What dimensions make text feel like AI slop?

The paper treats AI "slop" as a construct to decompose rather than a verdict. It adopts the Oxford Dictionary's definition of low-quality, LLM-produced material, then collects written definitions from 19 people across writing, journalism, linguistics, NLP and philosophy. The abstract calls these responses interviews; the method section describes survey responses. Deductive coding of those responses yields three axes. Information Utility combines Density, measured with token entropy and propositional idea density, and Relevance, judged by human annotators. Information Quality combines Factuality, which the paper says needs human annotation, and Bias (subjectivity), measured as the proportion of subjective words in a lexicon. Style Quality is read through Repetition (lexical repetition metrics) and Templatedness (syntactic structure).

The authors argue that slop "does not immediately permit measurement," because "low-quality" and "unwanted" are hard to quantify. They therefore propose "a composite measure over observable characteristics of text," elicited from people with relevant expertise, and offer the axes as a framework for "assessing writing across domains, beyond accuracy- or reference-based metrics." The paper also says granular codes "can vary in strength based on the domain, or the purpose of the text." The excerpt does not give the final code count after redundant codes were collapsed, or the tallies in its Table 1.

Against the nearest notes, this is a sharper version of the style-versus-substance argument. Does polished AI output trick audiences into trusting it? warns that presentation can stand in for expert judgment; the taxonomy keeps Style Quality as one axis beside Information Quality and Utility rather than letting it carry the whole judgment. Can imitating ChatGPT fool evaluators into thinking models improved? shows a related split in a different setting, with imitation matching surface style while the factuality gap stays open. The density axis sits near Can we measure reading efficiency as a quality metric?, though it uses token entropy and propositional idea density rather than an atomic-unit count. The paper's finding that neither LLM judges nor linear models "fully approximate" human slop assessments fits How much does rhetorical style shift AI review scores?, which shows LLM reviewers responding to rhetorical presentation.

The excerpt does not report the span-level annotation results behind its headline claim. The abstract says binary slop judgments are "(somewhat) subjective" but correlate with latent dimensions such as coherence and relevance; the excerpt gives no agreement statistic, correlation coefficient or annotation protocol, and the 150-article study appears only in the contributions list. The taxonomy rests on a small survey, all but one of whose respondents had three or more years' experience in their field. The implication is that the axes are a usable design space for critiquing text, while automatic scoring of slop should wait for the evidence the excerpt leaves out.