[AINews] The No AI Slop Manifesto

The No AI Slop manifesto is a short statement of values for people who use AI at work, and its central demand is that whoever sends AI-made work takes responsibility for it. It was written by speakers and attendees of the AI Engineer Code Summit 2026 (AIE CODE) and copies the shape of the 2001 Agile Manifesto: a list of things the signers "have come to value" over others, such as outcomes over output, understanding over speed, and their own thinking over a model's first draft, followed by beliefs and resolutions. According to Latent Space's AINews, written by swyx, Dex Horthy introduced it in the conference's closing keynote after many speakers had cosigned it, and Hacker News voted it onto its front page.

The manifesto measures slop by effort, and its test applies whether or not AI was used. Slop is "when we put less effort into producing something than we ask the consumer to spend understanding it." Its example of that imbalance: an agent can now produce a thousand-line PR or a ten-page document at almost no cost, while the cost of reviewing it stays the same. Using AI does not by itself make work slop. In the manifesto's account, slop happens when someone passes AI output along without taking responsibility for its quality or for the reader's experience. It also argues that slop compounds, because whatever sits in a team's knowledge base, chat or code is what the next agent reads and copies, and it names pressure from companies pushing their people to do ever more as one cause.

What it asks for is concrete. When an agent wrote something you are sending, say so, and check the parts that matter before it goes out. When someone sends you slop, tell them kindly and ask them to review it first. Keep reviewing at least some of what agents produce, with the stakes deciding how much, because agent reviewers, linters and verifiers help but are not enough on their own. When the same mistake shows up twice, fix the system and its guardrails, and let a linter, type checker or other automated check catch whatever it can. Judge work by what shipped and who it helps rather than by lines of code, merged PRs, tokens spent or agents running in parallel. The manifesto does not ban unreviewed output: it calls one-shot prompts and vibe coding "high leverage tools" and treats slop as a problem when it undermines a useful product, a maintainable system or a team's time. Among its suggested reading is Horthy's talk Why Software Factories Fail, which argues that review agents and extra tokens can raise the floor but stay limited by what models learn in training, so for now engineers still have to read code.

None of this is a rule set or a measured result. The manifesto says plainly that "none of us has this figured out," invites comments, and keeps collecting signatures through comments on a gist, so its list of signers is still changing. AINews says it was started to offer "a useful tool for discussion at work," and that is the best way to read it. Its most useful contribution is the effort test, which gives a team a way to argue about AI-assisted work in terms of what it costs the reader, and leaves each team to decide which outputs are worth a human's review.