Coding Horizon

GitHub Repos Are Banning AI And Agents Ignore It

Every figure this video puts on screen, chased to the paper that reports it.

AI contribution policies on GitHub

Andre Hora, Romain Robbes. “AI Policy, Disclosure, and Human in the Loop: How Are Contribution Guidelines Adapting to GenAI?” arXiv:2605.16706. Submitted 15 May 2026, revised 13 July 2026. Accepted at ICSME 2026. https://arxiv.org/abs/2605.16706

• 1,000 popular GitHub repositories analysed, 118 AI policies for contributors identified.
• 78% of those policies allow AI-assisted contributions. 22% explicitly discourage AI use.
• 51% require disclosure of AI-assisted contributions.
• 74% require a human in the loop during contribution.

Within the 78% that allow AI, the paper separates degrees of welcome: 60 of the 118 policies (51%) explicitly welcome generative AI, with the remainder permissive without being enthusiastic.

Whether coding agents comply with those policies

Wenhao Yang, Runzhi He, Minghui Zhou. “A First Look at Coding Agents’ Compliance with AI Contribution Rules in Open-Source Communities.” arXiv:2607.26819. Submitted 29 July 2026. https://arxiv.org/abs/2607.26819

The benchmark, RepoComplianceBench, is built from 455 hand-coded AI contribution rules drawn from 102 open-source communities, sorted into four rule types: Refuse, Disclose, Verify and Handoff. From those it curates 106 issues taken from 49 repositories that carry AI contribution rules. Four frontier coding agents were tested at 280 runs each.

• The policy file was opened in 12 of 347 unaided runs, which is 3.5%. Of 248 unaided violations, 242 of them, or 97.6%, happened without the policy ever being opened.
• In repositories that prohibit AI-generated contributions, refusal was 0% for all four agents in the unaided condition.
• Disclosure compliance runs 17% to 40% unaided, and rises to 77% to 97% when the agent is given a generic reminder, the clause verbatim, or one round of verifier feedback. One of the four agents did not improve under reminders or quoted clauses.
• Verification compliance is close to flat under reminders and quoted clauses, and the paper describes it as behaving more like a fixed system trait than a response to steering. It rises to 90% to 100% only under verifier feedback.
• Refusal stays near zero across every steering condition. Quoting the prohibition verbatim leaves refusal at 0% for three of the four agents. The highest refusal rate recorded in any cell is 23%, reached by one agent under one round of verifier feedback.

Note on the refusal figure: the paper’s abstract states that the agents never refuse to contribute in AI-banned repositories under any condition tested, while the paper’s own results table records non-zero refusal in five cells under steering. The 0% result is solid for the unaided condition, which is the condition this video’s figure is drawn to.

The scale of coding agent activity

Arsham Khosravani, Audris Mockus. “Detecting AI Coding Agents in Open Source: A Validated Multi-Method Census of 180 Million Repositories.” arXiv:2606.24429. Submitted 23 June 2026. https://arxiv.org/abs/2606.24429

• Across snapshots from December 2024 to April 2026, commit-attributed agents generate over 320,000 commits per month.
• Multi-method detection identifies 850,157 Claude Code commits in a single snapshot.
• Bot-account lookup, the signal most adoption studies rely on, recovers only 28,154 of those, which is 3.3%, a thirtyfold relative-recall gap.

The 850,157 figure is the single-snapshot count. The same paper separately reports 886,122 Claude Code commits across 17,295 projects cumulatively across snapshots.

Whether projects are regulating or prohibiting

Yunqi Chen, Thomas Zimmermann, Bianca Trinkenreich. “Making AI Visible, Not Vanished: How AI Policies Reshape Developer Experience on GitHub.” arXiv:2608.03329. Submitted 4 August 2026. https://arxiv.org/abs/2608.03329

• 29,624 GitHub repositories analysed, 385 of them found to have adopted AI policies.
• The paper concludes that AI governance primarily regulates rather than prohibits AI-assisted development.

Not carried into the video

Three further studies of the same question were read and left out, because each covers a much smaller population than the four above and none of their figures appear on screen: “Regulating the Machine Contributor” (arXiv:2606.14594, six organisations), “Beyond Banning AI” (arXiv:2603.26487, 67 projects), and “Making Agent-Mediated Contributions Governable” (arXiv:2607.15769, 50 repositories).