A place to cache linked articles (think custom and personal wayback machine)
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

hace 3 años
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263
  1. title: Is GitHub a derivative work of GPL’d software?
  2. url: https://drewdevault.com/2021/07/04/Is-GitHub-a-derivative-work.html
  3. hash_url: 0bb6418a58c977e7b142597213f6225d
  4. <p>GitHub recently announced a tool called <a href="https://copilot.github.com">Copilot</a>, a tool which uses machine
  5. learning to provide code suggestions, inciting no small degree of controversy.
  6. One particular facet of the ensuing discussion piques my curiosity: what happens
  7. if the model was trained using software licensed with the GNU General Public
  8. License?</p>
  9. <p><em>Disclaimer: I am the founder of a company which competes with GitHub.</em></p>
  10. <p>The GPL is among a family of licenses considered “copyleft”, which are
  11. characterized by their “viral” nature. In particular, the trait common to
  12. copyleft works is the requirement that “derivative works” are required to
  13. publish their new work under the same terms as the original copyleft license.
  14. Some weak copyleft licenses, like the Mozilla Public License, only apply to any
  15. changes to specific files from the original code. Stronger licenses like the GPL
  16. family affect the broader work that any GPL’d code has been incorporated into.</p>
  17. <p><a href="https://twitter.com/mitsuhiko/status/1410886329924194309">A recent tweet by @mitsuhiko</a> notes that Copilot can be caused to
  18. produce, verbatim, the famous fast inverse square root function from Quake III
  19. Arena: a codebase distributed under the GNU GPL 2.0 license. This raises an
  20. interesting legal question: is the work produced by a machine learning system,
  21. or even the machine learning system itself, a derivative work of the inputs to
  22. the model? <a href="https://twitter.com/eevee/status/1410037309848752128">Another tweet</a> suggests that, if the answer is “no”,
  23. GitHub Copilot can be used as a means of washing the GPL off of code you want to
  24. use without obeying its license. But, what if the answer is “yes”?</p>
  25. <p>I won’t take a position on this question, but I will point out something
  26. interesting: if the answer is “<em>yes</em>, machine learning models create derivative
  27. works of their inputs”, then GitHub may itself now be considered a derivative
  28. work of copyleft software. Consider this statement from GitHub’s blog post on
  29. the subject:</p>
  30. <blockquote>
  31. <p>During GitHub Copilot’s early development, nearly 300 employees used it in
  32. their daily work as part of an internal trial.</p>
  33. </blockquote>
  34. <p>— <a href="https://docs.github.com/en/github/copilot/research-recitation">Albert Ziegler: A first look at rote learning in GitHub Copilot suggestions</a></p>
  35. <p>If 300 GitHub employees used Copilot as part of their daily workflow, they are
  36. likely to have incorporated the output of Copilot into nearly every software
  37. property of GitHub, which provides network services to users. If the model was
  38. trained on software using the GNU Affero General Public License (AGPL), and the
  39. use of this model created a derivative work, this may entitle all GitHub users
  40. to receive a copy of GitHub’s source code under the terms of the AGPL,
  41. effectively forcing GitHub to become an open source project. I’m normally
  42. against GPL enforcement by means of pulling the rug out from underneath someone
  43. who made an honest mistake, but in this case it would certainly be a
  44. fascinating case of comeuppance.</p>
  45. <p>Following the Copilot announcement, many of the ensuing discussions hinted to me
  46. at a broader divide in the technology community with respect to machine
  47. learning. I’ve seen many discussions having to wrestle with philosophical
  48. differences between participants, who give different answers to more fundamental
  49. questions regarding the ethics of machine learning: what rights should be, and
  50. are, afforded to the owners of the content which is incorporated into training
  51. data for machine learning? If I want to publish a work which I <em>don’t</em> want to
  52. be incorporated into a model, or which, if used for a model, would entitle the
  53. public to access to that model, could I? Ought I be allowed to? What if the work
  54. being used is my personal information, collected without my knowledge or
  55. consent? What if the information is used against me, for example in making
  56. lending decisions? What if it’s used against society’s interests at large?</p>
  57. <p>The differences of opinion I’ve seen in the discussions born from this
  58. announcement seem to suggest a substantial divide over machine learning, which
  59. the tech community may have yet to address, or even understand the depth of. I
  60. predict that GitHub Copilot will mark one of several inciting events which start
  61. to rub some of the glamour off of machine learning technology and gets us
  62. thinking about the ethical questions it presents.</p>