All they could do is filter by the LICENSE file in the repo.
Unfortunately for them, by law copyright and license are determined by the authors and merely represented by a LICENSE file, which could be lying about both.
The court isn't going to accept that excuse when this goes to trial.
And you can have multiple licenses in the same repository, folders with copyright exceptions, etc.
It's hard enough for us human to find our way in this mess, I've little hope for an AI.
But maybe it's just the first step. The final step being able to sell an AI that understands Copyright management. I'm sure there is a big market for that.
I feel like a few guidelines and standards could help simplify a baseline process:
1) Require each repository to opt-in to be learned from.
2) Require any source file used for learning to have an SPDX license heading.
3) Have a list of approved permissive licenses to avoid any proprietary or copyleft arguments.
Using SPDX headings as the explicit guide would solve the problem of different code content using a different license within a project. An example being QtWayland: the client pieces are Proprietary/LGPL/GPL, whereas the compositor parts are Proprietary/GPL. That's not something you'd know from the license files at the root of the project (and post-6.3 they use SPDX instead of the prior license template heading).
Granted, this doesn't solve the problem of the chain of trust (is the individual publishing the code truly the copyright owner), but I think it would be a basic start for a program like this. The opt-in nature would make things... difficult, but I think that's a fair trade-off for something like this.
All they could do is filter by the LICENSE file in the repo.
Unfortunately for them, by law copyright and license are determined by the authors and merely represented by a LICENSE file, which could be lying about both.
The court isn't going to accept that excuse when this goes to trial.