This is exactly what I was thinking about. If Copilot is fair use, it means that all proprietary source code, as long as they're publicly available to read, will be free to use as training materials for a hypothetical free and open source machine learning project, which I think would be a good thing. An example is a proprietary program released under a restrictive "source available" license, you can read it but not reuse it under any circumstances (and I believe these projects are already included in Copilot's training data). This is why I said fair use can be a good thing and a ruling to reduce the scope of fair use can potentially be used by proprietary software vendors against the FOSS community.

It would be even better if training from all forms of available proprietary binary code can be fair use, too. It may allow the creation of powerful static binary analysis or code generation tools by learning from essentially all free-to-download proprietary software without copyright restrictions. However, the situation of proprietary binary code is more complicated here. Reverse engineering proprietary binary code is explicitly permitted by the US copyright laws, but the "no reverse engineering" clause in EULA overrides it, and this can be a bad thing. It makes FOSS's fair use right meaningless, meanwhile giving proprietary software vendors a free pass to ignore FOSS licenses.

Thus the outcome is unclear, it may go either way, this is why I said such an issue requires careful considerations.