I submitted one of my tools to a community software list last week. Standard drill: fork the repo, add one line, fill out their template, open the pull request.
What came back was a comment from the maintainer asking if I realized how it looked, creating a new GitHub account just to promote a project with no users. The account is over a year old. His comment lasted less than a minute before he deleted it himself. Then he proposed a new rule for the list banning AI-generated submissions.
At no point did anyone say a word about the tool itself. Not whether it works, not whether anyone on that list would use it. Everything got decided at the level of how my account looks to somebody skimming it for thirty seconds.
The honest part: I get it. AI slop is real. Machine-generated bug reports that cite functions that don't exist, bounty submissions that sound great and describe nothing real. Maintainers of big open source projects are drowning in this stuff, their time is the only resource they actually have, and a cheap screening heuristic that's right most of the time beats an expensive review they can't afford to run on everything. I'm not here to argue the guard rail down.
The phrase is drifting, though. Spam filters scored the message itself. You could point at the payload in a spam email, the link, the fake invoice. A slop call scores the sender instead: account age, follower count, stars, users, whether the prose has the wrong rhythm. All of that is proxy, and none of it touches the only question a curated list exists to answer, which is whether the thing is any good.
And proxies fail quietly. When the screen catches real slop, the maintainer saves an hour and never thinks about it again, but when it catches a real project from someone whose account doesn't look busy enough, nobody ever finds out. There's no error message. The submitter just goes away, and the list keeps drifting toward people who already looked right to it.
Even the automated version of the sniff test does no better. I fed the AI detector built into Substack, the platform you're reading this on, a piece that really was machine-generated, start to finish. It came back saying low 80s percent likely AI. Then I wrote on the same topic by hand, no tools, every word mine, and ran that through the same detector. 96 percent. It was more confident about the human than the machine.
There's a reason it comes out backwards. These models learned to write from our writing. That's the whole training set, enormous piles of human text, yours and mine scraped up along with everything else, so the rhythm a detector flags as machine-like is just averaged human rhythm. Hunt hard enough for the copy and you're eventually going to point at an original. A better detector doesn't get you out of that.
Nobody using the word can tell you what it means, either. If slop lives in the words themselves, the wrong ones in the wrong rhythm, then the thesaurus has a problem: a book that hands you lists of interchangeable words on demand, and nobody has ever called Roget's slop, and if it lives in the ideas underneath, people shipped bad ideas by hand for centuries before a model ever wrote one. Maybe it's about origin, except my 96 percent already showed you can't tell origin from the outside. The strongest version is that slop means output nobody cleaned up. Closer. But then someone who drafts with a model and rewrites every line in their own words isn't making slop at all, and the word stops being about AI. What's left is the only definition that works: nobody checked it before it shipped. That's a fact about effort, and effort doesn't show up in account metadata.
And the incentives all run one direction. Calling something slop takes ten seconds and can't be wrong in any way that costs you anything. Actually reading the submission takes an hour, and it might end with you having to write the harder sentence: I read it, and here's what's wrong with it. An accusation that skips the reading is low-effort content about low-effort content.
So what's being guarded? Attention, partly, and that's fair. A curated list is only worth the curation behind it. But some of it is older than AI. Rejection used to come with a stated reason. Wrong license, no docs, doesn't build, we already list three of these. A slop call comes with nothing attached, no standard cited, nothing you could fix.
None of the earlier writing tools set off anything like this. Spell check showed up and nobody went around yelling spellcheck slop at every clean paragraph. Grammar checkers and autocomplete got absorbed so completely nobody remembers them being new. Framing crews traded hammers for nail guns, and a nail gun in careless hands will absolutely build you a bad house, but no inspector ever failed one for nail gun slop. They checked whether the framing was square.
Open source spent thirty years bragging that it had solved this exact problem. Show up with a patch and nobody cares who you are, because the code either works or it doesn't. That was the whole pitch, and it's the thing that let self-taught nobodies route around every institution that wouldn't have them. The moment the test becomes "does your account look human," the credential system is back, except worse, because the old one at least told you which credential you were missing.
What holds up is the boring thing: read it. If something's wrong with it, you can say what, specifically. Quote the fabricated function. Tell me nobody needs another one of these. Either answer takes real time, which is the point, and I'll grant the math turns brutal at the volume the big projects see. I don't have a scalable answer for a maintainer staring at fifty of these a week. Neither does anyone else yet.
My pull request is still sitting there. Open. As far as I can tell, nobody has read it.
Support open, independent research at [icsacinstitute.org/community/donate](https://icsacinstitute.org/community/donate).icsacinstitute.org/community/donate.*