Start with the uncomfortable part: you cannot reliably tell what wrote a text by reading it. The research on this is unambiguous. Expert annotators disagree about borderline cases, humans misclassify human writing as machine writing, and the paper that built the most careful taxonomy for measuring slop in text concluded that scalable automated detection remains an open problem.
What you can judge is quality, effort and accountability. Those are the things the word slop actually points at, and they are observable. Here are ten signs, ordered by how much they tell you.
The ten signs
1. Generic specificity. The page is full of confident language and contains no names, dates, numbers, documents or places. Phrases like “many experts agree” and “studies show” appear without a study. This is the single strongest signal, because it is the hardest tell to fake and the easiest to check.
2. Uniform rhythm. Sentences arrive at roughly the same length, paragraph after paragraph, with the same internal shape. Human writing tends to breathe unevenly. Read a paragraph out loud and notice whether your voice settles into a metronome.
3. Structure standing in for substance. Three sections that make the same point in different words. A list of five items where items three and four restate item one. Formatting creates the impression of an argument without an argument having been made.
4. The answer is longer than the question deserves. You asked what a term means. You received an overview, a history, five benefits, a comparison table and a conclusion. Length is cheap to generate and expensive to read.
5. Confidence above the pay grade. Definitive claims about contested questions, stated without hedging, in a context where a specialist would hedge. The hedge-then-assert pattern can run the other way too: four qualifiers followed by a firm conclusion the qualifiers do not support.
6. Repetition dressed as emphasis. The same sentence structure recycled with synonyms. This is the visual cousin of uniform rhythm, and it shows up reliably in generated listicles.
7. Facts that dissolve on contact. You recognize a name and cannot find the person. You look up the statistic and it appears nowhere outside this page. You check the citation and the source does not contain the claim. Fabricated citations in legal filings are the most consequential version of this.
8. Tone mismatch with the venue. A product launch voice used for a death notice. A landing page structure used for a reference article. A cheerful summary of something grim. The content has not been adapted to where it is, which is what happens when nobody adapted it.
9. The effort ratio is upside down. You have spent twenty minutes checking something that took two minutes to produce, and you are still not sure. This is the experience the word was coined to describe, and it is worth trusting.
10. Provenance is missing. The page carries no author name, date or contact details, and there is no stated policy for corrections. The site's other pages follow the same template with the noun swapped. A publication that will not say who is accountable for a page is telling you something.
The tells that are not real tells
Four popular tests that produce more false accusations than true ones:
| Claimed tell | Why it fails |
|---|---|
| Em dashes | Humans who write for a living use them constantly. Machine output often avoids them. |
| Specific “AI words” | Word frequency shifts across tools and versions. Judging a person on one word is not analysis. |
| Perfect grammar and spelling | Clean copy is what an editor produces. This test punishes good work. |
| Long, formal sentences | Academic and legal writing looks like this and always has. |
There is a version of this argument that matters: the paper on measuring slop does note that social media discussion of the topic treats overuse of certain terms, low information density and list-shaped answers as indicators. The lesson is not that those features prove anything. It is that people are pattern-matching to surface features because the underlying question is genuinely hard.
The 60-second check you can actually run
Instead of guessing at authorship, ask three questions.
- Who stands behind this? A person's name, an organization, an email, a policy for corrections. If the answer is nobody, that is your finding, regardless of who wrote it.
- What is the specific claim? Strip the adjectives and try to write the load-bearing claim in one sentence with a number or a name in it. If you cannot, there is nothing to verify and nothing to use.
- Where would I check it? If the page does not tell you, and a search does not surface the source, treat the claim as unavailable. Not false. Unavailable.
Three questions, and none of them require you to know anything about language models.
Why the question is harder than it looks
Two facts pull in opposite directions.
Human writing can read as slop. A rushed human page stuffed with filler is slop in every way that matters to a reader, and the fact that a person typed it changes nothing about the cost to you.
Machine output can be good. An article drafted with a model, then cut, checked, sourced and rewritten by an editor, is not what the term describes. The definitions were written that way on purpose: the writers who popularized the word said explicitly that not all AI-generated content is slop.
That is why the useful question is not “was this written by AI”. It is “was anyone accountable for this before it reached me”, and its two follow-ups: what does it tell me that I could not have guessed, and what happens to whoever published it if it turns out to be wrong?
What to do when you find it
Mostly, nothing dramatic. Do not share it, which is the only currency it wants. If it is a factual claim in something you rely on, check it at the source and note that the page needed checking. If a company sent it to you, telling them is more effective than complaining publicly, because the sender has an incentive to fix the process and the platform does not.
The exception is harm. Fabricated photos attached to real events, invented citations in official documents, and confidently wrong advice about health or safety are worth reporting to whoever can act on them. That is a different problem from a boring article, and it deserves a different response.