Back to Reference
Contents
Reference
360 words
Jun 25, 2026

AI usage transparency report

The AI usage transparency report (Administration, /admin/ai-usage) is a read-only dashboard showing where AI has assisted with the catalogue and how much of that assistance a person has reviewed. It counts only - it makes no AI calls and changes no record.

It exists for honesty and accountability: AI here is an assistant that proposes metadata (entity recognition, summaries, handwritten-text recognition, translation, and so on). A person always remains responsible for what enters the record. Nothing in this report implies that AI decides anything.

What it shows

  • Inferences logged - the total number of AI inferences recorded against the catalogue.
  • Records touched - how many distinct records have had at least one AI inference applied.
  • By inference type - a breakdown of which AI tasks ran (for example entity recognition, summarisation, handwritten-text recognition, translation), each with a count, a share of the total, and a progress bar. Where an endpoint was recorded, the gateway host is shown as a small hint.
  • By model - which models produced the inferences, each with a count, share, and bar.
  • Human oversight - how many inferences carry a recorded human review or correction, shown as a count and a "reviewed share" percentage. This is the accountability headline: it measures how much AI output a person has checked, not how accurate the AI was.
  • Inferences over time - inferences logged per month over the trailing year, drawn as simple bars.

How to use it

Use it to answer "how much has AI touched our catalogue, and have we reviewed it?" A low human-oversight share is a prompt to review more of the AI-assisted metadata. The over-time bars show when AI assistance happened. The type and model breakdowns show which tools and models did the work.

Notes

  • The report is read-only and never edits a record or calls any AI service.
  • It is built from cheap aggregate counts, so it is safe to open on a large log.
  • It is jurisdiction-neutral and makes no country-specific assumptions.
  • With no AI activity recorded yet, it shows a calm "No AI activity recorded" state rather than an error.