So the code suggestion comes from the same data trained for the public version, which could include GPL code or have other issues?

I doubt any company would use this in their production code. Internal tools, maybe.