Hacker News Reader

Comment by bayindirh | Hacker News Reader

Tangential question: What's your take about Code AIs using GPL and source available code in their training sets, and breaching both licenses?

View on HN

Eschewing the ethics of such an action, my take would be that it happens (and will happen) regardless of license and no one can stop it unless those utilizing such code leave evidence that they disregarded the license.

The models can and will reproduce a lot of their training set given the correct prompt [0]. How you can know you're infringing a license if the model doesn't tell you from which repository it got "inspired" for generating that code?

GitHub was working on a feature which supposedly tells you which repository the snippet you just got is copied from, or IOW, which repositories have similar code, sorted by date. Effectively pushing the blame further on you by making you spend the time you just saved by investigating which repository provided the code you just got from "AI".

Supposedly, if the license is not friendly, you can delete the snippet and write your own version. :)

[0]: https://x.com/docsparse/status/1581461734665367554

Github, and others, have strategies to mitigate culpability. One way, as you note, is to hand off such responsibility to the consumer. Another is to suppress the training set from passing through too close to verbatim.

To be crystal, I am sympathetic with GPL and copyleft movements. I do think we've entered a time where it will be hard for license holders to enforce their rights. The laundering will only get more effective. And, per my original comment, competitive pressure will incentivize models to sail the most piratey tack.

It’s not obvious that training a model on GPL code would constitute a breach of the license.

Given the correct prompt, you can get the training set almost or completely verbatim [0]. Getting a GPL function is enough for GPLs virality, since you effectively lift the code from a GPL codebase and add to your codebase.

Plus, the stack's latest version contains at least one GPL repository which their license tool failed to detect. So it's not something hypothetical in the first place.

[0]: https://x.com/docsparse/status/1581461734665367554

Reply on news.ycombinator.com