Skip to content

AI-Assisted Archives Workflows

This page sets initial expectations for how AI tools may be used in Archives workflows. It is intentionally conservative and can be refined later as local practice develops.

Purpose

AI tools may be useful for repetitive support work, workflow automation, script drafting, and formatting assistance, but they do not replace professional archival judgment. This page defines a cautious starting point for approved and non-approved uses.

General Position

  • AI output should be treated as draft assistance, not authoritative archival description.
  • Staff are responsible for reviewing, correcting, and approving all final work.
  • AI tools should support workflows, not replace human decision-making about arrangement, description, context, or subject analysis.

Generally Appropriate Uses

Examples of lower-risk and more easily reviewable uses include:

  • drafting batch files, PowerShell scripts, or small utilities for file processing
  • generating checksum, rename, conversion, or metadata-helper scripts
  • creating citation-format helpers, including APA-style cover page support
  • generating draft transcripts from oral histories or similar recordings, especially when using locally run tools such as Whisper
  • reformatting lists into HTML, Markdown, tables, or spreadsheets
  • turning rough notes into a clearer draft for staff review
  • summarizing long OCR text for internal review
  • generating draft alt text, captions, or labels
  • normalizing metadata into a template or repeated structure

Uses That Require Caution

These tasks may be useful, but they require especially careful review:

  • reviewing and correcting AI-generated transcripts for names, dates, places, jargon, and speaker labels
  • drafting a first-pass description from an access PDF or OCR text
  • drafting item or collection descriptions
  • suggesting subject terms
  • extracting names, places, and dates from messy OCR
  • identifying likely formats, genres, or document types
  • creating public-facing language that implies certainty

Not Approved Without Further Review or Policy

  • uploading restricted, confidential, or sensitive records into external AI systems without approval
  • treating AI-generated text as final archival description without human review
  • allowing AI to invent names, dates, creators, provenance, or relationships
  • using AI output as evidence without checking the source material
  • using AI to make appraisal, access, privacy, or ethics decisions on its own
  • using AI-generated scripts for destructive file operations without careful testing

Human Review Requirements

Before using AI-assisted output in a final workflow:

  1. Compare the output to the source material.
  2. Correct factual mistakes, invented details, and misleading phrasing.
  3. Check for tone, bias, and inappropriate assumptions.
  4. Confirm that names, dates, identifiers, and citations are accurate.
  5. Make sure the final language reflects local descriptive standards and institutional practice.
  6. If the output is a transcript, review names, dates, places, unclear words, and speaker attribution carefully before reuse.

If the output is a script or automation tool:

  1. Test it on a small, safe sample before wider use.
  2. Confirm that it does not rename, move, delete, or overwrite files incorrectly.
  3. Review the logic before sharing it with other staff or institutions.

Privacy and Sensitive Material

  • Do not enter sensitive or restricted material into external AI tools unless that use has been approved.
  • When in doubt, remove or redact sensitive details before testing prompts.
  • If the records involve privacy concerns, donor restrictions, FERPA, or other access limitations, stop and ask before using AI.
  • Locally run transcription tools may be a better fit than external services when working with oral histories or other potentially sensitive recordings.

Documentation and Transparency

When AI meaningfully contributes to a workflow, staff should be prepared to document:

  • what tool was used
  • what kind of task it assisted with
  • what human review was performed
  • what final edits were made before publication or use

Starter Examples

Examples of reasonable prompt types:

  • “Write a Windows batch file that calculates MD5 and file size for a dragged PDF and copies the formatted result to the clipboard.”
  • “Create a small PowerShell script to help format citation text for VText cover pages.”
  • “Create a draft transcript from this oral history audio, but clearly mark uncertain words and do not assume speaker identities.”
  • “Convert this bulleted list into clean HTML.”
  • “Turn these notes into a checklist for staff review.”
  • “Extract names and dates from this OCR text and mark uncertain readings clearly.”

Examples involving archival description should be treated as more sensitive and reviewed more carefully than formatting or scripting tasks.

Status

This page is a starting policy draft. It should be revisited as local practice, institutional comfort, and professional expectations evolve.