Judge dismisses majority of GitHub Copilot copyright claims

Original Article ↗ Hacker News Discussion ↗

Scope of the ruling

  • Commenters note the judge mainly dismissed DMCA “copyright management information” (CMI) claims (17 USC 1202), not all copyright issues.
  • Two claims survive: breach of contract and open‑source license violations; some see these as potentially important but legally weaker.
  • Several emphasize the decision is about output behavior, not definitively about the legality of training on copyrighted code.

Reproduction vs training-time infringement

  • Many argue copyright law is primarily about unauthorized duplication, not merely reading or accessing works.
  • One view: Copilot rarely emits memorized code in “benign” situations, so plaintiffs struggled to show specific, infringing reproductions with removed CMI.
  • Counter‑view: even rare verbatim regurgitation matters; if a model can output copyrighted code, both provider and user may face infringement claims.

Liability: who is responsible?

  • One camp: machines have no agency; the human who copies model output into a product is the infringer, analogous to copying from Stack Overflow.
  • Another camp: the operator (e.g. Copilot provider) is distributing copies on request, similar to Napster or other services facilitating mass infringement.
  • Some expect corporate tools will add second‑layer scanners to flag outputs that match known copyrighted code.

AI as “copyright laundering”

  • Strong worry that LLMs let companies “wash” open‑source and GPL code into proprietary products, selling assistance built on unpaid community labor.
  • Others argue someone intent on stealing code can already just download it; using an LLM is a roundabout, weak “loophole”.

Open source, centralization, and power

  • Several express betrayal: open‑source contributions now fuel closed, capital‑intensive AI services they can’t replicate or audit.
  • There’s concern that AI breaks copyright only to re‑centralize control in large firms with the compute to train models.
  • Responses range from quitting open source or self‑hosting git, to deliberately using public domain or “no‑restrictions” licensing to support open models.

Human vs machine learning and clean-room analogies

  • Long debate over whether “training” is analogous to a human reading code:
    • One side says yes; if humans may learn from code, tools helping them should also be allowed.
    • The other side stresses that legal rights attach to humans, not models; scale and automation change the equation.
  • Clean‑room reverse engineering is repeatedly invoked as the traditional, more disciplined way to avoid infringement.

Practical risks and anecdotes

  • Multiple anecdotes describe LLMs outputting near‑identical code (including typos) from older online examples, sometimes from repos with no license.
  • This leads some developers to only use AI for guidance, not for directly copying suggested code.