LLM-generated code must not be committed without prior written approval by core
NetBSD has introduced guidelines that treat code generated by tools like GitHub Copilot and ChatGPT as “tainted” and require written approval before such code can be committed. Commenters debate whether this is a necessary legal safeguard against copyright and licensing risks—especially GPL contamination—or an unenforceable and overly cautious response to a common development aid. The conversation also explores how AI-assisted coding affects code quality, developer understanding, and long‑term maintainability, contrasting narrow autocomplete-style help with wholesale code generation.
Scope and intent of the NetBSD guideline
- The rule calls LLM-generated code “presumed tainted” and requires prior written approval from project core, explicitly naming tools like Copilot.
- Some read it as just short of a total ban; others stress it’s a guideline for good‑faith contributors, not a policeable prohibition.
- It’s framed by several commenters as primarily about protecting NetBSD from legal trouble, not about denying that AI tools can be useful.
Enforceability and norms
- Many say it’s essentially unenforceable: short snippets, restyled code, or heavily edited AI output are indistinguishable from hand-written code.
- Others argue unenforceability isn’t the point; like other project rules (testing, code familiarity), it sets expectations and gives grounds to act if abuse is discovered.
- Some see it as CYA: if AI‑derived infringing code shows up, NetBSD can show it tried to prevent that.
Licensing and copyright concerns
- Strong concern that LLMs are trained on GPL and source‑available code, then emit code without preserving licenses or attribution.
- Risk: GPL or otherwise incompatible code “bleeds” into a BSD‑licensed tree, echoing historic Unix/BSD legal disputes.
- Proposed mitigations include license‑segregated training (BSD‑only, GPL‑only models), but these don’t exist at comparable quality yet.
- Debate over whether small snippets are even copyrightable and how this compares to humans reusing patterns from memory.
- Some distinguish humans (who tend to absorb algorithms) from LLMs (which operate directly on copyrightable text and occasionally regurgitate it verbatim).
Code quality and understanding
- Critics fear AI encourages “copy without understanding,” making complex, safety‑critical or kernel code more fragile and harder to maintain.
- Others counter that competent engineers review AI output like any other code; tests and code review remain the key quality gates.
- Several note that reviewing unfamiliar, nontrivial AI code can be slower than writing it, and tests can’t capture maintainability or design clarity.
Perceived usefulness of AI tools
- Many describe concrete, positive uses: boilerplate React components, shell/PowerShell scripts, data parsers, small language translations, ESLint rules, and exploratory prototypes.
- AI is valued for:
- Reducing “yak shaving” and boilerplate.
- Acting as an interactive doc/learning aid.
- Helping in unfamiliar ecosystems or niche one‑off tasks.
- Others report poor reliability in less common languages or complex domains, requiring heavy rewriting and limiting production use.
Broader attitudes and project context
- Some fear such policies make NetBSD look out of touch and may harm its relevance; others say caution is warranted given its licensing ethos.
- Comparison is drawn to past automation (IDEs, static analyzers, autocomplete); supporters see LLMs as another step, critics say they introduce qualitatively new legal and ethical risks.
- Side discussion notes NetBSD still uses CVS (with plans/interest in Mercurial and a Git mirror), illustrating its conservative tooling culture.