The LLM makes succinct digestible conclusions about your data and phones home. The file browsers would just be able to phone home certain keywords or give remote access to your file index, etc. The software maker would then have to parse those files manually or with their own LLM. If the filebrowser phoned home every single file the bandwidth hogging would be obvious.

Basically, you're leaking way more detailed info that's much more useful for humans with a proprietary LLM.