Essay · July 21, 2026 · 9 min read
AI Slop: What It Means, and How to Not Produce It
"Slop" has jumped from niche internet slang into everyday speech. By 2025 it was showing up on word-of-the-year shortlists and, more tellingly, in ordinary conversation — it's the word your boss, your client, or a stranger reaches for to dismiss something in one syllable. If you use AI to help you write, design, or ship work, the fear underneath the term is real: that your output looks like the flood of generic stuff everyone is learning to scroll past. The reassuring part is that slop is a fixable problem, and the fix is mostly in your hands.
What is 'AI slop', actually?
The term was popularized by developer Simon Willison in May 2024. He originally proposed "slop" for unwanted AI-generated content, later sharpening it to material that is unrequested and unreviewed — "mindlessly generated and thrust upon someone who didn't ask for it." That's a useful test, not a complete definition. It isn't "content made with AI"; it's a claim about how a thing was made and pushed onto people, not whether a model touched it.
The test has limits, and they matter in an article like this. Requested, reviewed work can still be slop: merely running your eyes over model output isn't meaningful review, and being asked for something doesn't stop it from being hollow. So it helps to hold several markers at once — low substance (nothing checkable, no real point), displaced effort (cheap to make, expensive to read), missing verification (claims nobody confirmed), and indifference to the recipient (shipped without a thought for the person on the other end). One much-repeated formulation captures the cost: slop is content that "takes more human effort to consume than it took to produce." When a coworker forwards you raw model output, they're not creating — they're quietly moving their work onto your desk.
Hold onto that frame, because it's the one that helps a worried person the most. The tool is not the villain. Slop is a decision — the decision to not add anything, not verify, and ship the default anyway.
How do you spot AI slop (the tells)?
Most people hunt for lexical tells: "delve," "boasts," "rich tapestry," "in today's fast-paced world," "it's not just X, it's Y," heavy em-dash use. Those still exist, but by 2026 they're weak signals. Newer models were tuned away from them, and plenty of humans now nervously strip em-dashes out of their own real writing. You can remove every "delve" and still ship pure slop. So learn the structural tells instead — they last.
- The everything-sandwich. An intro that restates the prompt, three-to-five evenly weighted bullet sections that each state the obvious, and an "in conclusion, it's important to remember" wrap. Every paragraph is the same length. No paragraph carries more weight than any other, because no human chose what mattered.
- Confident hollowness. Fluent sentences that assert nothing checkable. No numbers, no names, no dated example, no opinion that could turn out wrong. Slop never risks being wrong because it never commits to anything.
- Both-sides on everything. "While there are benefits, there are also challenges." A person with an actual point picks a side.
- Zero friction, zero surprise. You can predict the next sentence. Real expertise has texture — a caveat, an aside, a "the common advice here is actually wrong."
- Un-anchored claims. Advice with no source, no "I tried this and it broke," no failure mode.
For images and video the tells are more physical: mangled hands and extra fingers, garbled embedded text, impossible physics, the plasticky over-lit AI sheen, and pure context absurdity — the "Shrimp Jesus" Facebook genre, or the fake "Willy's Chocolate Experience" whose AI-generated promo images lured families into a nearly empty Glasgow warehouse in 2024. Coca-Cola's 2024 AI-generated holiday ads drew heavy criticism for the same uncanny, slightly-off quality, even from viewers who couldn't name why they disliked them.
The real giveaway is often behavioral: volume and speed (one account posting 30 "guides" a day), fabricated authority, and citations to things that don't exist. Fabricated authority is worth a concrete example — in 2023 the writer Jane Friedman found several books she never wrote being sold under her name on Amazon and Goodreads, apparently AI-generated; the same year, AI mushroom-foraging guides with invented authors and genuinely dangerous advice turned up in online stores. And in early 2026 the curl project ended its long-running bug-bounty program after mounting low-quality, often AI-assisted vulnerability reports ate maintainer time — fluent, confident write-ups where a batch reviewed before the February 1, 2026 cutoff turned up no concrete vulnerabilities at all.
Here's the meta-test to run on your own draft. Don't ask "does this sound like AI." Ask: did this cost me nothing to make, and will it cost the reader something to get through? And: is there one specific, checkable, could-be-wrong claim in here that's genuinely mine? If the answer to the first is yes and the second is no, it's slop.
Why does so much AI output end up as slop?
Three separate engines, worth pulling apart because only two are your problem.
The economic flood. Platforms pay per engagement, so volume beats quality by design. A creator in a low-wage economy can net real money from AI images tuned for the feed. Music fraud runs the same play: in 2024, US prosecutors charged Michael Smith with using AI-generated songs and armies of bots to inflate streams and collect more than $10 million in royalties. Deezer reported in 2025 that a fast-rising share of the tracks uploaded to it each day — by its own estimate, close to a fifth by mid-year — were fully AI-generated, and that much of that AI music was tied to streaming fraud. This is deliberate slop, and it isn't the thing you're anxious about. But it's why the word carries such contempt.
The default-output problem — the accidental slop. Without strong context or constraints, general-purpose models tend to produce conventional, broadly acceptable answers: safe, balanced, comprehensive, generic. Left unsteered, a model regresses toward the average of its training data. Generic is the default; specific is the thing you have to ask for and add yourself. Most accidental slop is just someone shipping the first draft, which is by design close to the most average possible version of the thing.
Workslop — the workplace engine, and the one with data. A 2025 study from BetterUp Labs and Stanford's Social Media Lab surveyed roughly 1,150 US full-time employees. About 40% had received "workslop" — polished-looking AI output that masquerades as finished work — in the previous month. Each incident cost around two hours to untangle, which the researchers put at roughly $186 per employee per month, or about $9 million a year for a 10,000-person organization. The mechanism is the uncomfortable part: AI makes it cheap to look productive, so people ship convincing non-work and push the actual thinking downstream to whoever has to read it.
The through-line across all three: slop happens when someone lets the tool's default out the door without adding the one thing a model can't supply — a specific point of view, verified facts, and a judgment about what to cut.
How do you use AI without producing slop?
Ranked by leverage, roughly highest first.
- Bring the specifics AI can't invent. Feed it your real numbers, the thing that actually happened, the opinion you'd defend in a room. Those specifics — a true detail, a real figure, a take you'd stand behind — are what a model cannot supply for you, and they're the single biggest thing separating your draft from the default. This is the closest thing to the whole game.
- Steer away from the mean. "Write about X" gets you the average of the internet. Give it constraints instead: who it's for, the one thing it must land, what to leave out, a position to argue. "Argue that most advice on X is wrong, for readers who already tried the obvious, using this example, and cut anything a smart reader already knows" gets you something with a spine.
- Review — it's literally the definition. Unreviewed is a large part of what makes slop slop, and real review is more than skimming. Fact-check every checkable claim (models still fabricate with total confidence — the curl reports are the reminder), cut everything generic, and confirm you could personally defend each sentence. If you can't verify a claim, delete it. Don't ship it hedged.
- Cut hard, and kill the sandwich. Delete the restate-the-prompt intro and the "in conclusion" wrap. Make one paragraph obviously carry more weight than the rest. Uniform reads as slop; asymmetric emphasis is the fingerprint of a human who decided what mattered.
- Add friction on purpose. One surprising claim, one caveat, one "the common advice is wrong," one thing you're willing to be wrong about. Slop is frictionless. Expertise has edges.
- Edit late rather than prompt forever. Get a fast, rough draft and rewrite it hard, instead of chasing the perfect prompt. Your judgment is the anti-slop ingredient, and you can only apply judgment to output that actually exists.
The honest tradeoff: doing this well erases most of the time you thought you were saving on the writing. The gain moves upstream (getting unstuck, structure, breadth) and downstream you still own the review. Anyone selling "AI writes it, you ship it" is selling you the slop pipeline.
A few mistakes worth naming out loud: shipping the first draft; mistaking "sounds fluent" for "is good"; obsessing over lexical tells while leaving the hollow structure intact; using AI for the part that needed your judgment (the take) while hand-doing the part AI is actually good at; and disclosure theater — a "made with AI" badge doesn't turn unreviewed output into not-slop.
Does how you present and ship the work change how it's judged?
Yes — but not in the way people hope, and being honest about the limit is the whole point.
Presentation is a real judgment signal. Reviewers use "did they bother" as a proxy for effort, and effort is exactly what slop lacks. Two identical drafts — one a raw pasted block, one a clean, formatted, owned page — get judged differently, because the packaging signals a human cared enough to package it. How something arrives is also part of how people infer whether you reviewed it: a forwarded raw dump signals you didn't look; a finished, self-contained artifact signals ownership. That's the same asymmetric-effort axis slop is defined on, applied to the output side.
But here's the limit, stated flatly: presentation cannot rescue hollow content, and pretending it can is itself a slop move. Polishing empty work is precisely the workslop failure mode — something engineered to look finished that isn't. A beautiful wrapper on generic substance reads as more cynical, not less; the gloss makes the hollowness louder. We've all watched a slickly produced AI ad get dunked on precisely because it was well-made and empty. Presentation is a multiplier, not a substitute: multiply real substance and you win; multiply slop and you've just amplified the contempt.
That's the whole thing, really. Slop isn't a formatting problem or a which-model problem. It's what happens when nobody adds the specific, verified, willing-to-be-wrong judgment that a reader's time deserves. Add that, and it doesn't matter that a model helped. Skip it, and no wrapper will save you.
Go deeper: Simon Willison's original May 2024 post is where the definition starts. And the HBR "workslop" piece, written by the researchers behind the BetterUp Labs / Stanford study, lays out the workplace stakes and the numbers behind them.
Frequently asked
What does 'AI slop' mean? +
AI slop is AI-generated content that's unrequested and unreviewed — the definition from Simon Willison, who popularized the term. It isn't 'anything made with AI'; it's about how the output was deployed. If a human asked for it and actually read it before sending, it isn't slop, even if a model wrote every word.
How do you spot AI slop? +
The durable tells are structural, not lexical: the everything-sandwich (evenly weighted, obvious bullets), confident hollowness (no numbers, names, or checkable claims), both-sides-on-everything, and zero surprise. For images, look for mangled hands, garbled text, and that plasticky over-lit sheen.
Why does AI produce slop? +
Models are trained to produce the median acceptable answer, so generic is the default and specific is what you have to add. Combine that with platforms paying for volume and 'workslop' — looking productive by shipping unreviewed output — and slop becomes the path of least resistance.
How do you use AI without producing slop? +
Bring the specifics a model can't invent — your real numbers, a real example, an opinion you'd defend — steer it away from the generic default, and actually review it. Unreviewed is the literal definition of slop, so review is non-negotiable.
Is anything made with AI automatically slop? +
No. Slop is a decision — shipping the unreviewed default — not a tool. AI-assisted work with a specific point of view, verified facts, and human judgment about what to cut isn't slop, regardless of who typed the words.
Related: AI proposal & deck generators · how to make a one-pager
Ship work that doesn't get dismissed.
Do the human part — the specifics, the review, the point of view. And when you've built something good as an HTML page, Pagelive gives it a private, branded, tracked link. Free for 5 pages, no card.