Slop is a choice

A language model generates the median of everything it has seen. That is its nature and, correctly used, its virtue: a competent first draft of almost anything, in seconds. Median is the default. It was never the destiny. Blaming the model for slop is blaming the printing press for bad novels.

Before AI, mediocre work had a natural rate limiter: it still took effort to produce, so there was less of it, and the effort forced at least a minimum of examination. Generation removed the limiter. Volume with judgment is leverage. Volume without judgment is slop — and the volume was chosen by whoever pressed generate and shipped the result unread.

Taste is unglamorously concrete: knowing what good looks like in your domain, and being able to say why. It is not a gift. It is trained — by reading excellent code and excellent prose, by being reviewed hard, by rewriting until the difference between fine and right becomes visible.

The economics have inverted around it. Generation is now nearly free; discrimination is not. The scarce skill is the eye that rejects nine drafts and can articulate why the tenth is right. A person with taste plus a model produces more good work than either alone. A person without taste plus a model produces confident slop at unprecedented speed. The model is an amplifier, and amplifiers do not care what they amplify.

Taste in one head does not protect an organization. It has to be institutionalized, and there are exactly two mechanisms that work.

The first is review culture: human judgment applied to every artifact that ships, with AI-assisted work held to the same bar as human work. The moment a team maintains a separate, lower standard for generated output, it has chosen slop and outsourced the signature. The second is evals: taste written down as executable criteria — golden examples, failure taxonomies, acceptance thresholds — run against every generation and every model upgrade. Review scales judgment across people. Evals scale it across volume. Teams that build both ship AI-assisted work indistinguishable from their best human work, because it was judged by the same standard.

The strongest counterargument is arithmetic: if models let a team generate a hundred times more, humans cannot review it all, so slop wins by volume no matter what culture you build. Correct — if review means reading everything. It does not.

Manufacturing solved this a century ago: quality is controlled with sampling, tight specifications, and instruments — not by inspecting every unit by hand. The equivalent here is reviewing at the shipping boundary rather than the generation boundary, sampling deeply, and pushing the rest of your judgment into evals that run at machine speed. And the arithmetic hides a choice: nothing obliges you to ship everything the model produces. Teams drowning in their own AI output chose the volume before they chose the quality bar. Choose in the other order.