About three months after Debian voted to allow generative AI into its workflow, the project is arguing over what that permission looks like on the ground. LLM-generated reviews of packages in Debian's NEW queue have split discussion on the debian-devel mailing list, according to LWN.net, reopening a question the project believed it had settled with a general-resolution vote in August.
From a vote to a workflow problem
Debian contributors spent years arguing over generative AI before the dispute finally reached a ballot. The project's vote record shows eight options were put forward; the fifth carried, "Responsible Use of Generative AI." The title sets a narrower frame than outright endorsement: it governs how the tools may be used, not whether they are welcome.
With that question answered, attention has moved to where the tools actually touch Debian's machinery. The NEW queue is the staging area every incoming package must pass through before entering the archive, and it is where reviewers assess licensing, policy compliance and conformance with the Debian Free Software Guidelines. It is also where LLM-generated review comments have drawn objections from developers, per LWN's report.
The work overlaps directly with the remit of the DFSG, Licensing & New Packages team — precisely the body responsible for the judgement calls, particularly licence interpretation, that automated output handles least reliably.
The two sides of the queue
The dispute is less about whether the resolution permits these tools than about what responsible use means in a high-volume review queue. Those defending machine-assisted reviews point to a chronic shortage of human reviewer time: the queue is a persistent bottleneck, and they argue that pattern-matching checks — boilerplate licensing headers, policy boilerplate, obvious DFSG conflicts — are exactly the load-bearing but repetitive work that scales poorly with volunteer effort. Automated first passes, on this reading, free specialists for the judgement-heavy cases.
Critics counter that a fluent but wrong review is worse than no review at all. Package acceptance decisions carry long-term archive consequences, and an LLM comment that reads authoritatively but misstates a licence term forces a human expert to spend scarce attention disproving it. The burden of catching machine error, in other words, lands on the same people the tool was meant to relieve. Neither position has carried the day: debian-devel remains divided, and the resolution's general language leaves the operational rules to be negotiated in public.
Why it matters beyond Debian
For maintainers of any open-source project, this is an early and unusually clean case study. A policy resolution settles the principle; the operational questions remain. Who carries responsibility when an LLM review is wrong? How much reviewer time goes into triaging machine-generated comments? And how does a project tell careful human judgement from fluent automated text once the volume scales?
Debian makes a visible test case because its review processes are public, its standards are written down, and its contributors have already voted on the underlying principle. What the project now faces is the gap between a resolution permitting responsible use and the daily reality of queue hygiene once that permission is exercised.
There is a structural tension here that extends well beyond one distribution: generative AI gravitates toward exactly the high-volume, pattern-matching review work where a plausible-but-wrong output is most expensive.
What it means for Hong Kong readers
The immediate stakes for Hong Kong are indirect but practical. Ubuntu and the wider family of Debian-derived distributions inherit much of their package base from Debian's archive, so how the NEW queue adjudicates licence and policy questions eventually reaches every derivative — relevant to local teams running Ubuntu LTS in production who rarely track debian-devel. For organisations drafting their own internal AI-usage policies, Debian's experience is a useful template: a high-level approval resolves the principle in a single vote, but the hard governance work — attribution, disclosure, accountability for tool output — only begins afterwards. Hong Kong IT teams adopting LLM-assisted review or code-audit workflows can expect the same sequence.
LWN's full report is available to subscribers. Background is documented in the resolution text at debian.org/vote/2026/vote_002 and on the DFSG, Licensing & New Packages team page in the Debian wiki.
Debian 表決容許生成式 AI 加入其工作流程約三個月後，該項目正就這項許可在實際操作層面應如何落實展開爭論。據 LWN.net 報道，以 LLM 生成的審核意見處理 Debian NEW queue 中的套件，已在 debian-devel 郵件列表引發分歧，令項目原以為在八月以 general-resolution 表決解決的問題再度浮上檯面。
由一場投票變成工作流程難題
Debian 貢獻者就生成式 AI 爭辯多年，爭議最終才進入投票程序。項目投票紀錄顯示共有八項方案提出，第五項獲通過，題為「生成式 AI 的負責任使用」（Responsible Use of Generative AI）。標題所定下的框架比全面背書更為狹窄：它規管的是這些工具可如何使用，而非它們是否受歡迎。
這個問題既有答案，焦點便轉移到工具真正接觸 Debian 機制的地方。NEW queue 是每一個新套件在進入歸檔庫前必須經過的暫存區，審核員在此評估授權條款、政策合規性以及是否符合 Debian Free Software Guidelines（DFSG）。根據 LWN 報道，LLM 生成的審核意見同樣在此處引起開發者反對。
這項工作與 DFSG, Licensing & New Packages 團隊的職責範圍直接重疊——而該團隊正是負責作出判斷的機構，尤其涉及授權條款解讀，恰恰是自動化產出最不可靠的範疇。
queue 的兩面
爭議與其說是決議是否容許這些工具，不如說是在高流量審核 queue 中，何謂「負責任使用」。支持機器輔助審核的一方指出，人手審核時間長期不足：queue 是持續存在的瓶頸，他們認為模式匹配式的檢查——授權聲明的制式開頭、政策範本、明顯的 DFSG 衝突——正是關鍵但重複性高、難以靠義工投入擴展的工作。按此理解，自動化初審可令專家騰出時間處理判斷成分較重的個案。
批評者則反駁，一份流暢但錯誤的審核比沒有審核更糟。套件接納決定對歸檔庫有長遠影響，而一份讀來權威卻錯誤陳述授權條款的 LLM 意見，會迫使人類專家耗用寶貴注意力去推翻它。換言之，揪出機器錯誤的負擔，最終落在本應由該工具紓緩的同一群人身上。雙方均未佔上風：debian-devel 仍存分歧，而決議的概括性措辭令操作規則有待公開談判。
為何對 Debian 以外同樣重要
對任何開源項目的維護者而言，這是一個早期且異常清晰的案例研究。政策決議解決了原則問題；操作問題依然存在。當 LLM 審核出錯，誰須負責？審核員要花多少時間分類處理機器生成的意見？當數量擴大後，項目又如何分辨謹慎的人類判斷與流暢的自動化文字？
Debian 成為明顯的測試案例，在於其審核流程公開、標準寫成文字，而貢獻者亦已就底層原則投票。項目現下面對的，是一份容許負責任使用的決議，與該許可一旦行使後 queue 日常維護現實之間的落差。
這裡存在一個結構性張力，範疇遠超單一發行版：生成式 AI 恰好傾向投入那些高流量、模式匹配的審核工作，而在這些工作中，一份貌似合理卻錯誤的產出代價最高。
對香港讀者的意義
對香港而言，即時影響屬間接但實際。Ubuntu 及整個 Debian 衍生發行版家族的套件庫大部分繼承自 Debian 歸檔庫，因此 NEW queue 如何裁決授權與政策問題，最終會傳遞至每一個衍生版本——這與本地在生產環境使用 Ubuntu LTS、卻很少跟蹤 debian-devel 的團隊息息相關。對正擬訂自身內部 AI 使用政策的機構，Debian 的經驗是一份有用的範本：高層面批准一次投票便可解決原則問題，但艱難的治理工作——署名、披露、對工具產出的問責——其後才真正開始。香港 IT 團隊採用 LLM 輔助審核或 code-audit 工作流程，可預期同樣的次序。
LWN 全文報告供訂閱者閱覽。背景資料載於 debian.org/vote/2026/vote_002 的決議文本，以及 Debian wiki 中 DFSG, Licensing & New Packages 團隊頁面。