Heritage teams are right to be cautious about AI. The risk is not only that a model gets something wrong. The deeper risk is that the archive becomes easier to search while becoming harder to trust.
AI enrichment changes the evidence layer
Preservica's AI archiving material points to image analysis, metadata cleanup, PII detection, OCR and transcription as areas where AI can help. Those uses are reasonable when the archive keeps a clear distinction between source record, machine suggestion and approved metadata.
The danger comes when enrichment overwrites provenance. A generated description can be helpful, but it should not become indistinguishable from a curator's record. A suggested date can aid discovery, but it should not erase uncertainty.
SBL's heritage quality work treats AI assistance as part of a governed record workflow, not as a replacement for archival judgement.
Uncertainty should be preserved, not cleaned away
Historical records often contain damaged text, ambiguous names, obsolete geography, inconsistent spelling and incomplete context. These are not defects to hide. They are part of the record's evidential condition.
AI systems tend to prefer clean outputs. Archive systems need to preserve confidence, alternatives and unresolved exceptions.
A good interface can still be simple for users. Simplicity should not require false certainty in the data layer.
The review workflow matters as much as the model
Heritage AI should define who can approve machine-enriched metadata, what evidence they see, what gets logged and how corrections propagate.
The workflow should also separate low-risk enrichment from high-risk interpretation. Object description may be lower risk than attribution, provenance, rights status or culturally sensitive classification.
That risk-based distinction is often missing from generic AI-in-archives discussions.
A practical provenance test
Pick one AI-enriched record and ask five questions. What was the source? What did the AI add? Who approved it? What confidence or uncertainty remains? Can a future researcher see the difference?
If the system cannot answer those questions, the archive has gained convenience but lost accountability.
AI can make heritage collections more discoverable. It should not make them less defensible.
Questions teams ask before they start
Can AI be used safely in heritage archives?
Yes, when source records, AI suggestions, human approvals and uncertainty are clearly separated and logged.
What is provenance in an archive?
Provenance is the record of origin, custody, context and source relationships that allows users to trust the material.
What AI archive tasks are lower risk?
Metadata cleanup, duplicate detection and OCR assistance are often lower risk than attribution, rights interpretation or sensitive classification.