Hey guys, I have a contact who liked Fibery. They work with charities. Charities are very sensitive about donors’ personal/financial data being leaked by using AI. What solutions can we use so that AI doesn’t “steal” the sensitive data? In Fibery AI chat window, I think I could use # context. What else?
The only safe solution is to disable AI
Or self-hosted AI?
From ClickUp:
ClickUp AI is not trained on data from your Workspace. We’ve secured licensing with our partners to ensure they do not access your data for training purposes. We also have zero data retention agreements with all of the large language model (LLM) organizations we partner with. The agreements require our partners not to retain any data from your Workspace after your data is input and processed through the LLM. Additionally, we use in-context learning (ICL) to ensure that our models are not learning from data.
What do you think about this, how truthful is this?
I have no idea to be honest, but in the nutshell it is all about trust to OpenAI and Anthropic. In general they indeed state that data is not used for training, etc. Same for Fibery. But it is your choice to trust them or not.
local LLM is not an option so far.
Just adding a +1 on this topic.
Given that AI Search is much more powerful than standard Fibery search, it is unfortunate that there is no way to run it through a self-hosted (by user or by Fibery) model as @myg_ge suggested, or to ensure zero data retention. Currently, I cannot use Fibery’s AI search due to org policies. It would be great if there were “solutions […] so that AI doesn’t “steal” the sensitive data?” as @myg_ge said, that aren’t just to disable AI.
To clarify my request: it would be very useful to have either a) ability to use a self-hosted model for semantic search or, as a less ideal but probably easier-to-implement solution, b) Fibery self-hosts the model for semantic search with zero data retention OR gets a zero data retention agreement with the subprocessors.
I would assume many other Fibery users (or potential Fibery users) would like a solution to this too, but maybe I’m being selfish 
Thanks for considering!
Other relevant comments, for reference
For clarify, I can see this has been discussed before here:
And here:
And in the Expose Semantic Search through MCP request