The model itself is transformative and is considered separate from the outputs.
The outputs will always be a liability for a developer using the tool. So far the outputs are not covered by copyright due to the merger doctrine of the idea-expression distinction.
Not sure I follow. Forgetting how the tool works, it's sold as a tool that outputs usable code for customers. If it's outputting copyright encumbered code (even occasionally), then Microsoft/Gitlab is going to be liable for that.
I don't think an explanation of how it's okay since it's an AI model is going to impress a judge, if the plaintiff shows long passages of verbatim copyrighted code coming out of it.
The only thing I can say is that if there is a tool that is consistently outputting copyright protected works that this burden would not be worth it for most people. But as I have yet to see a single output from my own use or from Twitter that would pass the filtration test I am not worried about my personal liability.
That is, the way GitHub describes Copilot working might pass as fair use. The way it (sometimes) works in real-life will not.