How to — tell AI slop from writing worth reading

You have something in front of you — a newsletter draft, a vendor's "state of the industry" report, a freelancer's sample, your own first pass — and one decision to make: publish it, pay for it, forward it, or bin it. Five checks, about twenty minutes, and none of them involve guessing whether a machine wrote it.

That distinction is the whole method. Slop isn't a style, and it isn't a synonym for AI-assisted work. AI 101 — What is AI slop? defines it by two absences: nobody asked for it, and nobody checked it. Both are facts about process, which means both are observable. You don't have to inspect the prose for tells; you have to establish whether a person read this thing and is willing to answer for it. Style tells fail in both directions — plenty of human writers get accused of sounding generated, and plenty of generated pages pass because they read cleanly.

1. Read for claims, not for prose.

Go through with a pencil and mark every sentence that asserts something checkable: a number, a date, a name, a price, a causal claim, a quote attributed to someone. Then look at what you marked. Slop reads fluently and asserts vaguely — "many companies report", "studies show", "significantly faster", "experts agree". If you can strip every noun out of a paragraph and the paragraph still says the same thing, it said nothing.

The test is one question per paragraph: what do I know now that I didn't know before? If the honest answer is "that this topic is important", you're reading padding. Harvard Business Review has a name for the workplace version of this — "workslop", the polished, plausible document that arrives instead of an answer and pushes the checking cost onto whoever receives it, per HBR's write-up of the research. The reader pays for it either way, in money or in time.

2. Follow two claims all the way to a primary source.

Pick the claim the piece most depends on, and the one that would embarrass it if wrong. Then chase both to the thing itself: the paper, the filing, the model card, the dataset, the original announcement. Not the link in the sentence — the source behind the link.

The second hop is where assembled text dies. A press release cites a company blog, the blog cites a report, the report cited a survey nobody can find. If every citation in a piece is another article about the subject, nobody did the work; the piece is a summary of summaries, which is the structural signature of cheap generation. Two clean resolutions tell you a person went to the source. One dead link, or one number that doesn't appear where the piece says it does, tells you no one checked the rest either. If the piece is your own, this is the same discipline as checking an AI-generated citation before you rely on it.

3. Read the middle third and the last two hundred words.

Real writing accumulates. As a good piece goes on it collects specifics — names, versions, dates, dollar figures, what failed, who disagreed. Generated text does the reverse: it front-loads its strongest concrete detail, then restates the thesis in fresh vocabulary, then drifts toward implication.

So skip the opening, which is the easiest part to fake, and read where the writer has to deliver. Does the middle get more specific or less? Then read the ending, where a writer has to commit to something. If the last paragraph is "the future remains uncertain", "only time will tell", or a paraphrase of the introduction, the piece never had a destination — it had a length.

4. Ask one question only a person who did the work could answer.

This is the cheapest reliable test, and it works on colleagues, vendors and freelancers alike. Not "did you use AI?" — that question is unanswerable, and the answer distorts your judgment anyway. In a controlled study with 556 participants, people rated identical passages about 14 percentage points lower once told a machine wrote them, regardless of quality, per Haverals and Martin's study of attribution bias. Labels make you a worse judge, so don't collect one.

Ask instead: how did you get that number? What did you leave out? What surprised you? What did you try that didn't work? A person answers with detail, and usually with some irritation about the part you didn't ask about. A pipeline answers by restating the piece in shorter sentences.

5. Find the human who answers for it.

Look at the artifacts around the text rather than the text: a named byline with a person behind it, a date, sources linked where the claim sits, a way to send a correction, a stated method. For a research-flavored report, that means a methodology note — who was surveyed, how many, who paid for it. For a newsletter, a stated policy on how AI is used and what gets checked.

Absence is informative. No byline, no date, no method, no correction path, and a contact form that goes nowhere: that's a page with no one behind it. A process question is answerable in a way an authorship question never is, which is why it's the one worth asking.

Don't do this: don't run the piece through an AI detector and treat the score as a verdict.

Detectors answer the question you don't need answered — was a machine involved — and they answer it badly. A widely cited evaluation by Liang and colleagues at Stanford found that popular GPT detectors consistently misclassified writing by non-native English speakers as AI-generated while scoring native-speaker text accurately, and that light prompting was enough to slip past them, published in Patterns. So the tool's errors land hardest on careful writers with a smaller English vocabulary, and a clean score proves nothing about whether the text is true.

You are not the authorship police. You are deciding whether a piece is worth the slot it occupies, and that question is settled by evidence and accountability, not by a percentage.

How you'll know it worked: you can say three things without hedging. What this piece knows that you didn't before you read it. Which claim you verified at its source, and where. Who is answerable if it turns out to be wrong. If the first is vague, it's padding; if the second is missing, it's unverified; if the third is nobody, it's slop wearing a serif font — send it back with your question, or bin it and move on. The twenty minutes you spend is the entire difference between publishing something and publishing anything.

What's the tell that makes you stop reading first — a number with no source, or a paragraph that says nothing? Write to us via our contact page.