Hacker News Reader

Comment by kouteiheika | Hacker News Reader

If training Copilot was copyright infringement then essentially it'd make the majority of useful/interesting machine learning models illegal as they currently exist, and make it impossible for anyone except huge multi-billion corporations to train such models in the future, since only they'd be able to legally license enough data to do so.

Models like GPT-J or DALL-E mini could not legally exist anymore. So sure, please, make this illegal, regress the whole field of machine learning 20 years back, and make it something that only billion dollar corporations can do.

We should be striving to make copyright less draconian, not more.

View on HN

Not necessarily, the issue here isn't that training on public code should be illegal, it's merely that the models trained on such data should be considered derivative works under the licenses which the code is released under. GPL code must also be released under GPL, so Copilot must not charge users to use a model that is trained on and regurgitates GPL code.

> We should be striving to make copyright less draconian, not more. Agreed, and I'm not sure why you think forcing GPT and Copilot to respect open licenses will make them illegal instead of more open.

> Agreed, and I'm not sure why you think forcing GPT and Copilot to respect open licenses will make them illegal instead of more open.

I wasn't talking about Copilot. I was talking about the vast majority of other interesting models. It wouldn't make Copilot itself illegal because it was trained on explicitly licensed data (so it'd make it GPL-licensed), but it would make those other models illegal.

Take for example the GPT-J, which was trained on 825GB of data scraped off the Internet. If we assume the view that a machine learning model is a derivative work of its training data then that makes GPT-J illegal, because it was trained on a bunch of "all rights reserved" data, and there's no legal license under which it could be released. Most interesting models are like that.

> so Copilot must not charge users

That's not what the GPL says.

> Copilot must not charge users to use a model that is trained on and regurgitates GPL code.

I was with you until then. Charging for GPL code is perfectly within the licence as long as you make the source available.

Yea that's my mistake, but not really the point I was trying to make.

> If training Copilot was copyright infringement then essentially it'd make the majority of useful/interesting machine learning models illegal as they currently exist, and make it impossible for anyone except huge multi-billion corporations to train such models in the future, since only they'd be able to legally license enough data to do so.

Regarding the first part, I certainly wouldn't argue that training Copilot is copyright infringement. However, the code it spits out in its current state can in some situations be infringement.

Copyright infringement doesn't happen when you read War & Peace, it's when you take that and reproduce it verbatim or very close to it.

So to that point, your doom & gloom absolutist scenario could not play out if the product of the model was sufficiently different.

We saw Google sued by Oracle for infringement over copying APIs. Now imagine you're a not-Google sized company, are you going to take the chance that Copilot will spit out something that they consider copyrighted?

I think in terms of legal/business risk, it's just too high as it stands now.

> So to that point, your doom & gloom absolutist scenario could not play out if the product of the model was sufficiently different.

It would absolutely play out, because it is impossible to guarantee that such a model will always produce something "sufficiently different". These models are black boxes with billions of parameters. That's how they work. It's just as unrealistic as those politicians pushing through "lawful access to encrypted data" that'd effectively make strong end-to-end encryption illegal. There's no middle ground here. We either accept that such a model might sometimes output a snippet from its training data and benefit from the 99% of times it doesn't, or we can be copyright maximalists and ensure no one benefits. (Except ironically huge corporations like Microsoft, either because they can license the data for training, or because they have sufficiently well funded legal departments.)

> Now imagine you're a not-Google sized company, are you going to take the chance that Copilot will spit out something that they consider copyrighted? > > I think in terms of legal/business risk, it's just too high as it stands now.

This is a fair point, but you can say this about any other interesting model. Is that piece of text generated by GPT-NeoX-20B (which is a fully free and open model trained by essentially hobbyists) illegal to use because it might infringe on someone else's copyright? You don't know. And it was also trained on code from Github. Where are the posts calling for people to sue their authors because they're not respecting the GPL?

Here, I've just tried it and screenshoted it for you, spitting out GPL'd code: https://i.imgur.com/2T4uSJR.png

Again, this is not the Copilot. This is the free GPT-NeoX-20B model that anyone can download. The model's not under GPL, and yet it clearly "contains" GPL'd code. Anything which affects Copilot's legal status will also affect GPT-NeoX-20B, but even more severely since GPT-NeoX was also trained on a ton of "all rights reserved" data. So when you raise your pitchfork at Copilot you should also ask yourself the question - are you fine with also killing projects such as GPT-NeoX, or maybe a more lax copyright law is more beneficial to the society as a whole when it comes to machine learning?

> This is a fair point, but you can say this about any other interesting model. Is that piece of text generated by GPT-NeoX-20B (which is a fully free and open model trained by essentially hobbyists) illegal to use because it might infringe on someone else's copyright?

Again, you misunderstand. Copyright violations don't occur just by using the system, and as I noted, it's not a problem to use ML models. The actual output content is what matters.

In your example, yes, absolutely, if you used some editor feature that spit that code out verbatim, sans license adherence, you will be violating copyright. The issue isn't specific to copilot beyond the fact that Github has created and offered this model trained on source code and begun selling it as a product.

TL;DR: Is it illegal to use it? Of course not. Is it going to get you sued for blindly taking its output, packaging it and selling it? Possibly. If I were trying to manage the risk to my business, I certainly wouldn't be allowing developers to use it.

Reply on news.ycombinator.com