Trust & Transparency
A practical, transparent guide to how The Archiver uses artificial intelligence to help you manage your collections safely and securely.
We don’t just deploy AI blindly. Before any feature goes live, we rigorously test it in-house against known historical and archival test datasets to ensure both historical accuracy and the safe handling of sensitive subjects.
For European and UK archivists, data sovereignty is paramount. We prefer and actively default to European AI model providers (such as Mistral) for inference to keep data processing local, secure, and compliant with regional standards.
We balance sovereignty with performance. While European models handle text analysis and transcription effectively, they aren't fully optimised for complex video handling yet. In those specific tasks, we securely utilise advanced models like Google's.
AI at The Archiver is a tireless assistant, not a replacement. Every tag, metadata field, and transcription generated by our AI is placed into a review queue for a human archivist to approve, edit, or reject. Nothing is published or exported without archivist sign-off.
Trust Controls
Every AI-generated metadata field carries a confidence score (0–1). Low-confidence fields are visually flagged for archivist review, so you can focus attention where it matters most. No guesswork about what the AI is certain of and what needs a second look.
For each metadata field, the AI provides a brief explanation of why it suggested that value — what it saw in the document, photograph, or recording. This makes review faster and helps you understand decisions rather than accept them blindly.
Every action is logged: AI processing events, archivist edits, overrides, and approvals. The audit trail is timestamped, agent-attributed, and exportable. You can demonstrate exactly who did what and when — essential for institutional accountability.
Provenance
Every AI action and human review is recorded as a PREMIS event — the international standard for preservation metadata, maintained through the Library of Congress. These events are embedded in BagIt exports, creating a verifiable provenance chain from upload through to export.
EAD3 finding aid exports include maintenance history elements recording that AI-assisted cataloguing was used, when it was performed, and by which software agent. This follows EAD3's own provisions for documenting the creation and revision of archival descriptions.
Sensitivity & Care
Every item can be tagged with access conditions: public, restricted, or closed. The AI flags potential personal data (PII) and sensitive content, but the archivist makes the final decision on access restrictions.
Condition notes and archivist observations are injected into AI prompts — not overwritten by them. Your professional context shapes the AI's suggestions rather than being replaced by them. The archivist's knowledge is treated as primary.