Instances don't actually own the copyright to comments. The poster owns the copyright and licenses it to the instance. Which lets the instance use it, but not sublicense to others.

The current assumption made by these companies is that AI training is fair use, and is therefore legal regardless of license. There are still many ongoing court cases over this, but one case was already resolved in favor or the fair use position.

I don't think you can use gpl for anything but code. Creative commons license would be more appropriate.