3 ms·

While I understand the sentiment wasn’t Copilot trained on code not only hosted on GitHub, but found all over the Internet? Which means hosting your code yourse

by M4v3R 5y ago

While I understand the sentiment wasn’t Copilot trained on code not only hosted on GitHub, but found all over the Internet? Which means hosting your code yourself would not prevent GitHub from using it to train Copilot. That raises an interesting question though - how do you opt out? Is there even a way to do it?

bryanrasmussen 5y ago
robots.txt, or a copyright notice saying the code can't be used to train AI which bots will ignore and open their corporate masters to liability. on edit: fixed typo
- woodruffw 5y ago
  This kind of learning needs to be opt-in, not opt-out. I would also be extremely surprised if most open source copyright holders didn't already expect their licensing terms to protect against this kind of code/authorship laundering. Speaking individually, I know that it certainly surprised me to hear that GitHub thinks that it's probably okay to regurgitate entire fragments of the training set without preserving the license.
- ezoe 5y ago
  Bad news for you. Japanese copyright law, article 47-7 explicitly allow using copyrightable works for data analysis by means of a computer(including recording a derivative work created by adaptation) It should be considered as fair-use of USA except we don't use Common Law system so we explicitly state what exempt from the copyright protection.
  - bryanrasmussen 5y ago
    Thanks for the bad news! Not glad to hear it but glad to know something I didn't. That said - so they would be able to sell some things in Japan that they couldn't other places.
    - amelius 5y ago
      What if a Japanese software company uses it to write software, which is then sold in the US? It's still copyright laundering, if you ask me.
- blihp 5y ago
  robots.txt is a convention for those who want to be good 'web citizens' rather than legally binding. It does absolutely nothing to stop someone who ignores your wishes. For example, there are tons of bots that ignore robots.txt entirely or even go straight for the thing (i.e. 'hey, thanks for telling us where to look!') you're telling them to avoid in robots.txt. While copyright is a mechanism that can be used if you can make the case, and have the means, it will only work for entities that have something to lose and are within a jurisdiction where it matters.
rvense 5y ago
I'm not surprised. I imagine all images on the internet are used to train image classifiers as well. It's a shitty future, but it's the one we have.
- hojjat12000 5y ago
  Researchers in our lab created a huge dataset of facial expressions from images on the web, annotated it and published the URLs to the images and the annotations for research but made sure to search only for images with proper licenses. I don't think that you are allowed to just go download any old image and train on it. I understand the many many people do it, but it's not legal (as far as I know, please correct me if I'm wrong).
  - sillysaurusx 5y ago
    > I don't think that you are allowed to just go download any old image and train on it. My understanding as a two-year student of ML is that you are allowed in the US to go download any old image, train on it, and then release the model as long as the outputs are "sufficiently transformative." That last phrase is the key part, and has never been tested in court. It's entirely possible that either I'm mistaken here, or that the courts will soon say that I am mistaken here. https://www.youtube.com/watch?v=4FA_gt9w28o&ab_channel=guavaball https://www.youtube.com/watch?v=4FA_gt9w28o&ab_channel=guava...
    - saurik 5y ago
      To be clear: "transformative" not meaning merely "altered" but really meaning "repurposed"; if the new work is something people could feasibly use instead of the old work (harming the author's original market), it isn't "transformative".
      - sillysaurusx 5y ago
        Yes. For example, arfa ran into this question when launching https://thisfursonadoesnotexist.com/ https://thisfursonadoesnotexist.com/. Lots of furry artists had exactly the same concerns with his work there, but that work is decisively transformative. Copilot seems ... well, less transformative. I'm still not sure how to feel.
bsd44 5y ago
I would like to know this too. I understand that GitHub is a private company and you have to accept their T&C, but surely they aren't allowed to use source code found elsewhere on the internet to train their ML models without asking for permission first unless it's a B2B cooperation such as with Stackoverflow.
- lacker 5y ago
  According to the discussion at this link, you do not need permission to use copyrighted data to train AI models. Copyright prevents you from copying data, it doesn't prevent you from learning from it. https://twitter.com/luis_in_brief/status/1410985742268911631?s=20 https://twitter.com/luis_in_brief/status/1410985742268911631...
  - marcosdumay 5y ago
    To train your model, yeah, probably ok. But I don't think anybody will see people using the duplicated code that AI insert on your codebase the same way.
- lloydatkinson 5y ago
  Oh man I can’t imagine the consequences for certain languages and frameworks if it uses SO answers though. Imagine if it trained in all the dumb and ancient answers like “how do I get the length of a string in javascript” and took the first accepted answer of “use jquery”
  - bsd44 5y ago
    This raises an issue of trolling. What prevents developers to generate "inappropriate" code to feed it to this algorithm the same way they did with the Microsoft Chat bot for example? That will surely reflect on the quality of code generated by this AI system and therefore the stability and security of applications built.
    - kingofclams 5y ago
brobdingnagians 5y ago
I guess it goes back to closed source / trade secrets territory. If you have something you really don't want stolen, it is safer to never expose it and never trust that the law will fairly protect you. The irony is that copilot won't suggest its own source code, just everyone else's. It is open source without the benefits.
- axismundi 5y ago
  Smells like Microsoft
  - yumraj 5y ago
    More like OpenAI
perryizgr8 5y ago
> how do you opt out? Is there even a way to do it? Yeah, don't post your code in the public for everyone to read. If I am musician, and I play my song publicly, people will hum it if they like it. I can't do anything to stop that, except not play the song to anyone.
- M4v3R 5y ago
  It’s a different thing when random people start humming or even singing your song in private, and an entirely different thing when a big corporation uses your work to train a system that will at some points generate parts of your works and will make them millions of dollars of your and everyone else’s code without attributing anyone in any way. Plus, to be honest I’m not even sure whether I’m for or against that, I am just wondering if one was against it but still wants to do open source work do they have any recourse?

by M4v3R 5y ago

bryanrasmussen 5y ago

robots.txt, or a copyright notice saying the code can't be used to train AI which bots will ignore and open their corporate masters to liability. on edit: fixed typo

woodruffw 5y ago
This kind of learning needs to be opt-in, not opt-out. I would also be extremely surprised if most open source copyright holders didn't already expect their licensing terms to protect against this kind of code/authorship laundering. Speaking individually, I know that it certainly surprised me to hear that GitHub thinks that it's probably okay to regurgitate entire fragments of the training set without preserving the license.
ezoe 5y ago
Bad news for you. Japanese copyright law, article 47-7 explicitly allow using copyrightable works for data analysis by means of a computer(including recording a derivative work created by adaptation) It should be considered as fair-use of USA except we don't use Common Law system so we explicitly state what exempt from the copyright protection.
- bryanrasmussen 5y ago
  Thanks for the bad news! Not glad to hear it but glad to know something I didn't. That said - so they would be able to sell some things in Japan that they couldn't other places.
  - amelius 5y ago
    What if a Japanese software company uses it to write software, which is then sold in the US? It's still copyright laundering, if you ask me.
blihp 5y ago
robots.txt is a convention for those who want to be good 'web citizens' rather than legally binding. It does absolutely nothing to stop someone who ignores your wishes. For example, there are tons of bots that ignore robots.txt entirely or even go straight for the thing (i.e. 'hey, thanks for telling us where to look!') you're telling them to avoid in robots.txt. While copyright is a mechanism that can be used if you can make the case, and have the means, it will only work for entities that have something to lose and are within a jurisdiction where it matters.

rvense 5y ago

I'm not surprised. I imagine all images on the internet are used to train image classifiers as well. It's a shitty future, but it's the one we have.

hojjat12000 5y ago
Researchers in our lab created a huge dataset of facial expressions from images on the web, annotated it and published the URLs to the images and the annotations for research but made sure to search only for images with proper licenses. I don't think that you are allowed to just go download any old image and train on it. I understand the many many people do it, but it's not legal (as far as I know, please correct me if I'm wrong).
- sillysaurusx 5y ago
  > I don't think that you are allowed to just go download any old image and train on it. My understanding as a two-year student of ML is that you are allowed in the US to go download any old image, train on it, and then release the model as long as the outputs are "sufficiently transformative." That last phrase is the key part, and has never been tested in court. It's entirely possible that either I'm mistaken here, or that the courts will soon say that I am mistaken here. https://www.youtube.com/watch?v=4FA_gt9w28o&ab_channel=guavaball https://www.youtube.com/watch?v=4FA_gt9w28o&ab_channel=guava...
  - saurik 5y ago
    To be clear: "transformative" not meaning merely "altered" but really meaning "repurposed"; if the new work is something people could feasibly use instead of the old work (harming the author's original market), it isn't "transformative".
    - sillysaurusx 5y ago
      Yes. For example, arfa ran into this question when launching https://thisfursonadoesnotexist.com/ https://thisfursonadoesnotexist.com/. Lots of furry artists had exactly the same concerns with his work there, but that work is decisively transformative. Copilot seems ... well, less transformative. I'm still not sure how to feel.

bsd44 5y ago

I would like to know this too. I understand that GitHub is a private company and you have to accept their T&C, but surely they aren't allowed to use source code found elsewhere on the internet to train their ML models without asking for permission first unless it's a B2B cooperation such as with Stackoverflow.

lacker 5y ago
According to the discussion at this link, you do not need permission to use copyrighted data to train AI models. Copyright prevents you from copying data, it doesn't prevent you from learning from it. https://twitter.com/luis_in_brief/status/1410985742268911631?s=20 https://twitter.com/luis_in_brief/status/1410985742268911631...
- marcosdumay 5y ago
  To train your model, yeah, probably ok. But I don't think anybody will see people using the duplicated code that AI insert on your codebase the same way.
lloydatkinson 5y ago
Oh man I can’t imagine the consequences for certain languages and frameworks if it uses SO answers though. Imagine if it trained in all the dumb and ancient answers like “how do I get the length of a string in javascript” and took the first accepted answer of “use jquery”
- bsd44 5y ago
  This raises an issue of trolling. What prevents developers to generate "inappropriate" code to feed it to this algorithm the same way they did with the Microsoft Chat bot for example? That will surely reflect on the quality of code generated by this AI system and therefore the stability and security of applications built.
  - kingofclams 5y ago

brobdingnagians 5y ago

I guess it goes back to closed source / trade secrets territory. If you have something you really don't want stolen, it is safer to never expose it and never trust that the law will fairly protect you. The irony is that copilot won't suggest its own source code, just everyone else's. It is open source without the benefits.

axismundi 5y ago
Smells like Microsoft
- yumraj 5y ago
  More like OpenAI

perryizgr8 5y ago

> how do you opt out? Is there even a way to do it? Yeah, don't post your code in the public for everyone to read. If I am musician, and I play my song publicly, people will hum it if they like it. I can't do anything to stop that, except not play the song to anyone.

M4v3R 5y ago
It’s a different thing when random people start humming or even singing your song in private, and an entirely different thing when a big corporation uses your work to train a system that will at some points generate parts of your works and will make them millions of dollars of your and everyone else’s code without attributing anyone in any way. Plus, to be honest I’m not even sure whether I’m for or against that, I am just wondering if one was against it but still wants to do open source work do they have any recourse?