26/09/2026

Making AI Metadata Fit Archival Standards

Written for archivists, librarians, cataloguers and digital collections staff who have AI-drafted descriptions and need them inside the catalogue.

AI can now draft a title, a summary, names, places and dates for a scanned document in seconds. Projects rarely stall there. They stall when a few thousand AI-generated rows sit in a spreadsheet that nobody can import: the columns match no catalogue field, the names are spelled three ways, and the dates say "early 1920s?".

Machine-generated metadata is only useful if it arrives in the catalogue in the shape the catalogue expects. That shape is set by standards, and it has to be decided before the first record is drafted.

Decide the target schema first

The schema is whatever your catalogue system imports. Check that before anything else.

  • Dublin Core. Fifteen general elements (title, creator, subject, description, date, type, coverage and so on), maintained by the Dublin Core Metadata Initiative. It fits simple item-level records and is the baseline format for OAI-PMH harvesting. Structured names, date qualifiers and hierarchy do not fit well.
  • MODS (Metadata Object Description Schema, Library of Congress). Richer bibliographic description in XML: structured names with roles, several date types, subjects broken into topic, geographic and temporal parts, and attributes that link a value to its vocabulary.
  • EAD (Encoded Archival Description; EAD3 is the current version). The encoding for hierarchical finding aids: fonds, series, file, item. If your AI is describing a collection rather than isolated objects, this is usually the target.

Behind much archival practice sits ISAD(G), the International Council on Archives' General International Standard Archival Description (second edition, 2000). It is a content standard, not a file format: it defines the elements of a description and multilevel description from fonds to item. AtoM's archival description template follows it. The ICA's newer Records in Contexts (RiC), a conceptual model and ontology whose version 1.0 was released in 2023, describes records, people and activities as a network rather than a strict tree. It matters for linked-data planning; for most institutions the import route is still EAD, MODS or Dublin Core.

Map every AI output to a specific field

Before running a model, write a mapping. Every AI output needs exactly one destination field. A typical mapping:

AI output Dublin Core MODS EAD
Draft title title titleInfo/title unittitle
Summary description abstract scopecontent
Date date originInfo/dateCreated unitdate
People, organisations creator / subject name origination / controlaccess
Places coverage subject/geographic geogname
Topics subject subject/topic controlaccess/subject
Document type type genre genreform
Language language language langmaterial

Two rules follow from the table. If an output has no field, do not produce it. Speculative interpretation has nowhere to go and only creates review work. And full OCR text is not descriptive metadata. It belongs in a separate text layer (ALTO, for example), linked to the record, as the platform-versus-pipeline guide discusses under export.

Controlled vocabularies: propose a term, never invent one

A model left to write free-text subjects will produce "Jewish schools", "Jewish education" and "Hebrew schools" across three records that a cataloguer would give one heading.

The fix is to constrain the model to existing vocabularies and authority files:

  • Library of Congress Subject Headings (LCSH) for topics, and the LC Name Authority File for names.
  • VIAF, the Virtual International Authority File hosted by OCLC, which clusters name records from many national libraries.
  • Wikidata, for people, places and organisations, with links to many authority files.
  • National authority files. For Hebrew-language collections, the National Library of Israel's authority file (Mazal) allows a preferred form of a name in up to four scripts: Hebrew, Latin, Arabic and Cyrillic.
  • Your own local vocabulary, if the institution has one. It takes precedence.

The AI should propose a term from the vocabulary, with its identifier, not free text. In practice: retrieve candidate headings first, then let the model choose among them. When nothing fits, the value is routed to a cataloguer as a suggested new heading. MODS and EAD3 both have attributes for recording which vocabulary a term came from and its identifier, so the link survives import.

Dates and place names

Archival dates are often uncertain, and the uncertainty is information. Keep two forms:

  1. The display date, as a cataloguer would write it: "circa 1920s", "[1923?]".
  2. A normalised date for machines. Use ISO 8601 for known dates (1923-05-14). For uncertain and approximate dates, use the Extended Date/Time Format (EDTF), developed by the Library of Congress and later incorporated into ISO 8601-2. In EDTF, 1923? means uncertain, 1923~ means approximate, 192X means an unspecified year in the 1920s, and 1920/1929 is an interval.

MODS accepts EDTF directly (an encoding="edtf" attribute, plus a qualifier for approximate, inferred or questionable dates), and EAD lets you store a normalised date alongside the display text.

Place names need the same discipline: Vilna, Wilno and Vilnius are one city. Transcribe the place as the source gives it in the description; link the access point to an authority record. Do not let the model silently modernise names.

Mark what the machine wrote

Once imported, nobody can tell which values a person checked and which a model guessed unless you record it. For each AI-proposed field, keep:

  • that it was machine-proposed, and by which model and version,
  • the confidence score,
  • who reviewed it, when, and whether it was accepted, edited or rejected.

Standards give you some of the containers, not all. EAD3's maintenance history can record events with an agent type of human or machine. MODS has a record information section for notes on the record's origin. Simple Dublin Core has no place for per-value provenance, and dcterms:provenance is not the answer: it describes the custody history of the resource, not of the metadata. Keep field-level provenance in local administrative fields or a sidecar file, and add a plain public note such as "Description drafted with machine assistance and reviewed by archive staff". The audit trails post covers the wider chain of evidence.

Review workflow: a person approves every record

Nothing machine-generated reaches the public catalogue without a human decision. Order the review queue:

  1. Low-confidence fields first, especially names and dates, where one error misleads searchers.
  2. High-confidence fields sampled, not skipped. Confidence scores are only useful when calibrated against checked results; the confidence-calibration case study shows a "certain" category that was right only 56% of the time until its threshold was recalibrated, after which it reached 90%.
  3. Edit rates tracked per field. They show where the model helps and where it does not.

The AI photo cataloguing guide applies the same workflow to photographs.

Multilingual description

Collections in Hebrew, Yiddish or Russian are often described in two languages, and AI makes it tempting to generate the second one automatically.

  • Tag every value with its language. Dublin Core in XML takes xml:lang; MODS and EAD both carry language and script attributes, and MODS can group parallel versions of a field.
  • Keep names in the original script alongside the romanised form, using the romanisation table your catalogue follows (the ALA-LC romanization tables include one for Hebrew and Yiddish).
  • A machine translation of a description is a new machine-generated value. Mark it and review it like any other field.

Pilot on 100 records before scaling

Run the whole path on about 100 records that represent the collection, including the difficult ones:

  1. Draft metadata against your mapping.
  2. Have your actual reviewers check it.
  3. Measure time per record, edit rate per field and how often vocabulary matching found a term.
  4. Import into a test instance of your catalogue, then export and check nothing was lost.

If the import works and the edit rates are acceptable, scale. The backlog sprint model describes how to size the full run around your review capacity.

A short checklist

  • Target schema chosen by what the catalogue imports: Dublin Core, MODS or EAD.
  • A written mapping; every AI output has one destination field.
  • Subjects and names chosen from LCSH, VIAF, Wikidata, national or local authority files, with identifiers.
  • Display dates plus ISO 8601 or EDTF normalised dates.
  • Places as in the source, linked to authority records.
  • Field-level provenance: model, confidence, reviewer, decision.
  • A public note that description was machine-assisted.
  • Review queue ordered by confidence; high-confidence fields sampled.
  • Language and script tagged on every value.
  • A 100-record pilot, imported and exported, before the full run.

If you have AI-drafted descriptions that need to become catalogue records, see how we approach archive cataloguing, or get in touch with a sample of records and the name of your catalogue system.

Frequently Asked Questions

Which metadata standard should AI-generated descriptions follow?

The one your catalogue system already imports. Dublin Core suits simple item-level records, MODS suits richer bibliographic description, and EAD suits hierarchical finding aids. Decide this before any AI runs, because it determines what the model is asked to produce.

Can AI choose subject headings from LCSH or an authority file?

Yes, if it is constrained to do so. The reliable pattern is to retrieve candidate terms from the vocabulary and have the model choose among them, storing the term's identifier. When nothing fits, the value goes to a cataloguer as a suggestion, not into the record as free text.

How do you record uncertain dates like 'circa 1920s'?

Keep the human-readable date as written, and add a normalised form in ISO 8601 or, for uncertain and approximate dates, the Library of Congress Extended Date/Time Format (EDTF). In EDTF, 192X means an unspecified year in the 1920s and 1923~ means approximately 1923.

Should machine-generated metadata be labelled in the catalogue?

Yes. Record which fields were AI-proposed, with what confidence, and who reviewed them, and give users a plain note that the description was drafted with machine assistance. Future staff need this to judge which values to trust.