In a discussion about user data exposure in Discourse AI, BrianC asks what user information—like IPs or personal data—is shared with LLMs when using AI plugins on their subdomain-hosted Discourse site. merefield, developer of the Chatbot and AI Topic Summary plugins, clarifies that only usernames and raw post content (Markdown) are sent to the LLM. He notes that any personally identifiable info mentioned in the post content (e.g., names, addresses) will also be included, but metadata like IPs or user profiles are not transmitted. Verbose logging can reveal the exact queries sent. BrianC acknowledges the plugins are excellent and expresses concern about LLM privacy policies, noting usernames alone aren’t worrisome but context within posts could be. trusktr adds that for public content, exposure to AI is no different than to search engines or humans—but warns that private posts or business-sensitive discussions should not be sent to LLMs, as they could be reused or exposed to others. For open-source communities, however, sharing public content with AI is seen as beneficial for collective learning.
I’m using Discourse AI on my site, which runs on a subdomain (community.website.com), and I’d like to better understand what kind of user information might be shared with the language model (LLM) during interactions. Specifically, I’m curious about:
What types of user data (e.g., personal information, IP addresses) could potentially be exposed to the LLM?
Are there any safeguards in place within Discourse AI to limit or anonymize what gets sent?
For some additional context, my setup uses Caddy as a reverse proxy and Sucuri for DNS and firewall. If anyone has insights on how this configuration might affect what is exposed—or just general knowledge about how Discourse AI handles user data—I’d really appreciate the input!
Looking forward to hearing from those who have a better understanding of this.
I believe you’ve been using my AI plugins at some point, Chatbot and AI Topic Summary, since you’ve posted in those Topics, so I will respond for those, but if you want more information, please post in those Topics.
Both of my plugins send usernames and raw Post content (ie the markdown). NB if someone mentions someone’s name in a Post, or an address that will get sent in the markdown, of course, but otherwise Users are only depicted by Usernames.
Other metadata is not sent, e.g. IPs, User Profiles, etc.
You can see the queries being sent in logs if you select the option for verbose logging and divert the logs to Warn (there’s another setting) so they are visible in /logs.
Thank you Robert. Yes, I do use those plugins which are excellent. Appreciate the feedback. After reading some of the LLM privacy policies transferring sensitive data for users would be concerning. Obviously whatever the context within the chat will be sent and the username by itself really not concerning. Some of the LLMS terms are quite invasive so that’s what spurred my inquiry. Thanks again
That’s good its only usernames and post content. As long as it is only publicly-visible content, then really it makes no difference if it was a search engine, an AI, or a human, that saw some content and spread it or derived something from it.
I would be concerned about private posts/categories that are for certain logged-in users. If sensitive business discussion is happening, and that stuff goes to an AI, well now the AI can present those ideas to anyone else who may ask for business ideas. Or similar.
My site is for an open source project, so the more data gets sent out for AIs to learn from, the better for helping everyone out.