Recently I left a comment on a pull request. It said the change fixed a symptom; it explained, briefly, why the code was the way it was and how that had produced the symptom; it suggested a fix that respected the why; and it asked what people thought. It was as long as that needed and no longer, and it was the kind of comment I've been leaving on GitHub for over a decade. The replies didn't touch the argument. Too dense. Too long. Don't use AI prose in comments. I asked what in the comment was wrong and got the same answers back. It wasn't the first time, in that thread or others, and it isn't only happening to me.

I've heard this reply before, in a different costume.

Mechanism

The bracket

"Your English is so good." Anyone who has been on the receiving end knows the sentence has a second half the speaker never says: for someone who looks like you. The compliment isn't about the English. It's about the surprise that this person produced it, and it tells you exactly who was expected to produce it and who wasn't.

"This reads like AI" carries the same bracket. It isn't a claim about the text, because the text can be a correct, on-topic review comment and still draw it. It's a claim about who could plausibly have written that text unassisted, and the bracket reads: not you. The mechanism is old and measured: Lev-Ari and Keysar had people rate identical trivia statements and found them judged less true when read in a non-native accent, and the driver was processing effort, not stated prejudice. Harder to read, therefore less credible. The accent has moved from the ear to the page; the effect hasn't.

I wrote a long piece on the first half of this problem earlier this month. The short version: nobody can tell. 4,600 people in the largest study scored 50 to 52 percent, and paying them for correct answers did nothing. The checklist people use is a description of an average applied to an individual. In 2026 that checklist says: smooth sentences, structure, no filler, careful vocabulary, "delve".

Now look at who writes like that when no model is involved. Non-native English speakers, who learned the language from grammar books and write it more carefully than the people who grew up in it; Liang and colleagues ran seven detectors over TOEFL essays and got a 61 percent false-accusation rate, against near zero for American eighth-graders. Nigerian and other West African English, where "delve" was ordinary business vocabulary long before OpenAI's annotators, many of them in Lagos and Nairobi, taught a model to like it. Autistic writers, whose register is precise and unpadded, and whose forum posts a 2026 study found were flagged at a significantly higher rate than everyone else's. Anyone who was taught to communicate clearly on an asynchronous medium and then did.

Open source didn't need a new mechanism for treating these people worse. Rastogi and colleagues, on 70,000 pull requests, found that where you appear to be from predicts whether your patch is accepted: a submitter in Switzerland had 58 percent higher odds than one in the United States, and China and Germany sat near the bottom. Terrell's 1.4 million-user gender study found the same shape. The population most exposed is also the one growing fastest: GitHub's own numbers put India at 5.2 million new developers last year, more than any other country. The accusation is landing on the newcomers.

Is it landing as a rhetorical move? A June study of 25 million Hacker News and Reddit comments found AI accusations up more than tenfold since 2023, with their tone shifted from mockery to what the authors called gatekeeping. A control list of older accusations ("shill", "astroturf") didn't move.

That is the closest thing to a measurement I have. It's a preprint, and it isn't specific to open source. Nobody has yet counted whether OSS accusations fall disproportionately on non-native writers; someone should. Until they do, I'm arguing from mechanism, and I think the mechanism is obvious.

Steelman

What maintainers are right about

The complaint that produced the policies is real. Daniel Stenberg ran curl's bug bounty for seven years; the confirmed-vulnerability rate sat north of 15 percent, then in 2025 fell below one in twenty as generated reports flooded in, and in January he shut the programme. That is a volunteer being denial-of-serviced by people who couldn't be bothered to read their own submissions. Gentoo, NetBSD, QEMU and Zig have said no to generated code; CPython, the Linux kernel, Rust and Debian have said yes with a human accountable. All of that is governance, and governance is a maintainer's job.

The accusation isn't governance. A policy tells contributors what a project accepts. A reply that says "sounds like AI" and stops tells one contributor that their argument doesn't need answering. The first protects a project. The second protects the replier from having to think.

The first protects a project. The second protects the replier from having to think.

Cost

"Why should I read it"

The other thing I hear is that maintainers shouldn't have to spend their time on text a model produced. I think it's the wrong complaint, because reading a comment costs the same whether a person or a tool typed it. What costs you time is a comment that is too long, or wrong. Those are review comments, and you should leave them: "the second half repeats the first"; "this misreads how the resolver handles extras". A contributor can act on either of those, because they are about the text.

"Reads like AI" is about neither. If the comment is clear and right, and a model helped a Portuguese-speaking contributor get it to that state, the project got a better comment than it would have had without the model. That's the outcome we say we want. Complaining about it is complaining that the argument arrived in good order, and I struggle to read that as anything other than a complaint about having to engage with it.

Practice

The ask

Judge the comment on merit, the way this community has always claimed to judge the patch. Three questions are enough. Does it make its point? Is it professional and respectful? Is it technically right? If all three are yes, the reply is engagement, not a guess about provenance. If one is no, say which, and say why. That's a review. "Sounds like AI" isn't a review and it isn't a policy; it's a hunch, with a false-positive rate the research puts somewhere between bad and unfixable, aimed by default at the people with the least standing to push back.

Open source got where it is by reading the patch and not the passport. The patch has started arriving in better English. That isn't a problem to police.

Sources

What this leans on

  1. Jakesch, Hancock & Naaman, PNAS 2023
  2. Lev-Ari & Keysar, Journal of Experimental Social Psychology 2010
  3. Liang et al., Patterns 2023
  4. Willison, "How cheap, outsourced labour in Africa is shaping AI English", 2024, after Hern in The Guardian
  5. "The Misclassification of Autistic Writing as AI-Generated", Springer AIED 2026
  6. Rastogi, Nagappan & Gousios, "Geographical Bias in GitHub", ICSE/MSR 2016
  7. Terrell et al., PeerJ Computer Science 2017
  8. GitHub, Octoverse 2025
  9. "That's AI Slop, You Bot!", arXiv 2606.12073, June 2026 (preprint)
  10. Stenberg, "The end of the curl bug-bounty", 26 January 2026
  11. melissawm/open-source-ai-contribution-policies, a maintained list of project policies

Disclosure

In case you were wondering

An agent helped with this one too. It did the background research, checked the numbers against the originals, and ran two editing passes over my draft. The argument, the comment it opens with, and the irritation are mine.