---
title: AI Settings and House Style
description: Turn individual AI features on or off for your institution and teach the AI your cataloging style with house guidance and examples.
section: guides
order: 10
updated: 2026-07-11
verified: 2026-07-11
related: [guides/ai-analysis, guides/ai-tools-and-provenance, guides/cultural-sensitivity, guides/transcription-and-ocr]
features: [institution-ai-settings, ai-analysis]
---

# AI Settings and House Style

You'll learn how to control which AI features run for your institution and how to shape the AI's output to match your catalog's voice.

Both live in **Admin → Settings**, in the **AI & Cost Controls** section (the **AI Features** and **AI Guidance** groups). You need the admin role to change them.

## Turn AI features on or off

The **Enable AI features** switch controls everything at once. Below it, each analysis type has its own toggle:

- **Metadata analysis** — titles, descriptions, dates, and subjects ([AI Analysis](/docs/en-US/guides/ai-analysis))
- **Museum vision** — face detection, visual interest, and smart-crop regions that automatically frame browse thumbnails on the detected subject
- **Transcription** — audio/video transcription with speaker identification
- **OCR** — text recognition in scanned PDFs and page images
- **Audio tours** — narrated tour scripts and text-to-speech audio
- **Colorization** — colorizing black-and-white photographs
- **Image remix** — stylized derivatives such as coloring-book versions
- **Content screening** — flagging media that needs curator review ([Content Screening](/docs/en-US/guides/content-screening))
- **Sensitivity review** — flagging catalog language for reparative description ([Cultural Sensitivity](/docs/en-US/guides/cultural-sensitivity))
- **Admin assistant** — the Ask tab in the admin search modal, including its confirmed site changes ([Admin Assistant](/docs/en-US/guides/admin-assistant))

Everything is on by default. Turning a feature off stops **new** runs everywhere it can start — the artifact page buttons, batch operations, and background jobs. Existing results are untouched: transcripts, tours, and analyses that already ran stay visible and editable.

Two side effects to know about:

- Turning off **screening** or **sensitivity review** also stops the automatic passes that run when media is ingested or metadata changes. New items will accumulate as "unscreened" — relevant if you use child-safe browsing with the "exclude unscreened" policy ([Content Warnings](/docs/en-US/guides/content-warnings)).
- Turning off **transcription** also stops transcription during imports.

:::tip
Since every AI call has a per-run cost, these toggles are also your simplest cost control: turn off what you don't use. Spending by feature is visible in the AI usage view.
:::

## Teach the AI your house style

Museums have strong opinions about how their catalog reads. Rather than accepting generic output, write **AI guidance** — your own notes that are added to every AI prompt as institutional context.

**House style** applies to all AI features. Keep it short and concrete:

> Title photographs as "[Subject], [Place], [Year]" when known. Prefer "photograph" over "image". Use "Coast Salish", not outdated terms. Our region is Whatcom County, WA — assume local context.

**Per-feature notes** add to the house style for one feature:

- **Metadata analysis** — titling conventions, vocabulary preferences, what to emphasize or avoid.
- **Audio tours** — narration voice, audience, length, what to foreground.
- **Sensitivity review** — community self-names and institution-specific terms to watch for.
- **Transcription** — local proper nouns, terminology, and expected languages. This one works differently: it seeds the speech recognizer's context so unusual names and terms are transcribed correctly.

**Examples** are optional before/after pairs the AI can pattern-match against — an input like "a photo of a schooner at a dock" paired with the output you'd want: "Schooner at Sehome Dock, Bellingham, 1908". Up to five.

### What guidance can and can't do

Guidance shapes tone, terminology, emphasis, and local knowledge. It is deliberately **additive**: it cannot override Preservated's safety rules, change what fields the AI produces, or auto-write cultural advisories (those remain community-authored — see [Cultural Sensitivity](/docs/en-US/guides/cultural-sensitivity)).

Guidance text rides along on every AI call, so length is capped and shorter is better — a few high-signal lines beat a page of prose.

Every AI result records which version of your guidance was active when it ran, as part of the provenance trail ([AI Tools and Provenance](/docs/en-US/guides/ai-tools-and-provenance)). After you edit guidance, re-running analysis on an artifact produces a result attributable to the new text.