is finalised.

Archivers.ai is built around this gate. As the processing and access layer in front of your existing systems — AtoM, ArchivesSpace, CALM, Axiell, Preservica — it never writes a finalised record on its own. It prepares proposals; the archivist confirms, corrects, or rejects them. The way we use AI is organised entirely around keeping the human in control at every step, not around removing them to save time.

Confidence flags make review targeted

A review gate is only workable if review is efficient. If an archivist has to re-examine every field with equal suspicion, the gate becomes a bottleneck and the time saving evaporates. This is what per-field confidence flags solve.

Every field Archivers.ai proposes carries its own confidence indicator. A clearly printed date the system read with near-certainty is flagged differently from a faded handwritten name it found hard to parse. This lets the archivist direct attention where it is actually needed: scrutinise the low-confidence fields closely, verify the high-confidence ones quickly, and move on. Review becomes triage rather than a uniform re-keying of everything.

The honesty of this matters as much as the efficiency. A confidence flag is the system declaring its own uncertainty rather than presenting every answer with the same false assurance. An archivist can trust a tool that admits what it is unsure about far more than one that asserts everything with equal confidence — because the second tool hides exactly the errors you most need to catch. This is especially important for the difficult material archives increasingly process: handwriting, photographs, and oral histories, where uncertainty is genuine and pretending otherwise would be dangerous.

Per-field provenance: who, what, when

The third element is provenance, recorded at the level of the individual field. For every AI-assisted field, Archivers.ai records the provenance of the work: the model that produced the suggestion, the timestamp, and the review state — whether it was accepted as proposed, edited by a human, or rejected. This is provenance for AI-assisted fields including model, timestamp, and review state, captured as the record is built rather than reconstructed afterwards.

This is what makes the description defensible rather than merely present. A description with provenance can answer questions a bare description cannot. Where did this date come from? An AI proposal, reviewed and confirmed by a named cataloguer on a given day. Was this subject heading checked by a person? Yes — here is the review state and the timestamp. Did a human ever see this field? The record itself says so.

Crucially, this provenance is not a separate logbook that drifts out of sync with the catalogue. It travels with the record and exports with it in standards-aligned formats — alongside EAD3, Dublin Core, and PREMIS preservation metadata in BagIt packages — so the audit trail is part of the archival output, not an afterthought bolted on for a one-off inspection.

Standing up under FOI, donor scrutiny, and funder reporting

The reason all of this matters is that archival descriptions are tested in the real world, and the three most common tests each demand exactly what defensible AI provides.

Freedom of Information. A public archive subject to FOI must be able to account for how it produced and handled records. If an AI-assisted description is challenged, “the computer generated it” is not an acceptable answer; “an AI proposed it, a named archivist reviewed and confirmed it on this date, and here is the record of that” is. Provenance turns a vulnerability into a documented process.

Donor scrutiny. Donors care how their material is described, and rightly so — a misdescription can misrepresent a family, a community, or a life. Being able to show a donor that descriptions were not produced unsupervised by a machine, but proposed and then reviewed by a professional, is the difference between confidence and complaint. The review state on each field is the evidence.

Funder reporting. Funders increasingly ask how AI was used in projects they paid for, and expect a human-in-the-loop process with documented verification. A complete, timestamped, per-field record of AI proposals and human review is precisely the governance evidence a final report needs. We have written more on what heritage funders expect from human-in-the-loop AI, and the through-line is consistent: defensibility is the deliverable.

Defensibility is the point, not a bonus

It is tempting to treat confidence flags, the review gate, and provenance as nice-to-haves layered on top of an AI tool. For an archive they are the opposite — they are the part that makes AI usable at all. The speed is only valuable if the output can be trusted, and output can only be trusted if it can be defended.

For archivists this preserves professional authorship and accountability over every record. For directors and heads of service, it means a sign-off they can give without fear that they are introducing undocumented, unaccountable content into the collection of record. The aim of our whole approach to trust is that adopting AI never costs you the evidential integrity that makes your archive worth keeping.

AI proposes; the archivist disposes; the record remembers both. That is what defensible AI in the archive looks like.

To see confidence flags, the review gate, and per-field provenance working on your own material, join the Archivers.ai waitlist for early access.