Research Notes

Long-form research for directors of digital collections and working archivists. Each note examines one question in machine-assisted archival description, cites archival standards and professional guidance, and identifies the author's own proposals. Every note is versioned and includes an address for corrections. Read on the web or as a PDF, without signing up.

Archively Research Note 01

When Does an AI Proposal Become Archival Description?

A human reviewer is necessary, but not sufficient. Professional control depends on who may accept a machine proposal, what context and criteria they use, what choices they can actually make, and what evidence and correction path the institution keeps afterwards.

Machine systems now produce text that looks like archival description: transcripts, candidate dates, titles, scope and content notes, subject terms, biographical statements.

Saad Muhammad, 28 min read, 4 figures, version 1.0, 6 September 2026, PDF edition

Read the note

A scatter plot. The horizontal axis is the consequence of an error, labelled low, moderate and high. The vertical axis is interpretive distance from the source, labelled extracted at the bottom, summarised in the middle and inferred at the top. Three shaded bands run diagonally across the plot, light to deep, each carrying its own name: Tier A, provisional or internal, at the lower left; Tier B, accepted public description, across the middle; Tier C, high consequence or contested, at the upper right. A single vertical rule near the right of the plot marks the consequence floor, past which the tier is C however mechanical the output is. Ten output classes are plotted as small filled circles with their numerals set beside them: transcript text, dates and named entities, captions, title and scope note, authority links, relationships, administrative history, biographical history, sensitivity signal, and rights decision. A numbered key gives each class its name and the tier it falls in under the intended use currently shown. The figure opens on an example: class 1, OCR or transcript text, used as public archival description. That point is drawn larger, in archival blue, on a dotted line marked same output, four uses, whose four square stops are the positions the same class would take under each of the four intended uses. Choosing a different intended use, from internal discovery aid, staff working draft, public archival description, and rights or access decision, moves every point horizontally and can change its tier. A full table of tiers by output class and intended use follows the figure.

Figure 3 of 4. Review depth changes with use and consequence