I find the "trained on open data" line of reasoning really weird.
The datasets behind these models are from web scrapes, containing many sentences clearly taken from news websites that don't appear to have licensed their IP for that purpose - I don't see a difference between that and AI companies scanning in copyrighted books or art.
(Note: I'm not making a judgment either way on this, but if you are branding yourself on AI ethics, I think you need to be consistent in how you talk about this)
Firefox uses on-device downloaded-on-demand ML models for privacy-preserving translation.
They're not LLMs. They're trained on open data.
Should translation be disabled if the AI 'kill switch' is active?