[happymellon]

This is the fundamental difference between BSD and GPL.

GPL is is Open Source but for you to benefit, you have to make derivative works GPL.

BSD is Open Source and you can close derivatives so that you benefit and you don't have to share back.

I'm not sure how you don't see how an auto-suggest tool that inserts GPL code into yours is different to BSD.

[EdSchouten]

Does Copilot in its current form auto-inject copyright notices for all BSD licensed code from which output is derived? As far as I know, it does not.

This means that authors of BSD licensed code have their rights violated to the same degree as those of GPL licensed code.

[_ph_]

Yes, technically they do. But if the problem is discovered, it is very easy to comply with the BSD terms by adding proper attribution. With GPLed code it is much more difficult.

[kelnos]

> But if the problem is discovered...

The default shouldn't be "you can violate people's licenses until you notice or get caught". Sure, in practice it probably won't cause anyone legal trouble: it would be really hard for, say, the author of some BSD-licensed code to prove that a user of Copilot actually caused them monetary damages by not providing attribution.

But this isn't really a world I want to live in; we shouldn't treat other people's work with such entitlement. Copilot should be able to trace the provenance of a suggestion it makes, and notify the author of their licensing obligations.

If that's not possible, then Copilot should not exist.

[happymellon]

> If that's not possible, then Copilot should not exist

Github and Microsoft should have trained it on their own code because they could make it public domain. Using other peoples code with various licences and many different requirements is just going to be a complete shitshow.

[Houshalter]

All of these issues have existed since GPT-2, maybe earlier. I remember the AI dungeon guys narrowly avoided a lawsuit from training on copyrighted novels. I find it amusing that the tech community was very pro GPT, but anti-copilot. Now that it affects them personally. Especially ironic because this place is usually very skeptical of copyright matters.

All AIs are trained on copyrighted data scraped from the internet. What you guys want effectively amounts to making most AI illegal. At least outside of big tech companies with large private datasets. Is that the world you prefer to live in?

[happymellon]

Yes, in this case the licence is violated by default, by all code that CoPilot generates.

The point was that you can fix attribution errors pretty simply once they are uncovered, its a lot harder to pull GPL code out of your project to prevent you from having to relicence.

It doesn't sound like it is fit for purpose.