AI slop is a good phrase because nobody needs it explained. You know it on sight. The article that answers its own headline in the first line and then says the same thing four more ways. The listicle where all ten entries have the identical three-clause shape. The product page describing a category instead of a product.
Somewhere in the last two years it stopped being a description and started being an explanation, and that's where it went wrong. Slop gets talked about as a property of the machine. Something the model does to you, the way a printer jams. Read it that way and there's nothing to do but wait for a better model, or refuse to use one at all.
But no model has ever published anything. A person generated that page, read it or didn't read it, and pressed a button. Every step after the first one was a person.
What the argument rests on
The first figure is an industry number, taken from independent write-ups published in 2026 that agree with each other, checked on 11 August 2026. The other five are mine and were counted from the files on 18 August 2026 rather than recalled. Two of them had drifted since this piece was published and one was never right. See the foot of the page.
The Model Is Not the Variable
The same model that produces slop produced this. Nothing about the model changed between the two. What changed is everything wrapped around the request: what it was told about the reader, what it wasn't allowed to write, what it had to check before it could claim anything, and what happened to the draft once it existed.
Ungoverned generation has a signature because it's a system doing its default thing, which is producing the most plausible next sentence. Plausible is the word. It doesn't mean true and it doesn't mean worth reading, and a process that asks for nothing else gets prose that's fluent, evenly weighted and completely hollow.
A model is a drafting engine. Slop is what a drafting engine produces when nothing downstream of it has the authority to say no.
Which makes the useful question a different one. Not whether AI wrote it. What was standing between the draft and the publish button, and what it was allowed to throw out.
What AI Slop Actually Looks Like
Naming the symptoms matters, because a team that can't describe the failure can't check for it. Five properties, and a piece only needs three of them to read as machine output.
| Property | How it shows up | Checkable? |
|---|---|---|
| Uniform rhythm | Every sentence the same length, every paragraph the same shape | Yes |
| Vocabulary tells | Delve, tapestry, seamless, a testament to | Yes |
| Structural tells | Three-item lists, not just X but Y, summary closers | Yes |
| No position | Balanced on every question, committed on none | Partly |
| Nothing checkable | Paragraphs with no number, date, name or mechanism in them | Partly |
The top three are why a script is worth writing. The bottom two are why an editor still is.
The Last Two Are the Real Ones
A piece can pass every mechanical check and still be slop, because vocabulary and rhythm are surface properties. What actually makes writing worthless is having no argument, and no machine has ever detected that.
Which is the honest limit of everything below. The checks strip the tells out. They don't hand you a point of view, and a page with no point of view is slop with good punctuation.
Why Everybody's Output Converges
There's a mechanical reason competing sites now read identically, and it's nothing to do with the model being bad at writing.
Three companies in the same sector use the same tool, type near-identical prompts and ask for near-identical things. A seven-hundred-word post on a topic, professional tone, general audience. Give a system optimising for the most plausible next sentence the same request and it returns the most plausible next sentence, which is the average one by definition.
The Model Does Not Know Who You Are
It can't see your positioning, your customers, the argument you've been making for six years, or the thing your founder believes that your competitors think is daft. Without any of that, it writes for the average company in your category.
Which is why the fix isn't a better prompt. A better prompt gets you a better average. The fix is handing over the information that makes your output non-average, in a form the tool reads every time, rather than in a briefing call somebody half-remembers.
What Governs the Sentence You Are Reading
Four plain text files sit above this article. Between them they run to just under eleven thousand words, and not one of them contains any part of the argument you're reading. They're the constitution, not the content.
- WRITING-STANDARD.md, 5,980 words. Voice, banned patterns, fact discipline, structure. The long one, because it's the one doing the work.
- ULTIMATE-POST-FIX.md, 2,262 words. The protocol for repairing a page that already exists, which is a different job from writing one and fails in different places.
- SEO-POST-AUDIT.md, 1,769 words. A scored audit that's not allowed to change anything. Keeping the marking away from the writing is most of why it's worth having.
- CLAUDE.md, 937 words. The brief for this site: what it is, what's settled and can't be reopened, and the traps that are written down because each one has already caused a problem once.
The Rule That Outranks the Others
Section zero of the standard is one line long. Every claim has to survive an interview. If a reader stopped at any sentence here and asked for more, there has to be a real answer waiting. If the honest answer is a shuffle, the line comes out.
That rule cuts more copy than the rest of the document put together, and it's the one most content operations don't have, because no model can satisfy it alone. Something in the system has to know whether the answer actually exists.
Fifty-Nine Words That May Not Appear
The banned list is specific rather than tasteful. Delve, tapestry, seamless, robust, leverage, a testament to, unlock the potential. Fifty-eight entries, alongside fourteen structural patterns: the three-item parallel list, the not just X but Y construction, the summary sentence that ends a paragraph by restating the paragraph.
Em dashes are banned outright, right down to the separator in the page title. That one isn't a craft judgement. It's the most-cited punctuation tell going, and on a site arguing that I know what machine output looks like, shipping it costs something and buys nothing.
The Numbers Are a Floor, Not a Target
Average sentence length has to land between fourteen and twenty-two words. Two sentences per fifteen hundred have to run past thirty-five, because uniform rhythm is the loudest tell of the lot and the one that survives every vocabulary fix. Three fragments, minimum. Reading ease above forty-five.
A 129-line script checks every one of those before the file gets called finished. This article failed it twice on the first pass, for a whole sentence set in bold and a phrase that tripped one of the structural patterns. Small things. That's rather the point of spending a machine's attention on them.
What Google Actually Penalises
Worth being precise here, because the fear is vaguer than the rule and that vagueness sends teams to the wrong response.
Google's published spam policies name scaled content abuse: generating many pages primarily to manipulate rankings rather than to help people. The policy turns on purpose and quality, and it applies whether the pages were produced by a machine, by people, or by both.
So authorship isn't the trigger. A thousand governed pages that each answer a real question sit outside the policy. A hundred ungoverned ones restating the same average paragraph sit inside it, and a human byline changes nothing.
Which Makes Slop a Commercial Risk, Not Only an Aesthetic One
That reframing matters when you're arguing for budget. AI slop isn't just embarrassing. It's the observable property the policy describes, produced by the process the policy exists to catch, and the fix is the same governance that makes the writing better anyway.
A second commercial cost turns up sooner and hurts more quietly. Over half of consumers back off the moment they suspect a machine wrote what they're reading, so the penalty lands with your audience long before it lands with a search engine.
The Fix, in Order
The order matters more than the tooling, and doing these out of order is why most attempts produce a fortnight of enthusiasm and no change.
- Write down what only you knowPositioning, customer, the argument you have been making for years, the thing competitors get wrong. No model has access to any of it. The step teams skip
- Write down what may never appearStart with twenty words and phrases. You already know most of them, because they make you wince in a competitor's copy. An afternoon
- Automate the checkable halfVocabulary, structural patterns, rhythm bands, link integrity. A short script beats a long document because it runs whether or not anyone remembers.Mechanical
- Make it blockA warning is a style guide with extra steps. Until something stops publication you have documentation, not a control.The dividing line
- Keep an editor on the restArgument, position, whether the piece is worth publishing at all. No script reaches this, and it decides whether the output is any good. Human, permanently
Steps two to four remove the tells. Step one is what makes the writing yours, and doing these out of order is why most attempts produce a fortnight of enthusiasm and no change.
Step one is the one teams skip, and it's the only one that touches the convergence problem. Steps two to four strip the tells out. Step one is what makes the writing yours.
How to Audit an Archive for AI Slop
Most teams asking me about AI slop already have several hundred pages of it live and no idea which ones. The instinct is to run a detector over the archive and start at the top of the list. That gives you a queue sorted by the wrong thing, because a detector scores a probability and what you need is a queue sorted by damage.
Sort by traffic instead. A slop page nobody visits is an embarrassment. A slop page taking two thousand sessions a month is a commercial problem, and it's the one where a rewrite pays for itself inside a quarter.
Run a second sort alongside it, by exposure. The pages a regulator, a partner or a prospective client would plausibly read. Those are rarely the high-traffic ones. A dormant policy page carries more reputational load per visit than anything in your top ten, and it's exactly the kind of page that gets generated once and never opened again by anyone inside the business.
Four Passes, in This Order
The first pass is mechanical and takes an afternoon on a mid-sized archive. Run the banned list across every page and count hits per thousand words. You're not looking for pages to delete. You're looking for the shape of the distribution, because AI slop is rarely spread evenly. It clusters by author, by month, or by the week somebody picked up a new tool.
Second pass, take the top decile by hits and read ten pages properly. This is where you find out whether the vocabulary problem is the whole problem or the visible edge of a bigger one. Every archive I've opened, it's been the edge.
Third pass is the expensive one and there's no way round it. Take every page in the high traffic band and check the numbers in it. Not the prose. The numbers. The figure that turns out to be wrong is almost never the headline statistic that got fact-checked. It's the incidental one in paragraph nine that nobody thought needed it.
A page that reads beautifully and contains a false number is worse than AI slop, because it survives review.
Fourth pass, and only now, decide what happens to each page. Rewrite, consolidate, redirect or leave. Leaving is a perfectly good answer for a page with no traffic and nothing false in it, and a programme that can't say leave it will spend six months rewriting things nobody reads.
What the Audit Turns Up
Three patterns come up again and again. A cluster of pages published in one narrow window, all the same shape, which is somebody trialling a tool without telling anyone. A set where the vocabulary is clean and the argument is missing, which is a writer who learned the banned list and nothing else. And a long tail of pages that are fine.
That last group matters more than it sounds. If a slop audit condemns everything, what's being applied is taste rather than a control, and nobody in the building will trust the next one.
The Objections Worth Taking Seriously
Two arguments against all of this hold up, and pretending otherwise would be the same failure this article is about.
A Standard Can Ossify
A banned list written in 2026 encodes what reads as AI slop in 2026. Language moves, models move faster, and a list nobody revisits turns into a set of arbitrary prohibitions a new writer follows without understanding why. I go back to mine when a rule fires against writing that's good, which happens about monthly and is the only reliable sign a rule has aged.
The related risk is the list starts banning ordinary words because they turned up in one bad draft. Down that road the vocabulary narrows until every piece sounds identical, which is the original problem wearing a different hat.
It Can Become Compliance Theatre
A green tick is enormously satisfying and means very little on its own. The failure is a team that runs the checker, sees it pass, and skips the reading. I've watched it happen and what came out was cleaner AI slop. No banned words in it. No argument in it either.
The guard is that scoring is never the last gate. A human sign-off sits after the script, not before it, and a piece that passes every check can still be killed for being dull. If the person doing that sign-off is measured on throughput, the guard is decorative.
The One That Does Not Hold
The objection I hear most is that all this is overhead a small team can't carry, and it's the weakest of the three. The whole apparatus on this site is four files and one script. The files took a few days spread over months, most of it spent making decisions any content lead already carries around unwritten.
Set that against the cost of one wrong number reaching a regulated page, or one client noticing three of your articles open the same way. The overhead argument only works if you assume the alternative is free, and the alternative is where AI slop comes from.
The Part That Does Not Automate
Mechanical checks have a hard ceiling and it arrives early. A script can prove a sentence contains the word tapestry. It has nothing whatsoever to say about whether the paragraph is true.
So verification runs as its own protocol, and it's the slowest thing in the process by a distance. Disputable claims get listed before drafting starts. Comparative claims fail safe: any largest, first or fastest that hasn't been checked against named alternatives in the same session gets dropped or swapped for a plain figure.
That rule is in the document because of one specific failure. A review shipped calling an operator's game library the largest in its segment. Three competitors were bigger, and the operator's own published count was sixty per cent higher than the number in the piece. Both errors were about a minute of checking away, and nobody had a minute.
Somebody Has to Sign It
I read everything before it goes out. Not as a formality. It's the step where the thing becomes mine rather than the tool's, which matters because when a claim on this site is wrong it's my name at the top of the page, not a vendor's.
The pipeline doesn't remove that step. It clears enough mechanical work out of the way that there's attention left to do it properly, which is the argument I've made about rules that can fail a build and the one thing all of this is actually for.
What It Actually Costs
There's a version of this article that ends by claiming the best content money can buy, for no money. Tempting line, and false twice over, which is why the standard forbids it. There's a subscription. There are the hours that went into eleven thousand words of governing documents, and the further hours of keeping them honest every time a rule turns out to be wrong in practice.
The true claim is narrower and more useful. Drafting was never the expensive part of editorial work. The expensive part was the checking, the sourcing and the argument about whether a claim holds, and that cost hasn't gone anywhere. Some of it became automatable, which isn't the same as becoming free, and what refused to automate is now the entire job.
So when something reads like slop, the model is the least interesting suspect in the room. Ask what it was told. Ask what it wasn't allowed to write. Ask what checked the draft, what happened when a check failed, and whether anyone could switch it off at four o'clock on a Friday. Nobody has ever published slop by accident. They published it because nothing in the building was allowed to stop them.
Where This Standard Came From
Worth saying, because a standard invented in the abstract is a wish list and one built out of failures is a control. Mine is the second kind and it shows in what's in it.
It got merged out of three separate content operations: a firearms retail site, a gambling affiliate property and a regulated casino review estate. Three different voices, three commercial models, three sets of things that had already gone wrong. Where all three agreed, the rule went into the shared standard. Where they disagreed, the difference went into a per-site profile and the craft stayed common.
The Rules That Surprised Me
Several arrived from incidents rather than taste. The ban on comparative claims that haven't been checked against named alternatives came from a review that shipped calling an operator's game library the largest in its segment when three competitors were bigger. That rule has never been relaxed.
The contraction band came from a firearms buying guide, and I decided it was wrong for essays and rewrote the rule to say this site ran at zero per cent on purpose. That lasted a week. Put in front of the person who owns the site, the same prose came back as too stuffy, and the imported band was right after all. Overruling a rule using only the archive it governs is how you end up defending a decision nobody ever made.
Why It Is Four Files and Not One
A single long document gets read once, at induction, by somebody who then works from memory. Same failure as a style guide, just bigger, and splitting the standard was the cheapest fix going.
Four short files with distinct jobs actually get opened, because you know which one holds the answer. Voice and patterns in one, fact discipline in another, the repair protocol in a third, the site brief in the fourth. Nothing clever about it. It's the difference between a standard that gets read and one that gets cited in a meeting by somebody who hasn't opened it since March.
The split has a second benefit I didn't plan for. Because the craft sits in shared files and only the site profile changes, improving a rule improves it everywhere at once. Something learned on a firearms retail page lands on a casino review the same afternoon, which isn't how a per-client style guide behaves.
The Two Checks This Article Failed
The figure in the box at the top is real and worth unpacking, because a claim about governance is worth more with the receipts attached.
On its first pass this piece failed twice. Once for a whole sentence set in bold, which §2.4 forbids because bolding a sentence is emphasis applied without deciding what's load-bearing. Once for a phrase that tripped the from-X-to-Y pattern, one of the fourteen structural tells.
Neither would have been caught by a person reading carefully, and neither matters much on its own. Which is exactly the argument for spending a machine's attention on them. Too small to justify human time, too numerous to ignore in aggregate.
And One the Script Did Not Catch
An earlier version of the disclosure on this page said the article had failed seven checks. It had failed two. The seven was invented, in a note about fact discipline, on an article about ungoverned generation.
No script caught it, because no script could. It got caught by reading the claim back against what the checker had actually printed, which is the human half of the process doing the exact job it exists for. I've left it in the disclosure rather than quietly correcting it, because an argument about governance that hides its own near-miss is worth nothing.
And it happened again. Three of the six figures in the box at the top were wrong when I re-counted them on 18 August 2026. Two had drifted as the files grew, which is ordinary decay. The third, a claim that the checker ran to 343 lines, was never true of anything. It runs to 124. Nobody had counted, on a page whose whole argument is that somebody should.
Questions I Get Asked
Is AI Slop Just a Snobbish Word for AI Content?
No, and collapsing the two is what stalls the conversation. Slop describes a specific set of observable properties: uniform rhythm, vocabulary tells, no position, nothing checkable. Plenty of AI-assisted writing has none of them, and plenty of human writing has all four.
Can a Detector Tell Me If Something Is Slop?
It can tell you something about how the text was probably produced, unreliably, and that is a different question. Detectors return false positives on careful human writing and false negatives on edited machine writing, and neither error is rare enough to build a policy on. Read the five properties instead.
Does Removing the Tells Fix It?
Partly, and there is a trap in going too far. Strip every tell at full strength and you produce a second artificial voice, clipped and telegraphic, which nobody is screening for yet because it is newer. That failure has its own article, and my own draft fell into it.
Whose Job Is This?
Somebody named, in editorial, with the authority to stop a page. Split across a team with nobody accountable and the standard is stale within two months, which is the same failure mode as every other governance document.
How Long Does It Take to Set Up?
A first version that blocks the twenty worst offenders takes an afternoon. The full standard takes as long as the decisions take, and the decisions are the slow part because they are genuine choices about voice rather than a template to fill in.
Does This Slow Publishing Down?
The check adds seconds. What slows publishing down is finding a problem after publication, or an editor manually scanning for banned phrasing that a script would catch instantly. Net, every operation I have done this to got faster.
What If We Have No Written Standard at All?
Then that is the finding, and it is the most common one. Most teams have a tone-of-voice deck and no rules a machine could evaluate. Converting the deck into assertions is the entire first week of work, and it is worth doing even if you never automate a single check.
Will Better Models Solve AI Slop on Their Own?
Partly, and in the least useful direction. Each generation gets better at avoiding the obvious vocabulary, which removes the surface tells and leaves the convergence problem completely untouched. A model that has never heard of your positioning writes an average page whether it writes it well or badly.
Does Disclosing AI Use Make the Slop Problem Worse?
It makes it visible, which is different. Over half of consumers disengage on suspicion of machine authorship, so the reflex is to say nothing. My view is that suspicion is triggered by the writing rather than by the disclosure, and a note explaining the process reads as confidence when the writing underneath it holds up. It reads as an admission when it does not.