All public GitHub code was used in training Copilot • Hacker News Board

Re: All public GitHub code was used in training Copilot

#101

Post by Spivak » Thu, Jul 08, 2021, 2:37 PM UTC

mikehearn wrote: post #83

jonfw wrote: post #75

Earlier quoted context omitted.

It would be pretty concerning if people used GPT-3 while they were writing a novel, and it assisted them in plagiarizing a Steven King novel. We already have examples of copilot blatantly plagiarizing code

Right, but that sounds like the bigger issue here is that the model might spit out copyrighted material, not just that it scrapes it. The former seems like a technology problem that Microsoft can solve.

The issue is that not only might the model spit out copyrighted material verbatim (which it is) but that it might also spit out non-obvious derivative works that will get you in legal hot water years down the road.

Re: All public GitHub code was used in training Copilot

#102

Post by unanswered » Thu, Jul 08, 2021, 2:37 PM UTC

carlgreene wrote: post #78

While I’m not trying to lessen the implications of something like this, but didn’t we all agree to them being able to do this when we agreed to their TOS?

Yeah, but I don't think you're allowed to interrupt the circlejerk by pointing that out. Every piece of code pushed to GitHub comes with: an implied licence for GitHub and its users, which is an alternative to any explicit license in the code(!!!) and also a representation that you're authorized to grant such a license. Of course one imagines that in many cases uploaders are not actually authorized to grant such a license, such as if they're uploading something they themselves have received under GPL license, but IANAL.

Re: All public GitHub code was used in training Copilot

#103

Post by lukeplato » Thu, Jul 08, 2021, 2:37 PM UTC

_Understated_ wrote: post #92

Ok, my curiosity has been fired here... I have conjured up two scenarios here: Let's say I use copilot to generate a bunch of code for an app, something substantial, and it regurgitates a load of bits and pieces from many sources it got from GitHub, I'd assume there won't be any attribution in it... it will be as if Copilot made the code itself (I know it sort of does but lets not split hairs!). I'm guessing the prev…

Copilot is a commercial paid service that generates money for Microsoft

Re: All public GitHub code was used in training Copilot

#104

Post by BlueTemplar » Thu, Jul 08, 2021, 2:38 PM UTC

neom wrote: post #87

Curious what the consensus is on how GH should have approached this to avoid such blowback. Best case scenario, they explained in advance on the GH blog they're going to be doing some work on ML and coding, and they'd like people to opt into their profile being read via a flag setting/or put a file in the repo that gives permission like robots.txt? Second best case scenario, same as first but opt out vs opt in, and l…

Code (co)created with Copilot has to follow all the licenses of the source (heh) code. This generally means at the very least automatically including in projects getting help from Copilot a copy of all the licenses involved, and attribution for all the people the code of which Copilot has been trained on.

(Not sure for the cases where there is no license and therefore normal copyright applies, but AFAIK this isn't the case for any code on Github, which automatically gets an open source licence ?

EDIT : Code in public repositories seems to be "forkable" on Github itself but not copyable (to elsewhere). That's some nasty walled garden stuff right there, I wonder how legal that ToS is ? I could see how this could make them to incentivize people to stop using other licenses on Github, to not have to deal with this license mess... EEE yet again ?)

Re: All public GitHub code was used in training Copilot

#105

Post by ghoward » Thu, Jul 08, 2021, 2:38 PM UTC

richardwhiuk wrote: post #90

ghoward wrote: post #70

This is why I relicensed my code [1] yesterday to a license I wrote [2], which is designed to poison the well for machine learning. [1]: https://gavinhoward.com/2021/07/poisoning-github-copilot-and... [2]: https://yzena.com/yzena-network-license/

If it's allowed by fair use, your license is irrelevant. If it's not, your license doesn't matter.

In my blog post, I talk about how training is fair use, but we don't know about distributing the output. These licenses, even if they don't work, are designed to poison the well by putting enough doubt into companies' minds that they would not want to use Copilot if it has been trained with my relicensed code.

Re: All public GitHub code was used in training Copilot

#106

Post by CyberRabbi » Thu, Jul 08, 2021, 2:38 PM UTC

skinkestek wrote: post #56

Yoric wrote: post #13

Out of curiosity, how do we define license violation in that case? I, as a human being, have trained by reading code, much of which is covered by licenses that are somehow not compatible with code I'm writing. Am I violating licenses? Asking seriously. It's really unclear to me where law and/or ethics put the boundaries. Also, I'd guess it's probably country dependent.

> I, as a human being, have trained by reading code, much of which is covered by licenses that are somehow not compatible with code I'm writing. Am I violating licenses? As someone who has taught students in ICT a quick rule of thumb was that I picked a piece of text that I suspected, wrapped it in doublequotes and put it into a search engine. 9/10 times - possibly more - of the times I had that feeling it was true.…

It’s not AI it is ML. GPT-3 is a very large ML model. It does not reason. It’s a statistical machine.

Re: All public GitHub code was used in training Copilot

#107

Post by kevincox » Thu, Jul 08, 2021, 2:39 PM UTC

saint_abroad wrote: post #23

0x0 wrote: post #7

To the people arguing it's "fair use" to use this for training an ML network. Where do you draw the line? What if you train an "ML network" with one or two inputs... so that they almost always "generate" exact copies of the inputs? Five inputs..? Ten? A thousand? A million?

> Where do you draw the line? My simplistic view is that the following is legally equivalent: input -> ai network -> output input -> huffman coding -> output So, whilst: * compressing and decompressing a copyright work is permissible; * output and weights are deterministic transformations of the inputs; thus: * not eligible for copyright (lacking creativity); and * are derivative works of the inputs;

> output and weights are deterministic transformations of the inputs;

That may be true but I fail to see how any process that produces the same content that was input into it somehow strips the license. If the generated code is novel, then there is no copyright and it is just the output of the tool. If the code is a copy, but non-creative (example a trivial function) then it isn't covered by copyright in the source anyways, so the output is not protected by copyright either. However if the output is a copy and creative I don't think it matters how complicated your copying process was. What matters is that the code was copied and you need to obey copyright.

Again, I don't think that novel code generated from being trained on copyrighted code is the problem. I think it is just the verbatim (or minimally transformed) copying that is the issue.

Re: All public GitHub code was used in training Copilot

#108

Post by tazjin » Thu, Jul 08, 2021, 2:40 PM UTC

neom wrote: post #100

tazjin wrote: post #91

Earlier quoted context omitted.

> Second best case scenario Not really, consider for example repositories mirrored to Github. It seems unclear who has the rights to grant this permission anyways (with free software licenses). Probably the copyright holder? Who that is might also be complicated.

Good point, I would have thought GH requires you to agree in some TOS that you have permission to put the code on GH (but I don't know)? If so, could that point be put aside? (I'm not a software engineer so sorry if that made no sense. Super curious about the whole codepilot thing from a business and community perspective)

> that you have permission to put the code on GH

This is the complicated bit: All open-source licenses grant you permission to redistribute the code (usually with stipulations like having to include the license), so you are almost always allowed to upload the code to Github.

What it doesn't mean however is that you're the copyright holder of that code, you're merely redistributing work that somebody else has ownership of.

So who gets to decide what Github is allowed to do with it?

I expect this will end up in courts and we won't get a definite answer before that.

Re: All public GitHub code was used in training Copilot

#109

Post by Spivak » Thu, Jul 08, 2021, 2:40 PM UTC

CyberRabbi wrote: post #96

Tenoke wrote: post #74

Earlier quoted context omitted.

Yes, it would stiffle NLP research immensely and we likely wouldn't see anything better than gpt3 for years if such restrictions are put in place.

You’re free to privately research with this data but commercializing other people’s work using ML is theft. Edit: commercializing of the derived work is one explicit consideration used by US law in making a fair use determination. That said, even if it weren’t commercialized it may still be infringement and I believe it is.

Commercializing isn't really the issue, it's still copyright infringement even if you release it for free (i.e. piracy) -- it's unauthorized redistribution (i.e. copying).

Re: All public GitHub code was used in training Copilot

#110

Post by cool_dude85 » Thu, Jul 08, 2021, 2:40 PM UTC

jensensbutton wrote: post #98

Wouldn't it be the people publishing code written with Copilot that (potentially) violate any licenses? It doesn't seem to be that the tool violates anything, though it may put the _user_ at risk of violating something. Like, don't use it if you're worried about violating licenses, but I don't see how Microsoft could get in trouble for the tool. It doesn't write and publish code by itself.

Sorry, we built this tool for you that auto violates licenses. Sure, we're owned by a huge megacorp with billions of dollars, but it's your responsibility to confirm - and yes, we recognize it's impossible to confirm - that what you release using our tool isn't violating the license.

In short, github gets to make the license violator bot and push the violations off onto the small fry who actually use it? No thanks.