Is AI‑Written Code Considered a Contribution?

Is AI‑Written Code Considered a Contribution?\n\nDebian Developers Argue for Half a Month, Ending with Just Three Words: \"Wait and See\"\n\n\n\nToday, large models are rapidly changing the way software is developed, but the open‑source community is falling into a new dilemma: if code is written by AI, does it still count as a developer’s contribution?\n\nRecently, one of the most important Linux distributions, Debian, launched a heated discussion lasting several weeks over this question.\n\n• Some argue that AI is just a new development tool and does not need heavy restrictions.\n• Others worry that AI will undermine the open‑source community’s mechanism for training newcomers.\n• Still others, from an ethical standpoint, oppose generative AI, claiming these tools are essentially \"plundering the Internet\". \n\nBut after the debate, the Debian community reached a somewhat surprising conclusion: wait and see, postponing any decision for now.\n\n\n\n\n\nA draft \"AI contribution policy\" sparked a community uproar.\n\nThe starting point was a proposal by Debian developer Lucas Nussbaum in mid‑February this year.\n\n\n\nHe suggested that the Debian community discuss whether AI‑generated or AI‑assisted code should be allowed into the Debian project. To that end, he drafted a proposal hoping to clarify Debian’s policy on this issue, and said he would first collect a few days of community feedback before deciding whether to formally submit a resolution.\n\nAccording to the draft, AI‑assisted contributions (code partially or wholly produced by a large language model) are acceptable, but must meet a set of conditions, for example:\n\n(1) If the important part of a contribution comes from an AI tool and has not been manually modified, it must be clearly disclosed.\n(2) The contribution must include a clear statement or machine‑readable tag, such as [AI‑Generated].\n(3) The submitter must fully understand what they are submitting.\n(4) The submitter must take responsibility for the code, including its technical quality, security, license compliance, and actual utility.\n\nFurthermore, the draft stipulates that to avoid serious data‑leakage risks, no non‑public or sensitive project data may be fed into generative AI tools, including private mailing‑list contents and unpublished vulnerability reports.\n\nIn other words, AI can help you write code, but responsibility must remain with the human developer.\n\n\n\nThe first step of the argument: everyone could not even agree on what \"AI\" means.\n\nBut the discussion quickly exposed a more fundamental problem—people have not even reached a consensus on the term \"AI\".\n\nDebian developer Russ Allbery said bluntly that the word \"AI\" has become overly vague; today \"AI\" can refer to almost anything in the universe. Some mean ChatGPT, some mean code‑generation tools, and others even count ordinary automation scripts. He argued that if Debian wants to craft a policy, it must first define the exact object—for example, large language models (LLMs), reinforcement learning, or another specific technology—otherwise the policy will be hard to enforce.\n\n\n\n\"LLMs at least have a relatively clear meaning, whereas ‘AI’ often just stands for whatever the speaker wants it to mean, and even within a single discussion the meaning can shift.\"\n\nAnother developer, Sean Whitton, also advised that if a policy is to be written, the term LLM should be used instead of the vague word AI, and that different AI usage scenarios should be distinguished—code review, prototype generation, or direct production‑code generation. He suggested that voting options should treat these cases separately rather than applying a one‑size‑fits‑all approach.\n\nHowever, unlike those developers, Lucas Nussbaum believes the technical details are not the point; the real question is whether automated tools should be allowed to participate in code generation or analysis.\n\nHe gave an example: in the early days of the Linux kernel, the community argued fiercely over whether the BitKeeper version‑control system should be used; a similar debate arose in the security‑tool area—if a vulnerability scanner is closed source, should the community ignore the flaws it finds?\n\nNussbaum added that taking a completely \"anti‑tool\" stance would make it difficult to draw a clear line.\n\n\n\nCould AI damage the open‑source community’s mechanism for nurturing newcomers?\n\nThe conversation then turned to a deeper issue: might AI erode the way open‑source projects train new contributors?\n\nDeveloper Simon Richter offered an interesting observation: AI agents could, in some respects, replace junior developers. In many cases a junior developer can finish a task under a mentor’s guidance, and the same task can now be performed by AI—but the problem is that AI does not learn.\n\nThat means the mentoring effort invested by the community cannot be turned into new long‑term contributors.\n\nSimon Richter said: ‘The best outcome is that a small bug gets fixed but no new contributor is cultivated; the worst outcome is that the so‑called contributor merely becomes a middleman between the AI and the maintainer.’\n\nHe also warned about the cost of AI tools: if future development increasingly relies on paid AI services, the barrier to entry for ordinary users wanting to contribute to open source could rise.\n\nLucas Nussbaum, however, does not fully agree. He believes Debian will never run out of suitable tasks for newcomers, and AI might even help newcomers tackle more complex problems. He cited a study (co‑authored by an Anthropic employee) which found that human‑AI interaction patterns vary greatly, and different ways of using AI affect learning outcomes in completely different ways.\n\nProminent Linux developer Ted Ts’o also directly countered the anti‑AI argument: some people fear AI will reduce the number of future senior contributors, but if that leads to rejecting AI‑using contributors, that would be self‑harm.\n\n\n\n\n\nEthical concern: AI companies are ‘draining the Internet’\n\nPart of the opposition comes from ethical and copyright angles.\n\nDebian developer Matthew Vernon pointed out that generative‑AI firms have obvious problems when training their models—for example, massive scraping of Internet content, ignoring copyright and open‑source licenses, and directly using others’ intellectual property to train models.\n\nHe said these companies hoover up data with little regard for copyright. Matthew Vernon also mentioned other controversies, including AI’s energy consumption, deep‑fake problems, and AI‑generated bogus vulnerability reports—he believes a free‑software project like Debian should clearly oppose such AI tools.\n\nBesides ethics, the copyright issue remains unresolved, mainly concerning the copyright of the training data used for models and the ownership of AI‑generated output. For this, developer Jonathan Dowland suggested that, until the legal environment becomes clearer, it might be safer to simply ban such contributions.\n\n\n\nIs AI‑generated code low quality? Humans are no better.\n\nAmid the discussion another intriguing viewpoint emerged.\n\nMany people oppose AI‑written code because they think its quality is poor. But Debian developer Russ Allbery argues that this reasoning is flawed: humans can indeed write code better than AI, but they can also write code far worse.\n\nHe even joked that writing meaningless garbage code takes no creativity, whereas producing truly awful code actually requires human talent.\n\nOther developers noted that AI might merely be another step in the evolution of software development, yet it brings new problems—for instance, if code is produced via a prompt, what is the preferred way to modify it?\n\nNussbaum’s answer was to edit the prompt rather than the generated code. However, this answer is unconvincing because LLM output is usually non‑deterministic, model versions keep changing, and even using the same prompt can yield completely different code later on.\n\n\n\nFinal result: Debian decides to wait and see.\n\nLooking at the whole discussion, Debian developers have far from reached a consensus on this issue; they have not even agreed on the most basic question: what counts as an AI‑generated contribution?\n\nNussbaum said he originally proposed a resolution because some members of the community had started attacking AI users. After a period of dialogue, however, the conversation has remained generally rational, so there is no need to rush a policy.\n\nHe also speculated that if a vote were held in the future, the likely outcome would be to allow AI use but attach a series of strict restrictions.\n\nReference link: https://lwn.net/SubscriberLink/1061544/125f911834966dd0/