This article explains what your community's AI features send to AI providers, what Mateflow stores, what members can see in AI answers, and where AI can fall short. Read it before you turn on AI features.
Who can do this: Owners and admins set up AI features; signed-in members use them. · Where: AI features are set up in the web admin console; members use the AI assistant on the web and in the Mateflow app. · Plan: AI features depend on your platform plan.
What the AI assistant sends to an AI provider
Mateflow manages the AI models for you and processes AI features through OpenAI's API:
- The AI assistant (AI Copilot), the AI writing tools, and course outline and quiz generation use OpenAI's GPT-4o mini. These requests pass through Cloudflare AI Gateway on the way to OpenAI.
- AI moderation uses OpenAI's omni-moderation-latest model.
- The knowledge base uses OpenAI's text-embedding-3-small model to index your sources so questions can be matched to them.
Neither members nor admins can change these models. Under OpenAI's API data policy, data sent through the API isn't used to train OpenAI's models by default, and OpenAI may keep it for up to 30 days to monitor for abuse.
When a member asks the AI assistant a question, Mateflow first searches your knowledge base and, if it's turned on, your community posts. It then sends the question with the matching excerpts to the model. The assistant doesn't search the web. Other AI features send only what they work on: the text someone asks Enhance with AI to improve, the course details for AI Course Outline, or the lesson text or captions a quiz is built from. If your community already uses its own AI provider key (BYOK), these requests go to that provider.
What AI moderation sends to OpenAI
AI moderation is a separate feature, set up under Admin → Settings → Content Safety. The page says it's Powered by OpenAI Moderation. Content is sent to a third-party service — review your privacy policy before enabling.
- Only the title and body text of what you pick under Scan is sent: Top-level posts, Replies and comments, or both. Images, attachments, and videos aren't sent.
- When Enable AI moderation is off, nothing is sent.
- Each decision's scores and categories, without the post text, are kept for 365 days.
AI moderation can be wrong in both directions: it can flag harmless posts and miss harmful ones. It doesn't make final decisions on its own. In Enforce mode, content it hides or flags waits for a moderator to keep or remove it. See Use AI to moderate content.
What Mateflow stores from AI features
Mateflow keeps these AI records for your community:
- Members' AI chats. Each member's conversations are saved under Chat History, and only that member sees them. Chats expire after 180 days. Members can delete a chat sooner, on the web or in the app; on the web they can also rename or export it as Markdown or JSON.
- Hit Logs. Each question, the answer, and its citations are recorded for owners and admins under Admin → AI Copilot → Hit Logs. The page doesn't show which member asked. Deleting a chat doesn't delete its Hit Logs. See Review AI assistant answers and fill gaps.
- Knowledge base files and pages. Files you upload and pages Mateflow crawls are stored and indexed for the AI assistant only. They don't appear in your media library. Deleting a source removes its indexed documents.
Community posts aren't copied into the knowledge base; the assistant searches them where they live. When you remove a member from your community, their AI chats and Hit Logs are deleted. When a community is deleted, all its AI data is deleted with it.
Which content members can see in AI answers
The AI assistant answers each member only from community posts that member can already open. Posts in Private or Secret spaces are used only for members of those spaces. Drafts, posts waiting for approval, hidden or deleted posts, and posts in archived spaces are never used. Guests can't use the AI assistant.
The knowledge base is different. Knowledge sources have no member-level permissions, so anything in the knowledge base can appear in an answer to any member who uses the AI assistant. Don't add files or pages that some members shouldn't read.
To control which spaces are searched, see Choose which spaces your AI assistant can search.
AI answers can be wrong
The AI assistant can misunderstand a question or its sources. Members see AI can make mistakes. Please verify important information. below the chat box. Each answer lists its Sources, and hovering over its percentage shows AI confidence score for this response. When the evidence is too thin, members see Insufficient Evidence instead of a guess. An answer that stops early is marked Generation stopped before finishing — this answer is incomplete.
An answer is only as current and correct as your knowledge base and posts. Review low-confidence questions in Hit Logs and fix the sources behind them.
Related articles
Was this guide helpful?