5 ms·
It was trained on all public repos, and only public repos. They did not pay attention to licensing.
by NoraCodes 4y ago
It was trained on all public repos, and only public repos. They did not pay attention to licensing.
- Thev00d00 4y agoWow, that's crazy. Do we have something official that says as much?
- jacquesm 4y agoIt's infuriating, when you are large enough you can get away with anything when it comes to copyright violations but I remember the crazy pushback on things like format and time shifting against private individuals.
- voxic11 4y agoThose things were ultimately ruled as fair use though which is what Microsoft is claiming here as well.
- jacquesm 4y agoThose things were fair use, what Microsoft is doing is copyright violation pure and simple.
- voxic11 4y agoMaybe, but then every large AI project is also committing copyright violations because as Microsoft notes this is currently a common practice in the AI research community.
- jacquesm 4y agoOther parties doing is no excuse imo. Microsoft has been super litigious in the past when it came to copyright violation starting all the way back with Bill Gates' letter in Byte magazine about those pesky pirates. To see them do this makes pirating MS software fair game from here on. They could have asked nicely, instead they just took.
- searchableguy 4y ago> Other parties doing is no excuse imo It is. Laws are adapted based on widespread technological capabilities and progress. As an example, if it is easy to create real voice or signature using AI models - they should no longer be considered effective evidence for contractual reason instead of enforcing that it is illegal to forge it. That is not going to work. Past shouldn't dictate what we allow tomorrow.
- jacquesm 4y agoSorry, but that's not how the law works. Try that excuse the next time you're stopped for speeding and see how well it works.
- searchableguy 4y agoYour example is not good. Speeding is not a technical innovation that require any fundamental change. It is enforced in automated fashion and it is beneficial for the safety of public at large if reasonably implemented. All laws are made in interest of someone. Does the copyright apply to AI models since they are out of scope and weren't widespread when it came into force? Does the proposed benefit in the original law apply in practice? Are they more beneficial than the progress allowed by AI models who use them as training data? Is the copyright law practically enforceable on output generated by AI models?
- jacquesm 4y agoCopyright law is what it is today. Like it or not doesn't really matter. And yes, copyright law is practically enforceable, regardless of how copyright is broken. That's what the Berne Convention is all about. https://en.wikipedia.org/wiki/Berne_Convention https://en.wikipedia.org/wiki/Berne_Convention Copyright is what FOSS depends on. For Microsoft to shit all over GitHub contributors rights is despicable.
- Nanana909 4y agoIf Copilot reads copyrighted code to learn, and it‘s copyright violation… Does the same apply to a human? Do we now define copyright violation differently for computers? I don‘t know the perfect answer here. But I‘m not so sure we should have standards that change depending on if a program is doing it or a human is doing it. Perhaps a bad standard to begin with. I do tend to learn towards thinking „company uses publicly available, open source code in product“ is somewhat of a nothing-burger though.
- maleldil 4y agoFraming this as learning is unhelpful. Copilot isn't just learning. It can reproduce code it saw during training verbatim. It will even reproduce the GPLv3 licence itself. If you copy-paste code from a GPLv3-licenced sourced without attribution, it's a copyright violation. Why should it be different for Copilot? EDIT: After more reading on the subject, I'm willing to accept that copyright infringement is unlikely here. This link [1] was the one I found most convincing. However, I would still shift the goalposts and look at this ethically, and I still think it's wrong that Microsoft is profiting from code with licences like GPLv3. This is a whole other topic, though. [1] https://www.technollama.co.uk/is-githubs-copilot-potentially-infringing-copyright https://www.technollama.co.uk/is-githubs-copilot-potentially...
- temp-dude-87844 4y agoThis should absolutely be litigated to stop the blatant laundering of copyright under the guise of fair use. Copilot's API is surfacing snippets of work without licensing information attached alongside. It can be shown in discovery that Copilot does access the origin work. The sooner this is slapped down, the sooner we can avoid addressing the even more troubling question that exists today: is someone who used Copilot to throw together a bunch of code infringing copyright of works where those portions originate? This is a complex problem with no satisfying conclusions... how could one be violating copyright if they never accessed the 'copied' work to copy? Copyrights aren't patents. Infringement requires copying. Using Copilot launders the user's awareness of the origin works, yet making the Copilot users liable for widespread "accidental" copying would be troubling.
- jacquesm 4y agoIt's not complex at all: if you use Copilot to generate code for you you are engaging in copyright infringement. That you got the code from and entity that stole it somewhere else doesn't really matter. Generative models should respect copyright for their sources, and using a generative model to create new works that you intend to claim copyright on is stupid: someone may well show up one day with ironclad proof that you used their code without permission.
- tpmoney 4y agoA reminder that copyright infringement vs fair use is in part dependent on the amount of the copyrighted material that’s being used, the nature of that use and the transformativeness of the infringing work. Just because co-pilot suggests code snippets that can be found in a copyrighted work does not mean that the resulting produced product is in fact an infringement of that copyright. Also a reminder that outside the copilot debate, the online rights movement has largely been pushing for scraping, deep linking and transforming scrapped data to not be considered copyright infringement, regardless of any TOS on the site being scraped. To me, co pilot is a exactly that, a scraper that has scraped public websites and is now presenting me the scraped data in an alternative and often transformed form. It’s my responsibility as a developer to ensure that my released product complies with applicable copyright law, but copilot and the use thereof is not in and of itself copyright infringement. That a tool can be used to create infringing work or infringe on copyright in general is no more a valid argument against co pilot than it is against CD burners, de-drm tools, vcrs, kodi or plex, scanners or any number of day to day items that have the ability to infringe copyright if the user uses it for that purpose.
- avg_dev 4y agoWhat is format and time shifting? I am wondering if you mean like code formatting and reordering history, but I’m not really sure. Thanks.
- voxic11 4y agoYou can read between the lines > GitHub Copilot is trained on billions of lines of public code. > In one instance, GitHub Copilot suggested starting an empty file with something it had even seen more than a whopping 700,000 different times during training–that was the GNU General Public License. https://github.blog/2021-06-30-github-copilot-research-recitation/ https://github.blog/2021-06-30-github-copilot-research-recit... This indicates that they are training it on github's public repositories and at the very least including 700,000 GPL licensed projects or code files. Since the GPL is one of the most "restrictive" open source licenses one can assume they are not caring about the licenses much.
- NoraCodes 4y agoI asked, and that's what their spokesperson said via email.