I will no longer upload to NEW

Original Article ↗ Hacker News Discussion ↗

Debian’s decision to run new package uploads through an automated large language model review has triggered a broader debate over how AI should be used in open source workflows. Supporters argue that overworked volunteer reviewers need automated help and that LLMs can be effective code-review “interns,” while critics object to being forced to wade through long, unvetted, nondeterministic AI reports and see it as disrespectful of contributors’ time and Debian’s values. Many call for at least an opt‑out or opt‑in mechanism so that AI remains a tool for maintainers rather than an obligation imposed on uploaders.

Use of LLMs in Debian’s NEW Queue

  • Core change: Debian’s DFSG/NEW team is auto-running LLM-based reviews on new packages and sending the raw output to uploaders.
  • Many see this as clashing with Debian’s traditional caution and its “responsible AI” GR, which says automated processes must be overseen and owned by humans.
  • Some contributors say they will stop uploading to NEW unless there is an opt-out.

Consent, Burden, and Respect

  • Main objection is not “AI exists” but being forced to read unreviewed LLM output:
    • Long, nondeterministic “walls of text” (hundreds to thousands of words).
    • Uploaders may be penalized in queue order if they ignore the AI review.
  • Critics argue: if maintainers don’t have time to read the bot’s output, it’s unreasonable to expect contributors to.
  • Several posts frame this as a respect issue and a shift of labor from overworked reviewers to volunteers.

Quality, Determinism, and Comparisons

  • Some compare LLM output to compiler/linter/scripted checks; others say that’s misleading:
    • Deterministic tools are reproducible, debuggable, and stable.
    • LLMs are variable, can hallucinate, and add high-noise text.
  • Others report good experiences with modern models for code review:
    • Claim they catch real bugs and reduce human review load.
    • Emphasize that setup, prompts, and human curation matter.

Governance, Philosophy, and Bottlenecks

  • Some stress Debian’s formal governance and argue this process may violate the spirit of agreed policies.
  • Others insist the real bottleneck is scarce reviewer time, not contributors; any tool that helps maintainers is justified, even if contributors dislike it.

Broader AI Backlash and Slippery Slope

  • Thread widens to AI-heavy workflows elsewhere: job interviews, medical transcription, support bots, UI “AI slop.”
  • Concern about “soft coercion”: AI is “optional,” but opting out is punished via worse service or extra delays.

Suggested Compromises

  • Opt-in or clear opt-out (“No LLM” marker).
  • Separate queues for AI-reviewed vs non-AI-reviewed uploads.
  • Keep LLM as an internal assistant only; humans read the AI output, distill it, and take responsibility before contacting contributors.