Can we judge text quality without knowing who wrote it?

The paper separates two questions that detection work tends to fuse: whether a machine wrote a text, and whether the text reads as slop. It says slop "differs from AI-text detection in general, and can be applied to any text source (whether AI-written or not)." It adds that "Human writing can also read as 'slop'", although it adopts the definition and focuses on "(seemingly) LLM-generated texts." Its stated aim is to characterize "qualities of texts that contribute to them being categorized as 'slop,'" which it suggests may explain cases where humans mistake human-written text for AI-generated text.

The reasoning turns on what the judgment is about. Detectors such as DetectGPT and Binoculars score the likelihood of AI origin and, as the paper cites them, "report high discriminant performance (0.95 AUROC)." The paper states that "our taxonomy and annotations diverge from those used for AI-text detection in general." Its target is still framed as "stylistic patterns unique to LLM writing," but the definition it works from is built on observable quality dimensions that any text can have. The binary judgments that anchor the work are made by annotators and, per the abstract, "correlate with latent dimensions such as coherence and relevance."

This sharpens Can human judges detect measurable differences in AI text?, which finds that LLM text differs measurably in lexical diversity and that human judges cannot identify it. Slop judgments are a different kind of human response: people make them, and they are framed around quality rather than origin. That is also the contrast with Can humans detect AI text if machines can measure it?. The two notes are about different targets of the same kind of measurement, origin in one case and quality in the other. The excerpt does not report whether its annotators could tell which texts were AI-written, so the comparison stops at the framing.

The excerpt does not establish how slop judgments relate to origin in practice. It reports no test of the framework against detectors, and it gives no figures on how often human-written text is judged slop. Its own target, LLM-specific style, sits in some tension with a definition that applies to any text, and the excerpt does not resolve that tension. The implication is that the category can be used to critique writing wherever it comes from, but a claim that a text judged slop was machine-written goes beyond what the excerpt supports.