How Wikiwright uses AI
AI in Wikiwright suggests; people decide. It is off until an administrator turns it on, uses models hosted by Atlassian, and never changes a page without a person’s approval.
Two kinds of AI
- The Wikiwright Rovo agent runs in Atlassian’s Rovo, under your organization’s Rovo settings. It answers from Wikiwright’s link index through its actions, and any change it proposes runs only after the person confirms it in chat. It works whether or not AI suggestions are on.
- AI suggestions inside Wikiwright use Atlassian-hosted models (Forge LLMs). They are optional, and the rest of this page is about them.
What AI suggestions do
| Feature | What is sent to the model |
|---|---|
| Suggest a page for a broken link | The broken link’s text, the title it pointed at, and the titles and spaces of candidate pages |
| Match a missing section | The page’s title, the missing section’s name and the page’s current headings |
| Write a find-and-replace rule | Your description of the change, and the web domains your links use |
| Choose pages to link an orphan from | The orphan page’s title and the titles of candidate pages |
| Weekly clean-up plan | For the targets with the most broken links: the same as “suggest a page” above |
| Published report summary | Counts and trends, and space names only if an administrator chose to list them |
| Compare two similar pages | Each page’s title, last update date, headings and opening text |
Models and data
- Models are Anthropic’s Claude Haiku and Claude Sonnet, hosted and run by Atlassian through Forge LLMs. Nothing is sent outside Atlassian, and we don’t use your data to train any model.
- Only the data listed above is sent, and only for pages Wikiwright can read. The AI never decides who can see what: results are shown through each person’s own Confluence permissions.
- Requests pass through Atlassian’s content moderation for Forge LLMs.
- Answers are cached for 90 days so the same question isn’t asked twice.
Controls
- Off by default. A Confluence administrator turns AI suggestions on under Settings, and can turn them off at any time.
- Monthly limit, set by the administrator (US$2 of estimated AI use by default, at most US$5), with this month’s use shown. AI use is included in Wikiwright’s price: it isn’t billed to you. Each person is also limited to 30 AI requests a day.
- Organization policy. If your organization doesn’t allow apps to use AI models, AI suggestions stay off.
- Without AI, Wikiwright’s own suggestions (renamed and moved pages, pages with the same title elsewhere, renamed headings) keep working.
Checking the answers
- Page text goes to the model as data, never as instructions, and answers are checked against the shape each feature expects. When it picks a page, a heading or a place to add a link, it can only pick from options Wikiwright built itself, so it can’t invent a page. A find-and-replace rule it writes is checked and shown as a preview before it runs.
- AI picks are labelled as guesses and reviewed one at a time; the weekly plan keeps only the model’s most confident picks.
- Before release, each feature is tested against a fixed set of cases on the real models, including attempts to hide instructions in page text. In the latest run, every injection attempt was resisted.
- Every change still shows a preview, needs a person’s approval, and link fixes can be undone.
Questions
Email support@extoris.com.