But Is It Open?The online edition
Type & theme

Make yourself comfortable

Page color
Text size
20 px
Reading width
Ask about this chapter ↗

Your settings and place stay in this browser.

Chapter 9 · Timothy O’Brien

The Contribution Bottleneck

6 minute read12 source notesFree online edition

Recently I was in a discussion in California about the future of open source. Two questions came from the room. Why use open source when generative AI can build you a bespoke system from scratch? And why use libraries at all, when you can generate what you need?

My first reaction was about security. Anyone who went through the Log4Shell scramble in December 2021 knows how much work fixing a widely used library can take.1 Generating a replacement doesn’t make that work disappear. Someone still has to find its failures and keep it working after the first successful run.

The room was asking about something larger than replacing one library. If prompts and agents produce more of the code, the record of how a change was generated may become part of what a contributor submits. That changes the maintainer’s job as well as the programmer’s.

What remains scarce

Contributing to a serious project has always required context: its history, conventions, open problems, and the reasons earlier proposals were rejected. Trust builds through useful work and through staying around when that work needs revision.

Generative AI can help with both writing and review. The Linux Foundation’s October 2026 survey found self-reported time savings of 52 percent for code authoring and 44 percent for code review and feedback. About half of maintainers were already using AI for code review or bug checks.2

The difficulty is that faster review of one patch doesn’t guarantee a shorter queue. If submissions grow faster than a project’s ability to assess them, the backlog grows too. Someone still has to decide whether a change belongs in the project, understand its effects, and remain accountable when it breaks something. A tool can assist that judgment without taking responsibility for the release.

What maintainers are seeing

In the same survey, 59 percent of maintainers reported increased contribution volume since generative AI arrived; among large projects it was 69 percent. Respondents estimated that a quarter of incoming pull requests were low-quality “AI slop.” Only 13 percent of maintainers agreed that generative AI had improved community trust, while 53 percent disagreed.2 These are reports from a survey of 234 contributors and maintainers, not measurements of every open-source project.

The individual accounts show what the queue can cost. Daniel Stenberg ended curl’s bug bounty in January 2026 after low-effort AI-generated reports overwhelmed it.3 Godot’s Remi Verschelde described repeatedly having to question pull requests from new contributors.4 A CodeRabbit analysis found more issues in AI-co-authored pull requests than in human-only ones, though its small vendor-run study cannot establish the quality of AI contributions generally.5

Closing the door has costs too. A person who lacks time, confidence, or fluent English may use AI to make a useful first contribution. Apache Airflow allows AI-assisted pull requests when the contributor reviews the work and discloses the assistance. Its maintainers have also written an AGENTS.md file for tools to read.6 Clear project instructions give both the person and the tool a better chance of submitting something maintainers want.

The question is whether the project gains people who can learn and stay, or just more submissions. A patch can be helpful even if it was quick to write. It becomes expensive when nobody can explain it or respond to the review.

A replacement is another maintenance obligation

The California discussion also raised a harder question about what source code is worth when another implementation is cheap to attempt. A tool can help a team study a program and write a replacement, but “attempt” is doing a lot of work in that sentence.

Suppose a team uses a tool to reproduce one request-routing behavior it currently gets from nginx. A few tests may show the same result for ordinary requests. That leaves compatibility, malformed inputs, performance, security, and years of accumulated edge cases to investigate. Matching one observed behavior doesn’t establish that the team has replaced nginx. It has created a smaller component with its own maintenance burden.

The licensing question is separate. Translating existing code into another language doesn’t by itself establish that it may be relicensed. An independent implementation and a translated or adapted copy raise different questions. The generated output and the record of how it was produced would need examination; a prompt saying “rewrite under MIT” settles neither.

Code may become cheaper to produce without becoming interchangeable in practice. The original project’s tests, reviewers, release process, and accumulated knowledge remain valuable. A team choosing a replacement has to decide which of those obligations it is prepared to assume.

What projects are starting to record

The Apache Software Foundation permits generative-tool contributions that meet its license and intellectual-property criteria. Its guidance recommends commit tags such as Generated-by: and Co-authored-by: and considers future tooling to collect them into provenance records.7 OpenInfra has adopted Generated-By: and Assisted-By: labels alongside its sign-off process.8

Those labels establish a starting point. They don’t tell a reviewer whether the output is correct, whether the contributor read it, or what material the tool copied. A useful disclosure needs to support a review decision rather than merely announce that a model was involved.

GitHub’s pull request templates let a project ask for that information, while editor tools can add AI co-author trailers. The records depend on the contributor’s tools and the project’s expectations.9 GitHub’s 2026 pull-request controls also let maintainers restrict access or cap outsiders’ open requests, addressing the volume reaching the queue.10 Limiting submissions can protect reviewers, though a limit alone doesn’t help a new contributor understand how to improve one.

DSPy, the Stanford NLP framework led by Omar Khattab, shows why the record may need more than a prompt. Developers write modules describing tasks and supply a metric; an optimizer adjusts prompts, and sometimes weights, against that metric.11 To explain how such a program was built, a contributor would need to describe the program, optimization settings, and evaluation, rather than just paste the final prompt. What counts as useful provenance depends on the process that produced the artifact.

CISA and G7 partners’ voluntary 2026 guidance on SBOMs for AI covers dataset provenance, hashes, licenses, and model details.12 That is a record of components. It doesn’t require contributors to publish every prompt or agent transcript. A project deciding to ask for those records still has to explain what it will inspect and how the disclosure helps.

Who carries the review

I am uneasy about requiring contributors to publish all their context and prompts. Open source didn’t usually ask people to submit their scratch paper. There may be private data, unrelated work, or a long sequence of failed attempts in that record.

But an overloaded maintainer needs more than a plausible diff. The contributor can carry part of the burden by explaining the intended change, reviewing the generated code, running the project’s tests, identifying the tool assistance, and answering questions afterward. Where provenance bears on licensing or security, the relevant record matters too. Disclosure should be proportionate to that need.

This preserves an open door without treating every arrival as free work for someone else to assess. AI can help reviewers, but it doesn’t supply an unlimited number of trusted people willing to maintain a project. Shared security teams, reproducible builds, and coordinated disclosure can take some of that work off their desks. They are the subject of the next chapter.

Notes

  1. Cybersecurity and Infrastructure Security Agency, “Apache Log4j Vulnerability Guidance,” CISA, https://www.cisa.gov/news-events/news/apache-log4j-vulnerability-guidance; National Vulnerability Database, “CVE-2021-44228 Detail,” NIST, published December 10, 2021, https://nvd.nist.gov/vuln/detail/CVE-2021-44228. ↩

  2. Sam Boysel and Anni Lai, Navigating Open Source Software Development in the Generative AI Era: Challenges and Opportunities for Project Communities, LF Research, October 2026, https://www.linuxfoundation.org/hubfs/Research%20Reports/Navigating-Open-Source-Software-Development-in-the-Generative-AI-Era.pdf. The survey includes 234 contributors and maintainers; individual questions have smaller respondent counts. Figure 3 reports contribution-volume responses from 111 maintainers. Figure 8 reports community trust. Figure 19 (p.23) gives self-estimated time reductions of 52 percent for code authoring and 44 percent for code review and feedback; Figure 20 separates maintainers and contributors. These are respondent estimates, not a controlled timing experiment. See also the Linux Foundation announcement, October 20, 2026, https://www.linuxfoundation.org/press/new-lf-research-survey-shows-generative-ai-cuts-open-source-coding-time-in-half-but-25-of-incoming-pull-requests-are-considered-ai-slop. ↩1 ↩2

  3. Daniel Stenberg, “The end of the curl bug-bounty,” daniel.haxx.se, January 26, 2026, https://daniel.haxx.se/blog/2026/01/26/the-end-of-the-curl-bug-bounty/. The program officially stopped on January 31, 2026, after the confirmed-vulnerability rate fell below 5 percent in 2025. ↩

  4. Chris Kerr, “Godot veteran says ‘AI slop’ pull requests have become overwhelming,” Game Developer, February 17, 2026, https://www.gamedeveloper.com/programming/godot-co-founder-says-ai-slop-pull-requests-have-become-overwhelming, reporting Verschelde’s Bluesky post at https://bsky.app/profile/akien.bsky.social/post/3meyev4v2tc2p. The “draining and demoralizing” wording is Game Developer’s paraphrase (secondary source). In the post, dated February 16, 2026, Verschelde writes that maintainers “find ourselves having to second guess every PR from new contributors, multiple times per day.” ↩

  5. CodeRabbit, “State of AI vs Human Code Generation Report,” CodeRabbit blog, December 2025, https://www.coderabbit.ai/blog/state-of-ai-vs-human-code-generation-report; press release via Business Wire, December 17, 2025, https://www.businesswire.com/news/home/20251217666881/en/. The sample was 320 AI-co-authored and 150 human-only PRs. Vendor analysis, not peer reviewed. ↩

  6. Adrin Jalali, “Maintaining open source in the age of generative AI: Recommendations for maintainers and contributors,” probabl.ai blog, 2026 (undated; its timeline runs to February 2026), https://www.probabl.ai/blog/maintaining-open-source-age-of-gen-ai. Secondary source for Potiuk’s view. Airflow’s file is at https://github.com/apache/airflow/blob/main/AGENTS.md. Apache Airflow, “Pull Requests,” contributing docs, section “Gen-AI Assisted contributions,” https://github.com/apache/airflow/blob/main/contributing-docs/05_pull_requests.rst#gen-ai-assisted-contributions, says “Generally, it’s fine to use Gen-AI tools to help you create Pull Requests for Apache Airflow” if contributors follow its guidelines, and asks contributors to “State in your PR description that you have used Gen-AI tools.” ↩

  7. Apache Software Foundation, “ASF Generative Tooling Guidance,” https://www.apache.org/legal/generative-tooling.html. ↩

  8. OpenInfra Foundation, “Policy for AI Generated Content,” version 0.12.0, September 25, 2026, https://openinfra.org/legal/ai-policy. The policy now recommends Assisted-By going forward while still approving Generated-By. ↩

  9. GitHub Docs, “About issue and pull request templates,” https://docs.github.com/en/communities/using-templates-to-encourage-useful-issues-and-pull-requests/about-issue-and-pull-request-templates (issue forms can mark inputs as required; pull request templates are Markdown that fills the pull request body); GitHub Docs, “Creating a commit with multiple authors or on behalf of an organization,” https://docs.github.com/en/pull-requests/how-tos/commit-changes/creating-a-commit-with-multiple-authors. Checked against GitHub Docs and the GitHub changelog on 15 November 2026. Some code editors can add the trailer for AI: Visual Studio Code’s git.addAICoAuthor setting “can automatically append a Co-authored-by: trailer when you commit code that includes AI-generated contributions,” Visual Studio Code release notes, version 1.110, https://code.visualstudio.com/updates/v1_110, and version 1.118 turned it on by default for chat and agent workflows, https://code.visualstudio.com/updates/v1_118. That is an editor setting, not a GitHub requirement. ↩

  10. GitHub Changelog, “New repository settings for configuring pull request access,” February 13, 2026, https://github.blog/changelog/2026-02-13-new-repository-settings-for-configuring-pull-request-access; “Limit open pull requests for users without write access,” June 17, 2026, https://github.blog/changelog/2026-06-17-limit-open-pull-requests-for-users-without-write-access (GitHub says pull requests opened by Copilot or another AI agent count toward the limit; see https://github.blog/open-source/maintainers/how-pull-request-limits-are-cutting-down-the-noise); “Pull request limits: Exempt contributors by age or history,” October 28, 2026, https://github.blog/changelog/2026-10-28-pull-request-limits-exempt-contributors-by-age-or-history; GitHub Docs, “Available rules for rulesets,” section “Additional approval for unattributed Copilot pull requests” (public preview, on by default), https://docs.github.com/en/repositories/configuring-branches-and-merges-in-your-repository/managing-rulesets/available-rules-for-rulesets. Status as of 15 November 2026. ↩

  11. stanfordnlp/dspy, “DSPy: The framework for programming, not prompting, language models,” GitHub repository (MIT license), https://github.com/stanfordnlp/dspy; Omar Khattab et al., “DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines,” ICLR 2024, as cited in the repository README. ↩

  12. G7 Cybersecurity Working Group, Software Bill of Materials (SBOM) for Artificial Intelligence: Minimum Elements, BSI, May 12, 2026, https://bsi.bund.de/SharedDocs/Downloads/EN/BSI/KI/SBOM-for-AI_minimum-elements.html; CISA announcement, which shortens the title to Software Bill of Materials for AI – Minimum Elements, https://www.cisa.gov/resources-tools/resources/software-bill-materials-si-minimum-elements. The guidance is voluntary and builds on the G7’s June 2025 shared vision for SBOM for AI. ↩

From But Is It Open? by Timothy O’Brien (Discursive, 2026). © 2026 Timothy O’Brien. All rights reserved.

Chapter 96 min leftNext →

The book, chapter by chapter.

Source note

Go to the notes & sources ↓

The vocabulary.

A short guide to the book’s recurring terms. Follow each chapter for the argument and its sources.

Open weights
A release makes a trained model’s learned parameters available. The accompanying terms decide what you can do with them. Access to weights alone does not provide the training data, code or recipe. Read the chapter →
Copyleft
A way of using license conditions to preserve recipients’ freedoms when covered software is passed on. The obligations depend on the license and on what you distribute or make available. Read the chapter →
Permissive licenses
Licenses such as MIT and Apache 2.0 grant broad reuse rights with comparatively few conditions. Notice obligations still apply, and a software license does not automatically cover a release’s data or other components. Read the chapter →
Reproducibility
The ability to reconstruct and examine how an artifact was produced using the resources that are published. Running an existing checkpoint is a different achievement from rebuilding its training. Read the chapter →
Governance
Who can participate, who makes decisions, and how authority over a project or release is shared. A right to use the code does not by itself create a voice in upstream decisions. Read the chapter →
Training data
The examples used to train a model. Publishing the data, documenting its provenance and describing how it was filtered are distinct kinds of disclosure; privacy and other rights may limit what can be shared. Read the chapter →

A passage worth sharing.

Select a passage in the chapter, or share the one you’re reading.