You are correct that the problem is hard, because adding licensing info to training data and subsequently use it is something not yet accomplished by anyone (who at least has spoken publicly about it). It might be the only way out would be to have different training sets according to license, but then you lose the advantage of scale.
On the other hand, I believe it's a little much to ask SFC to do the cutting-edge research and propose solutions to Microsoft. Let's not forget that, when faced with the problem, their chosen path was "completely disregard licenses (for now), ask legal later". Many legal opinions are that there would not be any issue if copies of the input code were not reproduced verbatim; unfortunately this is not the case.
Looking forward to more exciting developments!
I get the impression that this article and it's reaction is largely a lot of people wanting to cry big brother and mostly just shit on Github & Git. IMO it would be more appropriate to look for solutions and post a public letter to github and/or just pursue actual litigation before advocating FOSS projects spend inordinate amount of effort across the board to vacate Github.