Reading the FAQ it seems to come down "people copy/paste from publicly available code anyway so who cares. We'll look into attribution in the future maybe".
I've never read an FAQ that avoids it's own questions this much. They say "permissive licenses", they could've just listed what licenses were deemed acceptable by their scraper. Licenses like WTFPL, MIT-0 or CC0 are perfectly usable for ML because they require nothing in exchange for the code. There are projects out there using these licenses, though most will require attribution at the very least.
Microsoft went full "licenses don't apply to machine learning, sue us", just like most AI companies when it comes to things like copyright, but I don't think claiming to care about licenses and then refusing to name any permissable licenses for your model is much better. At least MS doesn't lie about their intentions.