A million 'AI slop' reports in two weeks, and not one of them measures accuracy

Four questions hide inside the phrase "AI slop", and no shipped tool answers more than one of them. Authorship asks how a document was made, accuracy asks whether its claims hold, originality asks whether it adds information, and quality asks whether it helps a reader finish something ^[7]. An AI-content detector usually answers the first and stops ^[8]. Anthropic works the same problem from the generation side and lands in the same place: token selection is nudged so the text reads normally to a person while a detector can see the statistical pattern ^[12], which establishes that output probably came from a particular system and says very little about whether it is correct ^[13]. Two families of tooling, then, cover one of the four dimensions and leave the other three to whoever is assembling the context window ^[2].

The LinkedIn figure is worth holding at arm's length. More than a million clicks across fourteen days runs to something above seventy thousand a day ^[1], and the count is reports rather than verified generated posts or even unique users ^[3]. Reader standards differ, some responding to formatting or vocabulary, and writing polished with tools by a non-native English speaker can be read as machine output ^[9]. The button can also be worked by competitors, critics or coordinated groups ^[10]. What each click reliably records is that one post produced a negative quality judgement in one reader ^[11].

That is still the most interesting label in the story, because people catch things classifiers struggle with, such as a technical post with no working details or statistics with no identifiable source ^[19]. Both of those are retrieval failures before they are moderation failures. Index the second kind of page and you have added a document without adding a fact, which is the formulation to keep from the dev.to write-up: the web can contain millions of pages without containing millions of independent facts ^[6]. A pipeline that dedupes by URL and counts documents will read ten copies of one unsourced statistic as ten confirmations ^[3].

Which is why generated status is the wrong field to hang a ranking on, and belongs in the metadata instead ^[16]. Authorship fails as a proxy in both directions: a person can hand-publish an empty piece of familiar talking points, while an AI-assisted piece can carry original benchmarks, customer interviews and checked sources ^[18]. The provenance mark also decays exactly where a crawler meets it. Marked text gets shortened, translated, paraphrased or passed through another model before publication ^[14], and an empirical evaluation cited in the same piece found paraphrasing substantially weakened several watermarking approaches ^[15].

So the score that survives contact with this material is corroboration: how many mutually independent sources assert a claim, with any provenance marking treated as one input rather than the verdict ^[16]. The volume of slop reports is a message to people building search, RAG and agent systems that the failure sits upstream of the feed ^[17].