Slop Is a Process Property

The term "AI slop" is commonly treated as a property of the generator: text is slop because a model produced it. This paper argues that slop is a property of the process. It names generation that reached a reader without the review it needed, and reviewed generation is a draft, whoever generated it. On that account the slop critique is largely mitigable by review processes run with or by language models, but not on the grounds usually offered. The argument is not that models review better than they write, nor that they review better than humans. It is that one large class of faults, provenance, is invisible to a reading human at population scale and cheap for a machine to check: across four top-tier venues, papers carrying hallucinated citations received reviewer ratings indistinguishable from clean papers, accepted and rejected submissions whose ratings differed by two points carried the same share of them, and a venue-scale audit cost about four cents a paper (Russinovich, Siva Kumar and Salem 2026). Review is the better-suited task for a model for five structural reasons: a fixed target, a brief that says where to look, permission to answer "cannot verify", no stake in the document, and the generated text as a map of where to check. The remedy for slop is therefore a division of labour that assigns the provenance pass to the machine and the reasoning pass to the human. The paper names five mechanisms by which unreviewed generation acquires faults that are invisible from inside the text, carries four corrections as explicit limits, and reports one case in which two reviewers of different kind, a human on a phone and a model under a provenance brief, found disjoint error classes with no overlap, every provenance fault sitting on a source the generating model had not opened. The case is N = 1 and is offered as a specimen, not a result; the paper's own first draft supplies a second specimen, by the mechanism the paper describes. The division of labour is then placed under its binding constraint: when an AI agent is institutionally positioned as an employee, managers catch fewer errors themselves and pass the review on (Wiles, Hsu, Bedard and Kropp 2026), so the design must survive the effect it is meant to offset. Seven hypotheses are stated, each with the observation that would falsify it.