Disparate Impact of Current AI Policies on Contributors Using AI as Assistive Technology

[ritorhymes]

I have strong concerns that several policies as written in https://www.kubernetes.dev/docs/guide/pull-requests/#ai-guidance have a clear disparate impact (a form of discrimination) against contributors with disabilities who rely on AI as an assistive technology.

Here are the 3 policies of concern (the first quote containing two separate policies):

Large AI generated PRs and AI generated commit messages are not allowed.

When responding to review comments, you must do so without relying on AI tools. Reviewers want to engage directly with you, not with generated responses. If you do not engage directly with reviewers, the PR will be closed.

The policies do not distinguish unaccountable delegation to AI from accountable AI-assisted or assistive use. Instead, they categorically prohibit several core uses of AI that contributors with disabilities may rely on to participate.

Policy Intent

I understand the intent of the policies. I’ve contributed to the Linux Kernel repo before and have heard Jon Corbet raise concerns about maintainers being overwhelmed by spam and low-quality AI-generated contributions that strain their ability to manage the review pipeline effectively.

We can distill the policy intent here into two legitimate concerns:

1. Preventing careless, haphazard, unverified, low-quality AI-generated contributions, that may include bugs, code smells, sloppy logic, untested changes, or poorly scoped architectural directions.

2. Protecting maintainers from spending limited review bandwidth on an overwhelming volume of low-quality generated work, often pejoratively described as “AI slop.”

Accessibility Definitions

Under U.S. accessibility law, AI can function as assistive technology or an auxiliary communication aid when a contributor uses it to increase, maintain, or improve their ability to draft, type, organize, explain, review, or respond in a technical contribution process.

The Assistive Technology Act, codified at 29 U.S.C. § 3002(4), defines an “assistive technology device” broadly as:

“any item, piece of equipment, or product system, whether acquired commercially, modified, or customized, that is used to increase, maintain, or improve functional capabilities of individuals with disabilities.”

Source: https://uscode.house.gov/view.xhtml?path=/prelim@title29/chapter31&edition=prelim

ADA regulations under 28 C.F.R. § 35.104 define “auxiliary aids and services” to include “accessible electronic and information technology,” “acquisition or modification of equipment or devices,” and “other similar services and actions.”

Source: https://www.ecfr.gov/current/title-28/chapter-I/part-35/subpart-A/section-35.104

The DOJ’s ADA guidance on effective communication says the appropriate aid or service depends on:

“the nature, length, complexity, and context of the communication” and the person’s “normal method(s) of communication.”

Source: https://www.ada.gov/resources/effective-communication/

Disparate Impact of Policies

The policies as written effectively filter out some unserious contributors submitting low-quality work through haphazard use of generative AI tools. That is a legitimate goal.

But they also filter out high-quality contributions from contributors with disabilities who rely on AI as assistive technology, even when those contributors understand the architecture, implementation, and changes made; test thoroughly; respond thoughtfully; and assume full accountability for the work.

Large AI generated PRs ... are not allowed.

A contributor who does not rely on AI as assistive technology is welcome to submit a large PR, but contributors with disabilities who do rely on AI as assistive technology are prohibited from making equivalent contributions through the tools they need, even if both contributors could produce work of the same quality and demonstrate the same understanding.

If the answer is that such a contributor should simply break the work into smaller PRs, that imposes an additional procedural burden that does not apply equally to contributors who are allowed to produce the same large contribution without AI assistance. That burden can make the contribution harder to sequence, harder to maintain, more vulnerable to rebasing churn, and less likely to land. It can also force the work into an unnatural shape for reasons unrelated to the quality, coherence, or reviewability of the contribution itself.

The remaining two policies are even more problematic because they impact the fundamental ability to participate at all:

AI generated commit messages are not allowed.

When responding to review comments, you must do so without relying on AI tools. Reviewers want to engage directly with you, not with generated responses. If you do not engage directly with reviewers, the PR will be closed.

For contributors with disabilities who rely on AI as assistive technology for writing, drafting, organization, or communication, they are effectively prohibited from participating at all since they are not allowed to use their assistive tooling for the most fundamental tasks required. A contributor may fully understand, verify, and own a change while relying on AI as an assistive tool to document it or communicate during discussions around it. Furthermore, AI-generated outputs can still be reviewed, refined, corrected, and selectively submitted by the contributor so that the final response accurately represents their own understanding, decisions, and intent.

The policies should distinguish unaccountable AI delegation from accountable AI-assisted or assistive use.

Disclosure and Reviewer Bias

The disclosure requirement also needs guardrails for guidance like in https://www.kubernetes.dev/docs/guide/pull-requests/#large-or-automatic-edits :

To make it easier for reviewers to handle such Pull Requests, please explain how it was generated in the “Special notes for your reviewer” section of the Pull Request description. Reviewers may then be able to reproduce those steps (search/replace, linters) or can start the review with the right expectations (AI tools).

It is reasonable to ask contributors to disclose workflows that affect reviewability or reproducibility. But language about reviewers starting with “the right expectations” when AI tools are disclosed risks inviting subjective bias against AI-assisted contributors. The policy should make clear that disclosed AI use does not lower the presumption of contributor competence by itself, and that review scrutiny should remain tied to objective signals: quality, verification, tests, maintainability, and whether the contributor can effectively explain and revise the work.

Disclosure should also not create an informal second-class review lane where AI-disclosed contributions are deprioritized, ignored, or presumed lower quality before review. If additional scrutiny is needed, it should be tied to objective review concerns such as size, complexity, test coverage, reproducibility, maintainability, or whether the contributor can explain and revise the work. It should not be based on the mere fact that AI was disclosed, especially where AI is being used as assistive technology.

Policy Considerations

I honestly don't have an exact answer for what the policy should say to strike the right balance: filtering low-quality or unaccountable submissions without excluding legitimate contributors using the same underlying tooling. This is a challenging issue, the technology is evolving quickly, and policies written today could become outdated as AI-generated code quality improves and requires less human guidance.

I don't think there is a straightforward way to capture every edge case, but I do know that the policy should not fundamentally block an entire segment of contributors outright, especially regarding commit messages and review communication.

The policies should attempt to distinguish unaccountable delegation to AI from accountable AI-assisted or assistive use. They should not prohibit contributors with disabilities from using AI as assistive technology for basic contribution tasks when the work is otherwise reviewable, meets the project’s quality standards, and the contributor understands, verifies, owns, and can explain the substance of their work.

The policies should also guard against disclosure becoming a source of harmful bias. If disclosing AI assistance causes contributors to be deprioritized, presumed less competent, or placed into an informal lower-trust review lane, the policy will discourage honest disclosure and create worse incentives. Contributors should not be penalized for transparency.

[k8s-ci-robot]

This issue is currently awaiting triage.

SIG Docs takes a lead on issue triage for this website, but any Kubernetes member can accept issues by applying the triage/accepted label.

The triage/accepted label can be added by org members by writing /triage accepted in a comment.

Details

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes-sigs/prow repository.

[lmktfy]

Because this is project-level feedback:
/sig contributor-experience
/transfer kubernetes

[dims]

/transfer community

@lmktfy project-level but community related :)

[BenTheElder]

Most of what is written above infers things that simply aren't written in the policy nor occurring In practice. Quite the opposite. Large, obvious "slop" PRs that don't follow policies after being made aware of them are a signal to close.

PRs that disclose AI usage are not being shunted to "second class" and I'm on both ends of this submitting (AI-disclosed assisted PRs) and reviewing them as a long time highly prolific contributor.

PRs from people that bot their accounts to spam us and refuse to follow our policies after it has been raised however may be closed this extends to other merge requirements.

There are "right expectations" issues when the contributor does not engage with our documented contribution guidelines but sends us an enourmous volume of content.

A contributor who does not rely on AI as assistive technology is welcome to submit a large PR, but contributors with disabilities who do rely on AI as assistive technology are prohibited from making equivalent contributions through the tools they need, even if both contributors could produce work of the same quality and demonstrate the same understanding.

No. Usage of AI assistance is permitted. Wholesale generation of large PRs is what the policy refers to, because there is an existing policy for large PRs which applies generally.

If the answer is that such a contributor should simply break the work into smaller PRs, that imposes an additional procedural burden that does not apply equally to contributors who are allowed to produce the same large contribution without AI assistance. That burden can make the contribution harder to sequence, harder to maintain, more vulnerable to rebasing churn, and less likely to land. It can also force the work into an unnatural shape for reasons unrelated to the quality, coherence, or reviewability of the contribution itself.

No we expect large PRs to be broken up generally. There are exceptions for again, disclosed, deterministic bulk fixes (eg adopting formatters) further up in the policy that can trivially be independently verified (eg by the CI). These are the referred to exception.

[BenTheElder]

Happy to add further clarifying language, but this should be read in the context of the entire policy page, not as an out of context snippet, and specifically as an addendum to the large edits policy immediately preceding the AI section.

The large edits policy is long preexisting and already covers breaking up large PRs for all contributors, with specific essential carve-outs for reviewer agreed reproducible bulk edits such as find-and-replace or gofmt updates which must necessarily be submitted atomically to function.

[ritorhymes]

Resolution on Large PRs

Thanks for the context on the large PRs. If the preexisting large-edits expectations apply equally to everyone, AI-assisted or not, then there may not be an asymmetry there and my concern on that point is largely addressed.

The ambiguity is that the AI section does not make that relationship clear. If the general large-edits policy already governs large PRs regardless of tooling, then restating it as if it is a separate ban on large AI-generated PRs can read as an additional AI-specific restriction rather than as a cross-reference to the existing large-edits standard. Because that section may be linked, quoted, or read in isolation, the bare sentence can reasonably come across as stricter than the general policy, or as creating a double standard around what is considered a large PR.

I appreciate your willingness to add clarifying language here. With language in the AI section itself explaining that large AI-assisted PRs are governed by the same reviewability/reproducibility expectations as other large edits, and that the target is unreviewed, unaccountable, or non-reproducible generated work, I would consider this part resolved.

---

Review Communication Clause

There are a few remaining items to consider but for this comment I will focus on the biggest unaddressed concern which is the review communication clause:

When responding to review comments, you must do so without relying on AI tools. Reviewers want to engage directly with you, not with generated responses. If you do not engage directly with reviewers, the PR will be closed.

Reasonable Use Case Ban

A reasonable use case ban I can imagine would be: unattended automation, where an agent monitors the PR and responds to reviewers on the contributor's behalf without the contributor even being aware of the specific interaction.

How the Current Ban Language Overreaches

However, as written, the clause categorically bans relying on AI tools in review communication even for contributors with disabilities who use AI as a fundamental assistive technology to draft, organize, or express responses. Because responding to review is a fundamental part of the contribution process, that language can effectively prevent those contributors from participating at all.

A contributor can fully understand, verify, and own a change while relying on AI as an assistive tool to communicate during the discussions around it. AI-drafted responses can still be reviewed, refined, corrected, and selectively submitted by the contributor so that what gets posted accurately represents their own understanding, decisions, and intent. In every way that matters to a reviewer, that contributor is engaging directly. The mediation is in the articulation, not the substance.

Direction of Remedy

Remove the Categorical Ban

The blanket ban on AI-mediated review communication should be removed.

Address Quality Problems Through Existing Standards

If the goal is to reduce low-quality participation consuming maintainer bandwidth, a problem AI tools may exacerbate, the AI guidance should reinforce the project’s existing contribution standards rather than create a separate AI-specific trust penalty. Relevance, accuracy, responsiveness to feedback, appropriate disclosure, and contributor accountability should apply consistently to everyone, regardless of whether their communication appears AI-assisted.

Generic replies, failure to answer a reviewer’s question, inability to explain or revise the work, spam-like response velocity, and repeated failure to engage meaningfully are all legitimate review concerns. But they are legitimate because they are quality and accountability problems, not because they are “AI behaviors.” If the same conduct would be handled as inexperience, low competence, or failure to meet project standards when no AI use is apparent, it should not become a separate AI-policy violation merely because the contributor used AI assistance or because the response appears AI-mediated.

AI may make low-effort contributions easier to produce at scale, so it is reasonable for the project to enforce its quality bar more rigorously. But that quality bar should rise consistently for everyone. The standard should be reviewability, responsiveness, accountability, and ability to explain and revise the substance of the work, not whether a contributor’s communication appears to have been mediated by AI. Unlike a tool-based rule, that standard does not require anyone to determine whether or how AI was used. It can be evaluated from observable behavior in the review thread: whether the contributor answers the question, responds to follow-ups, explains the work, and revises it meaningfully.

Preserve a Narrow Ban on Unattended Automation

The distinct case that should remain prohibited is unattended automation: responses made on a contributor’s behalf without the contributor actually having awareness or agency over the interaction. The policy should be careful not to conflate that with AI-mediated writing, assistive drafting, or translation support. Those can produce stylistically similar text, but they are not the same conduct.

The clearest signal of unattended automation would be a long, polished response landing faster than a human contributor could reasonably read, understand, and respond to the reviewer. Outside of that kind of automation concern, the policy should treat quality and accountability issues as ordinary review problems, not as AI-specific violations.