Developer Jakub Jankiewicz argues that code produced with AI help should default to open source or CC0, because frontier models learned from commons code and because machine-only output may not be copyrightable in the US. His November 15, 2026 essay hit the Hacker News front page with 156 points on Hacker News, reopening a fight maintainers already feel in OSS "no AI" catalogues.
This post separates ethical claims, US Copyright Office guidance, and what builders can actually enforce in 2026.
TL;DR
| Claim | Status |
|---|---|
| Models trained heavily on OSS and Stack Exchange | Broadly true as industry practice; per-repo licensing still varies |
| Purely AI output is uncopyrightable in the US | Mostly aligned with Copyright Office and Thaler v. Perlmutter |
| Therefore all AI-assisted repos must be OSS | Normative, not a US legal rule today |
| GPL automatically infects Copilot output | Rejected in Doe v. GitHub discussion cited on HN |
| Companies will comply voluntarily | Unlikely without contracts or audits |
What Jankiewicz argues
The essay makes four moves:
-
Training debt. Frontier models ingested decades of GPL, MIT, and Apache code plus Stack Overflow under Creative Commons terms that require attribution. Proprietary shipping without reciprocity feels like "license laundering" to the author.
-
Copyrightability. Purely machine-generated works lack human authorship. He cites the monkey selfie dispute and the US Copyright Office AI guidance, paralleling Wikimedia's rule that purely AI images belong in the public domain.
-
Global patch with CC0. The UK and other countries grant limited copyright to computer-generated works. Defaulting AI-heavy code to OSS or CC0 avoids a patchwork where some countries allow proprietary enclosure.
-
Vibe-coded minimum effort. When developers do not review or edit model output, the human creative contribution is thin — so the result should be commons.
The tone is prescriptive ethics, not a lawyer's memo.
An open source license badge on a folder, illustrating the AI-generated code open source debate
What US regulators and courts say
The Copyright Office AI initiative published Part 2 on copyrightability on January 29, 2025. It stresses human authorship: prompt-only generation without meaningful selection or arrangement generally does not qualify.
Thaler v. Perlmutter (AI artwork registration) affirmed that wholly AI-generated works without human authorship do not receive copyright. Jankiewicz's chain is directionally right for the US, but it does not automatically void licenses you choose when you do edit.
For software teams, the Office also ran literary works listening sessions that explicitly included software — another signal that repo licensing and training are live policy files, not settled code.
Does training force open source on outputs?
On Hacker News, viraptor summarized Doe v. GitHub (Copilot) to mean LLM output does not automatically inherit OSS notices from training corpora. That breaks the essay's strongest legal compulsion line: ethics can demand reciprocity even when copyright law does not.
Meanwhile g-b-r argued the opposite extreme — any GPL-touched lineage should infect shipping binaries — showing how quickly the thread splinters.
Practical takeaway for builders:
- Your company's CLA and export policy matter more than a blog manifesto.
- Scan dependencies and document human edits the way you already would for Asahi's GPL relicensing fight.
- Do not assume Copilot, Claude Code, or Cursor output is clean-room.
What developers are saying on Hacker News
The thread is argument-rich:
- ubercow13: If AI code cannot be copyrighted, OSS licensing is meaningless for that slice — you already have no exclusive rights to grant.
- delichon: Humans also learned from OSS; should every human-trained developer be forced to open everything?
- anself: Agent loops with real-world testing still embed heavy human effort — ratio matters.
- blagie, an AGPL maintainer, says a clone of his work still felt like his creative content after license-stripped shipping and an LLM rewrite.
- rvz: Unenforceable; economics push toward $0 regeneration instead of waiting for OSS maintenance.
- jacobgold: Shift burden to open weights by law, not open all application code.
None of these comments are legal advice; they show where practitioners disagree.
How this hits Claude Code and agent teams
If you ship with Claude Code, Codex, or Cursor agents:
- Treat prompts and checkpoints as audit logs — who accepted which diff?
- Set a default license on new repos before agents open PRs; read open-weight license families when models are part of the product.
- Mirror Fable ownership norms in prompts when you care about refactors sticking — see Fable 5.2 code ownership guidance.
Agents amplify volume; they do not remove maintainer liability.
When defaulting to CC0 helps anyway
Even if you reject mandatory OSS, CC0 or permissive defaults can help internal glue code:
- One-off migration scripts with no competitive moat.
- Workshop examples tied to explainx.ai courses.
- Reproducibility bundles next to eval harness posts.
Keep proprietary only where human architecture and non-public data are the asset — the same split Menlo's enterprise AI survey shows in spend, not ideology.
The copyleft community reaction
Jankiewicz is not the first voice. LLM reverse-engineering class-action parallels already asked whether training ethics and binary RE law rhyme.
Maintainers who ban AI PRs, catalogued in open-source-says-no, want less slop, not more CC0 dumps. A rule that everything model-touched must be OSS increases noise in GitHub search — ozgung made that point on HN.
A moderate policy template for teams
| Situation | Suggested default |
|---|---|
| Hackathon demo, no moat | MIT or CC0 |
| Product core with human design docs | Proprietary + audit trail |
| Library extracted from agents | Apache-2.0 if patents matter |
| AGPL service backend | Do not let agents rewrite without legal review |
| Gov or healthcare client | No AI in export-controlled modules until counsel signs |
Run license scanners on merged PRs the same way you run secret scanners after Fireworks-class incidents.
International licensing snapshots
Jankiewicz is right that the US is not the whole map:
- United Kingdom can grant copyright to computer-generated literary works where there is no human author — exactly the patchwork his CC0 proposal tries to dodge.
- European Union transparency rules focus on disclosure and documentation, not mandatory OSS on apps — see how we covered Claude moral constitution reporting and watermark expectations on Fable.
- India and APAC enterprise buyers increasingly ask for training-data statements in RFPs, even when copyright does not force open code.
If you ship globally, write a one-page AI authorship memo: what was model-generated, what was edited, which license applies, and where humans signed off.
Enterprise procurement questions to ask legal
- Does our indemnity clause cover training-data allegations?
- Do customer contracts ban GPL contamination from generated patches?
- Are we certifying human authorship on export filings?
- Does insurance cover license laundering claims?
Those questions matter more than blog morality because procurement gates move faster than court precedent.
What would change the debate
Legislation forcing open weights — as jacobgold suggested — would shift incentives more than mandatory OSS on applications.
Until then, Jankiewicz's essay is a moral bumper sticker: reciprocate the commons you trained on. US law today gives you more nuance: human authorship, chosen licenses, and contracts.
Document your human edit ratio, pick a license on purpose, and treat viral essays like this one as inputs to policy, not build-breaking rules.
FAQ overlap with maintainer AI bans
The essay's maximal openness collides with projects that ban AI patches outright. Maintainers want fewer unreviewed megadiffs; Jankiewicz wants more public goods. Both can be true:
- Ban lists target quality and review load, not licensing theory.
- CC0 dumps without tests still waste maintainer time.
A sane contributor rule: disclose AI use, attach test logs, and pick a license matching your employer — even if philosophers wish everything trained on GPL became GPL again.
Classroom and workshop framing
If you teach with Claude Code workshops, use the essay as a debate prompt, not homework law. Students should leave able to:
- Read Copyright Office Part 2 summaries.
- Explain why Doe v. GitHub matters to Copilot users.
- Draft a team policy in ten lines.
That skill matters more than winning an internet argument about RMS and the GNU manifesto's moral descendants.
Quick answers for searchers
Should all AI code be open source? US law does not require it today; Jankiewicz argues it should be ethical default.
Is Copilot output GPL? HN commenters citing Doe v. GitHub say not automatically — still check your edits.
Is AI code public domain? Pure machine output may lack US copyright; mixed human work can still be protected.
Related reading
- Open source says no: AI ban catalogues
- Asahi Linux GPL relicensing and LLM laundering
- LLM reverse engineering vs OSS licenses
- Open-weight AI model licenses explained
- Claude Fable 5.2 code ownership prompting
- Replica Skill and cloning apps
This article is educational, not legal advice. US Copyright Office materials, the Jankiewicz essay, and Hacker News comments were reviewed on November 15, 2026.