Copilot is weird and can give out very weird responses that have little to do with your conversations.
And of course it might just grab context depending on what you do (e.g. clicking the copilot button might already do that).
I found it works best as GPT model if you disable the fancy stuff like search. It too easily looses track of what happened or completly goes off the rails.
(i believe disabling search is a beta feature in some regions, but its a hidden flag you can theoretically set, i made a tampermonkey script to add a button).
I hate the slow UI of Copilot, so i translate requests from a different GPT interface.