The Anti-Slop Manifesto
Producing a contribution can now cost almost nothing. Evaluating it still takes time. The problem is not that a machine was involved. It is that nobody did the work required to stand behind the result.
The handshake
Collaborative work runs on an implicit exchange. A contributor spends time understanding the work, reproducing a problem, or preparing a change. A reviewer spends time evaluating the result. Neither side owes the other agreement, but each arrives having done some work.
The effort required to produce a detailed report or a polished patch was never proof of correctness. It was a weak but useful signal that someone on the other side had read the code and might still be there when questions followed.
That signal is gone. A plausible report, patch, review, or security advisory can now be produced without the understanding its surface implies. Production became cheap. Evaluation did not.
The asymmetry
A generated vulnerability report can take minutes to submit and hours to disprove. A generated pull request can take seconds to open and an afternoon to understand, test, and reject. The author pays almost nothing. The recipient pays the full cost of determining whether any of it is real.
Slop is work submitted by someone who has not evaluated it, transferring the cost of evaluation to someone else. It can be generated by a model, copied from a forum, or written carelessly by hand. The production method is secondary. The abdication is the problem.
A mistake is different. Someone can reproduce a bug, study the code, make a good-faith argument, and still be wrong. Mistakes are part of the work. Slop skips the work and asks the reviewer to perform it after submission.
What it is already costing
Daniel Stenberg ended curl's bug bounty after the work of rejecting AI-assisted reports began to displace useful security work.
Ghostty changed the intake path. Users begin with a Discussion; maintainers turn a report into an issue only after it becomes reproducible and actionable. A polished submission is not evidence of understanding.
Some projects prohibit generated contributions. Others permit private assistance but require public work to be tested, understood, and owned. Reviewers need not absorb unsolicited machine output. Tools do not replace authorship.
The downstream cost
The immediate cost is time: reports debunked, patches re-tested, claims traced to sources that do not exist. At enough volume, maintainers spend their hours rejecting submissions instead of maintaining the project.
The larger cost is trust. When polished reports routinely conceal an absence of understanding, polish stops carrying information. Review slows down. First-time contributors face more suspicion. Projects add gates, close intake channels, or restrict participation to people already known.
Those defenses reduce noise. They also make open source less open. Good contributors inherit the friction created by people who transferred their work to the review queue.
Our position
Slop is not AI, and AI is not slop. A developer who uses a model to explore an implementation, then reads it, tests it, understands it, and takes responsibility for it has retained responsibility. A developer who submits output they did not run or cannot explain has submitted a liability with their name attached.
The useful question is not which tool touched the work. It is whether its author can stand behind it. Was the bug reproduced? Are the claims sourced? Do the tests cover the change? Can the author explain the decisions? Will the author remain present when review finds a problem?
This document was shaped with the assistance of a language model. We use language models throughout our work. We treat generated output as a draft or a hypothesis, never as evidence. When a model is wrong, it is not accountable to our users. We are.
The standard
If you submit it, you own it. The tool is not responsible for the report, patch, paper, or image. You are.
If you cannot explain it, do not submit it. A contribution you do not understand is review work assigned to someone else without consent.
If you have not verified it, label it as a question. Guess privately. Ask honestly. Report a fact only when you have evidence for it.
If review finds a problem, remain present. Authorship includes responding, correcting, and learning. It does not end when the output is posted.
The barrier to entry was never the point. The effort was.
Tools scale. Responsibility stays human.