--- title: Cultural Sensitivity Review description: Review AI-flagged harmful language and culturally sensitive subjects in catalog records, install acknowledgments, and add community cultural advisories. section: guides order: 14 updated: 2026-07-17 verified: 2026-07-17 related: [guides/content-screening, guides/content-warnings, guides/working-with-artifacts, guides/browse-views] features: [cultural-sensitivity] --- # Cultural Sensitivity Review Historical catalog records often carry language that was ordinary when written and is now recognized as racist, demeaning, or outdated — original captions, donor descriptions, superseded subject headings. Separately, some material carries cultural-protocol obligations: sacred or restricted subjects, human remains, and names or images of deceased persons. Cultural sensitivity review surfaces both automatically so you can contextualize them, while the historical record stays unaltered. This is the **text** counterpart to [content screening](/docs/en-US/guides/content-screening), which looks at images. Sensitivity reads an artifact's metadata and transcribed text — title, description, original captions, subjects, OCR and transcripts — and never the pixels. The guiding rule throughout: **AI proposes, curators approve, communities author.** Nothing AI flags is shown to visitors until a person acts on it. ## How findings are produced After an artifact is analyzed, when new transcribed text is added, and (debounced) when you edit metadata, Preservated runs a text-only pass over the record. Each finding lands in one of two tiers, which route to two separate queues with different obligations: - **Descriptive language** — slurs, outdated terms, superseded subject headings, colonial naming, and stereotyped period description. These become a viewer-facing acknowledgment once approved. - **Community protocol** — sacred or restricted subjects, human remains, and named or depicted deceased persons. AI only *surfaces the candidate*; the wording must come from the relevant community. A descriptive finding rated caution or stronger also bumps the artifact's content rating to `caution` and stages a provisional warning — held in the queue, hidden from visitors until you approve it (unless you turned on immediate AI warnings in [Content Safety](/docs/en-US/guides/content-warnings#the-sensitive-content-statement)). Review is ongoing and never complete: not every record with harmful language is flagged, and some flags are wrong. That is expected. ## Review the queues Open **Admin → Operations → Sensitivity**. Use **Run sensitivity review** to queue the pass over records that have never been analyzed. ### Descriptive language Findings are grouped into one card per artifact — a thumbnail and a link to the item (opens in a new tab), then each finding listed beneath it. Cards are ordered so the items with the most severe, or the most numerous, concerns sit at the top. Each finding shows the exact quote, the field it appears in, why it was flagged, and a severity. - **Approve** installs a viewer-facing acknowledgment in the institution's provenance-and-context voice — for example, that a term "was used at the time it was created but is now recognized as offensive" and is "preserved unaltered as part of the historical record." Use **Edit text** first if you want to tailor the wording for that item; you are the author of what visitors read. - **Dismiss** removes the finding. Dismissals are remembered — re-running the pass will not flag the identical quote again. Approving an item adds its acknowledgment to the [advisory banner](/docs/en-US/guides/content-warnings#what-visitors-see-on-a-flagged-item) on the artifact page. Dismissing the last caution-level finding on an item lowers its rating back if nothing else holds it up. ### Community protocol Sacred or restricted subjects, human remains, and deceased-person references require consultation with the relevant community — so these findings carry **no AI-drafted text**. When you install a disposition you can attach any combination of: - **A community-authored label** — text supplied by or in consultation with the community. This is never AI-generated. - **A cultural advisory** — a prominent, community-addressed notice on the item (see below). Choose *pending consultation* for the honest in-between state where you have flagged a candidate but community review hasn't happened yet, or *in consultation* once it has. - **Hide from public browsing** — circulation control that removes the item from public browse, search, and similarity results entirely, pending consultation. This is independent of the content-rating ladder. For a multi-page document, it also limits the hover page-preview in [Browse views](/docs/en-US/guides/browse-views#preview-a-documents-pages-without-opening-it) to the cover page only — so administrators who can still see the hidden item never get a hover preview of its interior pages either. :::caution Hiding an item from public browsing affects every visitor-facing surface at once. Administrators still see it. Use it as a holding state while consultation is underway, and lift it once the community has decided how the material should circulate. ::: ## Cultural advisories on the item page A cultural advisory is distinct from a content warning: it is not gating for shock, it respects a cultural prohibition, and by convention it addresses the affected community's members. The default treatment is a prominent banner; the item is not blurred and nothing is gated. Set the regionalized defaults in **Admin → Site Settings → Content Safety → Cultural advisory**: - **Addressed community** — the default community named in the advisory when an item doesn't specify one. - **Advisory text** — the body. Leave it empty for the platform default. The placeholders `{CONTAINS}`, `{COMMUNITY}`, and `{CONSULTATION}` are filled in per item. The deceased-persons convention is strongest for Aboriginal and Torres Strait Islander materials; institutions working with other communities should rewrite the default to match. - **Show advisories as a full interstitial** — an opt-in that presents the advisory as a text-first "Continue / Go back" screen before the item, instead of a banner. There is still no blur. Enable it only if your community partners request it. ## What gets stored Findings and their decisions live with the artifact, so deleting AI history never loses your review work. Community labels and advisories are recorded as community-authored and are never overwritten by a later AI run. This is the durable schema for community-protocol labels; a future Local Contexts Hub or Mukurtu integration needs no migration.