And then every snippet ever created with that trained data would have to include an acknowledgement for every repository included in the training set.

The LICENSE file would be longer than the rest of the code.

(FWIW, I agree with you theoretically, but practically it's hard to get your head around what the ramifications of that would mean)

Many permissive licenses (including Apache 2.0) require attribution.

If Joe Bag’O’Donuts copies and pastes LGPL code into his own personal repository that has MIT license attached, is it safe for Copilot to train on it?

I’m really of the opinion that MS needs to document the training set and include a high bar for inclusion of additional repos.