It seems that Copilot could address this issue by searching for matches in its source repositories for the strings it generates, with appropriate criteria, and give the user a link describing the origin of the code, who wrote it, and what the license is for cases where a match length exceeds a threshold. So, you wouldn't just get the Quake fast integer square root routine, you'd get a pointer to the Quake repository and license info from which it came. A separate model could be trained up that would find the closest match in source code repositories. A user could then use Copilot safely, attribute code correctly, and avoid code with incompatible licenses.

This would be a better approach than "shut it down".