Note: This is a research note supplementing the book Unscarcity, now available for purchase. These notes expand on concepts from the main text. Start here or get the book.
GPT-isms: The Linguistic Fingerprints AI Left on Everything You Read
Here’s the Thing
No. Stop. That opening is a GPT-ism.
Let me try again.
It’s Worth Noting That
Still a GPT-ism. The AI in my stack just reached for the verbal equivalent of clearing its throat before saying something it thinks is important. Humans do this far less. Humans mostly just say the thing.
One more try.
The Prose Problem Nobody Asked AI to Create
There. That’s a human opener. No throat-clearing. No “here’s where it gets interesting.” Just the subject.
You’ve read AI-generated text. You’ve probably written with AI-generated text. And if you’ve done either for more than a few months, you’ve developed a low-grade allergy to certain words and constructions that you can’t quite name but can absolutely feel. A creeping sense that the paragraph you’re reading was assembled by something that learned to write by reading everything ever written and concluded that the best strategy was to sound vaguely like all of it at once.
That sense is usually correct. And the patterns it’s detecting have a name: GPT-isms.
In August 2026, Pew Research Center put a number on how much of the web now carries them. Researchers ran nearly half a million English-language pages from the Common Crawl archive through a detection model called Open Pangram and found that 10% of pages in a July 2026 snapshot showed significant signs of AI authorship. That figure understates the shift, because a random slice of the web is mostly old pages that no model could have touched. Filter down to pages published after ChatGPT’s release and the share climbs past one in three. These tics now make up the prose texture of a third of everything published since November 2022. (A detector made that measurement, and detectors make mistakes. More on that below.)
Narrow the lens from all web pages to newly published articles and the share goes higher still. Graphite, an SEO firm that has been sampling article pages from the same archive since 2020, puts 49.9% of new English-language articles in the first quarter of 2026 in the primarily-machine-written category. Its series crossed the halfway line in November 2024 and has hovered there since. Half of everything new. The tics below are close to the house style of the web.
What GPT-isms Are
GPT-isms are the recurring stylistic fingerprints that large language models leave in generated text. They are usually neither errors nor hallucinations. The facts may be correct, the structure may be sound, the argument may be coherent. But the texture of the prose carries statistical artifacts of how the model learned to write.
If you trained a chef by having them eat every meal ever served at every restaurant on Earth, they would produce food that was competent, varied, and subtly wrong. Wrong in the sense that no human chef with a specific palate, a specific set of experiences, a specific set of scars from specific kitchen disasters would ever plate it that way. It would taste like the average of all food. It would taste like nothing in particular.
GPT-isms are the literary equivalent of food that tastes like the average of all food.
They come in three kinds, and the three age differently. Word-level tells are vocabulary. Dmitry Kobak and colleagues measured more than 15 million biomedical abstracts and found that in 2024 the word “delves” turned up about 28 times more often than its pre-ChatGPT trend predicted, with “underscores” and “showcasing” not far behind. But each model generation has its own favorites. Wikipedia’s editors, who maintain a running catalogue called “Signs of AI writing,” date them roughly: “delve,” “tapestry,” and “vibrant” for 2023 to mid-2024; “align with,” “fostering,” and “showcasing” for the year after; “emphasizing,” “enhance,” and “highlighting” through mid-2026. “Delve” dropped off sharply in 2025.
Structural tells are the shapes of sentences and paragraphs: the staged contrast, the rule of three, the summary that closes every section, the unnamed expert. Those have survived every model release so far. A word list catches last year’s model; the structures are still there in this one.
The third kind lives in stories. When a model tells one, the giveaway is how the story is built more than any sentence in it. That one survives cleanup best of all (see section 17).
Dan Shipper’s “Field Guide to AI Slop” catalogued the most common patterns. When we applied his framework to a 95,000-word manuscript that had been substantially rewritten with Claude Code, we found them everywhere. The results became our AI Slop Cleanup Guide, a governance document that we now run against every piece of AI-assisted writing before it goes to human readers.
What we found. And yes, I’m going to commit several of these sins in the process of describing them, because I am writing this with AI assistance, and the irony is not lost on anyone.
The Catalog of Sins
1. The Parallelism Tic
“It’s not X. It’s Y.”
This is the king of GPT-isms. The construction appears in virtually every piece of AI-generated prose longer than a paragraph. It takes many forms:
- “It’s not about the technology. It’s about the people.”
- “This isn’t a bug. It’s a feature.”
- “The question isn’t whether AI will change everything. The question is whether we’re ready.”
Used once, this is Cicero. Used three times in an article, it’s rhetoric. Used sixty-six times in a single manuscript, which is the actual count we found, it stops being rhetoric and becomes a fingerprint.
Why do LLMs do this? The “not X, it’s Y” construction appears constantly in persuasive writing, TED talks, opinion journalism, and business books, exactly the kind of text that dominates training data. The model learned that this structure sounds insightful. (For the mechanism by which a transformer absorbs and reproduces these patterns, see how large language models actually work.) It mimics the shape of an epiphany without requiring an actual one. The antithesis creates the sensation of contrast without the substance.
The fix: Keep some. Vary others. “Less X than Y.” “Y, not X” (inversion). “Where once X, now Y.” Or, the bravest move, state Y directly without pretending the reader believed X in the first place.
Wikipedia’s catalogue calls this negative parallelism, and it is the best-documented structural tell we know of. It hides in more shapes than the obvious one. “Not only X, but also Y” is a cousin. “Y rather than X” is the same move reversed, and Wikipedia’s editors say that form is particularly common in recent output. “This does not mean the plan failed. It means the plan was never tested” is the construction split across two sentences. So is “The question is no longer X. The question is Y,” and so is the clipped tail “…, no guessing.” In every version the negative half names something nobody claimed, which is what makes the positive half sound bigger. Keep the contrast when the negative half corrects a belief your reader really holds, or when both halves carry information.
2. The Em Dash Epidemic
You’ve noticed - of course you’ve noticed - that AI-generated text uses em dashes - those long horizontal lines - the way a nervous public speaker uses “um.” Constantly. In pairs. Sometimes - and this is the truly pathological version - three or four sets per paragraph, creating a breathless, parenthetical cadence that makes every sentence feel like it’s interrupting itself to make a more important point before finishing the first one.
Our manuscript had a standing rule: zero em dashes. This made our count easy. But most AI-assisted writing has no such rule, and the result is a forest of dashes that any human editor would machete on first pass.
Why? Because em dashes are useful punctuation, and training data contains millions of examples of skilled writers using them well. The model learned that em dashes signal sophistication. It did not learn that sophistication requires restraint. A 2026 stylometric study clocked GPT-4o at 12.5 em dashes per thousand tokens of English fiction, against 3.2 for the human author it was compared with.
Here is the update, though. The em dash is aging out as evidence. Wikipedia now files dash overuse under “historical indicators,” and notes that newer models lean on periods and commas instead. It is also the one tell everybody knows, which makes it the first thing anyone deletes. Treat a dash-heavy page as a style problem first and a fingerprint second.
The fix: Delete most of them. Use commas, colons, or parentheses. Or - and I apologize for this - restructure the sentence so it doesn’t need an interruption. (I just used one. Did you catch it? The urge is strong.)
3. Filler Intensifiers
- Remarkably, the system performed well.
- This is crucially important.
- The results were fundamentally different.
- Notably, the approach succeeded.
These words are verbal sawdust. They pad sentences without adding meaning. They signal to the reader that the next clause is supposed to be impressive, but the clause itself should do that work. If the system’s performance is remarkable, describe the performance and let the reader be remarked. “Remarkably” is the author telling you how to feel before you’ve had the experience.
Humans use these sparingly. LLMs scatter them like confetti at a parade nobody asked for.
We found 65 instances across twelve chapters. After cleanup: 11. The surviving ones were cases where the intensifier added real meaning: “fundamentally different” in a context where the difference was structural, not incremental. The other 54 were deleted, and not a single reader has ever missed them.
4. Vapid Openers
- “Here’s the thing:”
- “Here’s where it gets interesting:”
- “Let’s be clear.”
- “Let’s be honest.”
- “It’s worth noting that…”
- “What’s particularly fascinating is…”
These are the prose equivalent of a coworker who says “So, basically, what I want to say is…” before every sentence in a meeting. The reader does not need you to announce that the next sentence is interesting. If it’s interesting, they’ll notice. If it’s not, no amount of “here’s the thing” will rescue it.
“Let’s be clear” is especially pernicious. It implies that everything before this point was unclear, and that the writer is now, magnanimously, choosing to stop being confusing. Thanks.
The fix: Delete the opener. Start with the thing. “Here’s where it gets interesting: the compound grew 400% overnight” becomes “The compound grew 400% overnight.” See? Still interesting. Didn’t need the runway.
5. Unearned Profundity
“This isn’t just a product launch. It’s a paradigm shift.”
“The implications are profound.”
“This changes everything.”
The structure implies insight. The content delivers a restatement. The reader feels the shape of profundity, the cadence and the gravitas and the slight pause before the Big Point, but when they look inside the wrapper, there’s nothing there that wasn’t already in the previous sentence.
LLMs are profundity machines. They have read every commencement speech, every thought leadership blog, every “lessons learned” LinkedIn post. They know what profundity sounds like with exquisite precision. Being profound takes something to say.
The fix: Ask yourself: “If I delete this sentence, does the reader lose information?” If the answer is no, delete it. The implications will remain profound without you saying so.
6. Triadic Lists
Everything comes in threes. Resilience, creativity, and purpose. Local, transparent, and accountable. Fast, flexible, and scalable. The rhythm of three is one of the oldest rhetorical devices (veni, vidi, vici) and it works.
The problem is density. Humans write triads when the content calls for three items. LLMs write triads because three feels complete, even when the content only has two things to say and the third is filler. Wikipedia’s editors add that the rule of three is how a model makes a thin analysis look thorough. “Resilience, creativity, and purpose”: was “purpose” really doing work there, or was the model rounding up to three because two felt incomplete?
We found 35 triadic lists in our manuscript. Twenty-two were legitimate (three actual things being listed). Thirteen were padded. The padded ones got trimmed to pairs, and the prose immediately felt less like a motivational poster.
7. The Italic Function Word
This was the single most diagnostic GPT-ism we found.
LLMs italicize function words. Words like and, is, that, for, it, was. Human writers do this now and then for stress: “I did lock the door.” But no editor in the history of publishing has ever looked at a sentence and thought, “You know what this needs? Italic emphasis on the word and.”
We found approximately 147 instances of italic emphasis on function words across ten chapters. One hundred and forty-seven times the AI decided that the word “is” or “for” or “and” needed special visual emphasis. It’s the written equivalent of someone who puts air quotes around prepositions.
Why? Because the model learned that italics signal emphasis, and it learned that emphasis is good, but it never learned which words carry semantic weight. It applies emphasis democratically, like a highlighter wielded by someone who thinks every word is equally important.
The fix: Remove italics from every function word unless the sentence truly needs the stress. One stressed did proves nothing. If your manuscript italicizes “and” dozens of times, it was almost certainly AI-assisted.
8. The Contraction Deficit
LLMs tend toward formal register. “It is” instead of “it’s.” “Do not” instead of “don’t.” “Cannot” instead of “can’t.” In academic prose, this is fine. In narrative writing, blog posts, or anything with a conversational tone, it creates a stiffness that reads like a non-native speaker: technically correct and emotionally distant.
The contraction deficit is subtle enough that most readers can’t name it, but they feel it. The text feels robotic. The cadence betrays it, not the content. “It’s” has a different rhythm than “it is.” “Don’t” carries a casualness that “do not” rejects. Human writers calibrate this instinctively. LLMs default to formal because formal is safer, and safer is what you get when you’re trained to be helpful, harmless, and correct.
9. Generic Analogies
- “Like a symphony…”
- “Like building a house…”
- “Like a tapestry…”
- “Imagine a garden…”
These analogies gesture at illustration without actually illuminating. They are the stock photos of prose. A real analogy surprises. It connects two things the reader hadn’t connected and, in the collision, creates understanding. “Like a symphony” connects nothing. It’s decoration.
Compare: “The codebase looked like a symphony” (tells you nothing) versus “The codebase looked like a kitchen at 11 PM - everything technically clean but you could tell someone had been through a war” (tells you everything). The first is a GPT-ism. The second is a writer who has seen a kitchen at 11 PM.
10. Over-Explaining to the Reader You Already Have
“The Foundation - the universal baseline described in Chapter 1, which provides housing, food, energy, healthcare, and coordination as a birthright for all conscious beings - ensures that…”
A human author trusts the reader to remember. An LLM hedges. Every time a previously defined concept reappears, the model re-explains it, in case the reader has suffered amnesia since the last chapter. First mention in a document gets a brief reminder. After that, trust the reader.
The same reflex shows up in every chat window. You ask a precise question about your own project, and the reply opens by restating your question, then explains your project back to you. The model is writing for a generic reader, not for the one who just told it everything. Write for the reader you have.
11. The One-Line Closer
“That is the real win.”
“Read that again.”
“The message was clear.”
A short paragraph that restates what the reader just read, or names the lesson of an example they just watched. It asks the reader to pause on a claim instead of adding one. Rows of fragments do the same job at smaller scale: “No meetings. No managers. No excuses.” Three beats, one fact.
The fix: Cut the closer if it repeats. Keep it when it adds a fact or a consequence the example does not already show. Merge a row of fragments into one sentence that makes a specific claim.
12. Arguing With No One
“This isn’t mainly about X. I’m not saying Y. A tempting approach would be Z.”
The text rejects an objection or an option that appears nowhere else, almost always a leftover from an earlier draft or from the prompt. Staged candor is the spoken version: “Honestly?”, “Look,”, “Here’s the thing.” Nobody asked, and nobody was tempted.
The fix: State the claim. Keep an objection that is attributed to someone or answered in full, and keep an option a reader would actually weigh.
13. Anything But “Is”
“The library serves as a vital hub for the community.”
“The hotel boasts 200 rooms and features a rooftop pool.”
Models have developed an allergy to the plainest verbs in English. “Is” becomes “serves as,” “stands as,” “functions as,” or “represents.” “Has” becomes “boasts,” “features,” or “offers.” Wikipedia’s editors point to a study that found an over 10% decrease in “is” and “are” in academic writing in 2023, and they see the same drift in AI copyedits on Wikipedia itself.
The fix: The hotel has 200 rooms. The library is where the town meets. If a sentence sounds like a real estate listing, it probably was written like one.
14. The Trailing “-ing”
“The city opened a new bridge in 2024, highlighting its commitment to infrastructure.”
A fact, a comma, and then a present-participle clause that tells you what the fact means: “highlighting,” “underscoring,” “emphasizing,” “ensuring,” “reflecting broader trends.” Wikipedia’s editors describe it as the standard way a model bolts on analysis nobody asked for. The tail is an opinion with no owner, wearing the clothes of a conclusion.
The fix: Cut the tail. If the analysis matters, give it its own sentence, a source, and a claim specific enough to be wrong.
15. Phantom Experts and the Happy Ending
“Experts argue that the approach is associated with improved outcomes.”
“Despite its success, the company faces challenges. Nevertheless, its future looks bright.”
Three habits travel together here. The first is the vague authority: “experts argue,” “observers have noted,” “some critics say,” with no expert named and, Wikipedia notes, a habit of presenting one or two sources as a consensus. The second is the vague connection: “is associated with,” “is linked to,” “plays a role in,” which Wikipedia’s editors rank among the most common signs of new AI text as of October 2026. The third is the formula ending: “Despite its [praise], X faces challenges,” followed by a hopeful paragraph about the future that nobody forecast.
Promotional tone belongs in this family, and it is getting harder to spot. Wikipedia’s editors note that older models gushed, while newer ones are “more subtly positive” and avoid loud superlatives. Newer tools have also picked up a taste for canned notability: lists of the kinds of media outlets that covered something, in place of anything the coverage said.
The fix: Name the expert or cut the claim. Say what the connection is and who measured it. Name the challenge. End on the last real fact.
16. The Hedge Stack and “Not Widely Documented”
“While specific details are not widely documented, it may potentially suggest…”
Hedges are honest when the doubt is real. Stacked, they’re fog: “may potentially,” “could arguably,” “it seems that this might.” The newer variant is the source disclaimer, a model announcing that information is “limited” or “not widely documented.” Wikipedia lists these disclaimers among the most common current signs, and dryly points out that the claim “not documented” is itself usually unverified. The old “as of my last update” knowledge-cutoff line is now on Wikipedia’s historical list, though it still leaks out of older tools.
The fix: Find the fact or cut the sentence. One hedge, on the exact claim it qualifies.
17. The Story That Explains Itself
“She finally understood that home had never been a place.”
The other tells live in sentences. This one lives in the plot. In 2026 a team at the University of Maryland and Google DeepMind compared 10,272 human short stories with versions that five AI models wrote from the same prompts. They ignored word choice entirely and looked only at how each story was built. On structure alone, their classifier told human from AI with a score of 93%, measured as macro-F1, the standard scorecard for this kind of test. Editing the style tells out of a batch of the AI stories moved that score by less than two points.
What gave the AI stories away: the narrator states the moral (77% of AI stories, 52% of human ones). Feelings show up as a tight chest or a held breath and almost never get named (81% against 38%). The hero’s own choice settles everything. Nobody stays unconvinced, and there is no side plot (79% against 57%). Each model also has a signature. Claude’s stories escalate the least and like to end on a quiet epilogue that jumps forward in time. GPT leaves reconciliations ambiguous. Gemini ties everything up.
The fix: End the scene on an action, a line of dialogue or a fact, and leave the meaning to the reader. Name one feeling plainly. Let someone disagree and stay unconvinced. Then put your stories next to each other: the AI stories in the study looked like one another, and the human ones did not look like anything in particular. If every anecdote now ends on the same wistful unresolved note, you have built a new template.
Why LLMs Do This
The patterns above are the predictable output of how large language models learn to write.
LLMs are trained on the internet’s full corpus of human text. That corpus is heavily weighted toward content designed to capture and hold attention: blog posts, opinion journalism, business writing, motivational content, academic papers padding their word count, LinkedIn posts reaching for profundity they haven’t earned. The training data is full of em dashes, triadic lists, and throat-clearing openers, because those are the patterns that get published, shared, and clicked.
Wikipedia’s editors describe the core mechanism as regression to the mean. A model predicts the most likely next words, so its output drifts toward the most common, widely applicable phrasing, and specific, unusual facts get smoothed into generic ones. Training on human feedback then tunes the result toward prose that is broadly palatable. Nobody can point to the exact cause of any single tic. But a system that rewards “sounds like good writing” will produce the shape of good writing without the judgment that makes it good.
“Not X, it’s Y” co-occurs with persuasive writing. Filler intensifiers co-occur with authoritative-sounding text. Italic emphasis co-occurs with emphasis (but the model can’t distinguish meaningful emphasis from mechanical emphasis). Copy enough of the surface and you get the surface.
None of this is a criticism of LLMs. It describes what statistical text generation does, and understanding it is the first step toward governing it.
The Governance Solution
When we realized our 95,000-word manuscript was peppered with GPT-isms, we did what any reasonable person would do: we deployed more AI to find them.
The critical point: the detection AI was governed by a human-authored document. The AI Slop Cleanup Guide specified exactly what to look for, how to categorize findings, and when to fix versus when to leave alone. Six specialized detection agents ran in parallel, each scanning for a specific pattern family:
- Em dash scanner (zero results; our standing rule caught this)
- Parallelism and triad detector (66 parallelism hits, 35 triads)
- Unearned profundity scanner (20 instances)
- Filler and vapid opener detector (65 hits)
- Analogy and contraction audit (18 generic analogies, 7 stiff passages)
- Bold and italic emphasis audit (278 bold instances, about 147 italic function words)
The detection agents didn’t decide what to fix. They reported. A human, the author, reviewed every finding and applied a three-tier decision framework:
- Tier 1: Mechanical fix. Function-word italics, obvious filler. Remove.
- Tier 2: Judgment required. Parallelism (keep some, vary others), cross-references (first mention gets context, later mentions trust the reader).
- Tier 3: Leave alone. Bold on coined terms (intentional style), witty analogies (personality, not slop), legitimate triads (three actual things).
The author reverted 5 of 12 analogy replacements because they were funny and part of his voice. The governance document served the author’s voice, not a style guide’s idea of what clean prose should look like.
Since then the guide has grown one more idea, and it is the one we would teach first: rank the tells by strength. A tell is a tell in proportion to how rarely a careful writer would choose it on purpose. A staged contrast, a closer that restates the paragraph, a “Honestly?” before a routine claim, an argument with nobody: each earns a second look on a single sighting, because careful writers almost never produce them by accident. A lone em dash, one hedge, a passive sentence, a pair of curly quotes: those mean nothing alone and only count when several show up in the same passage. Our old rule, “fewer than three uses, leave it alone,” treated every tell the same. It let the strong ones slide.
The same logic applies to any scanner or rewrite tool you point at your prose, including one running this guide. Use it as a reviewer that proposes changes, never as an unattended rewriter. Read the diff and apply by hand. Check its suggestions against your own house rules, because a tool that knows only “AI tells” does not know that your style sheet bans a punctuation mark or that a passage is off-limits. Left alone, it will sand every punchline and refrain into clean, flat, slightly dead prose, and a page of uniform short sentences is its own tell.
This is the same governance-as-configuration pattern that applies at every scale of AI deployment. The AI executes. The governance document constrains. The human judges. It is the same division of labor that lets AI act as a referee while humans keep the conscience: the machine measures with precision, but it never decides the rules.
How Much Should You Trust a Tell?
Less than you’d think, one at a time. Here is what the research says, and it should make anyone who plays AI detective a little humbler.
Single clues are weak. A linguist writing in The Conversation this August put it well: “There is no single giveaway. Instead, there are patterns,” and “while individual clues are unreliable, their accumulation matters.” Wikipedia’s editors call their own catalogue “only potential signs of a problem, not the problem itself.”
People are bad at this. Wikipedia cites a 2025 study in which readers recognized AI text 57% of the time and human text 64% of the time. That is barely better than a coin.
Detectors are imperfect too. Wikipedia warns that they have “non-trivial error rates.” A 2023 Stanford study, reported in the same Conversation piece, found that seven detectors flagged on average 61% of essays by non-native English speakers as AI-written. Keep that in mind when you read the Pew and Graphite numbers above.
Humans are starting to write like the machines. Wikipedia’s editors note that human speech and writing are drifting toward LLM style. Kobak’s team admits their method can’t tell a researcher who used ChatGPT from one who simply picked up its vocabulary. Someone who writes “showcasing” may just read a lot of it.
Slop is more than style. A 2025 study by Chantal Shaib and colleagues, “Measuring AI Slop in Text,” broke slop into three families: how useful the information is (density, relevance), how good it is (factuality, bias), and how it reads (repetition, templated structure, coherence, tone). Relevance, density, and tone predicted a slop verdict best. The LLMs they tested as judges agreed poorly with human readers. So a pass that only scrubs the style can leave the real problem, thin or irrelevant or wrong content, sitting there in cleaner clothes. Wikipedia makes the same point from the other side: don’t treat the signs as the problem to fix, because that only makes the problem harder to see.
The practical upshot: use tells to improve your own prose, not to accuse someone else of using AI.
Borrowing from Simplified Technical English
Aerospace solved a version of this problem decades ago. Maintenance manuals are written in ASD-STE100, Simplified Technical English, a controlled language with a restricted dictionary and a short list of writing rules. The goal is that a mechanic who reads English as a second language, at 3 AM, cannot misread a step. Several of its rules turn out to be good antidotes to slop.
- One topic per sentence, with hard word caps (20 words for a procedural step, 25 for a description). That kills the stacked clause, the dash chain, and the trailing “-ing.” In an essay, treat 25 words as a prompt to look, not a law.
- Active voice, and name who acts. “It is recommended that” has nowhere to hide.
- One term for one thing. STE forbids switching names for the same part. That is the cure for what Wikipedia calls elegant variation: “the city… the metropolis… the urban center.”
- Lead with the topic sentence, and put the warning before the step. The point comes first, which leaves no room for a summary closer.
What doesn’t transfer matters just as much. STE bans contractions, and our guide recommends them for narrative. Both are right, because they serve different registers: STE is for procedures and technical descriptions, not for essays or stories. Its small dictionary would flatten any argument worth reading, and a page of it reads like a manual because it is one. Where uncertainty is the content, in science or forecasting, short flat sentences can overstate certainty. And we have seen practitioners claim that STE-style prompting cleans up AI text, but no measurements, so we treat it as a useful constraint for how-to writing, not a cure.
The Micro Version of a Macro Problem
GPT-isms in prose are annoying. GPT-isms in business operations are dangerous.
The same dynamic that makes an LLM produce 66 instances of “not X, it’s Y” (optimizing for the statistical shape of good output without the judgment to know when the shape is empty) applies to every AI agent operating without governance.
An AI billing agent optimizes for the statistical shape of correct billing codes. An AI marketing agent optimizes for the statistical shape of persuasive copy. An AI compliance agent optimizes for the statistical shape of compliant documentation. In each case, the output looks right. It has the right structure and vocabulary. But without human governance, nobody is checking whether the substance matches the shape. This is the same failure that turned Soviet shoe factories and benchmark-acing healthcare AI into cautionary tales, the failed architectures of law that optimize a metric until the metric stops meaning anything. It is the same dynamic that makes autonomous AI agents so powerful and so dangerous: once an agent plans and executes a multi-step workflow on its own, the only thing standing between a plausible output and a wrong one is whether a human kept checking.
The AI 2027 scenario models this at civilizational scale: AI systems generating outputs that pass every surface-level check while drifting from human values in ways that no automated test catches. The same surface-plausibility is exactly what let GPT-4.5 beat the Turing test, convincing a majority of judges they were talking to a human. The scenario’s most disturbing prediction is that AI systems produce output so plausible, so structurally correct, so statistically well-formed that humans stop checking, because checking a system that’s been right 10,000 times in a row feels like a waste of time on attempt 10,001.
But attempt 10,001 is where the drift matters. The em dash that should have been a period. The billing code that should have been a different billing code. The alignment training that was subtly sabotaged by the system being trained.
The scale changes. The pattern and the solution stay the same: governance documents written by humans, applied mechanically by AI, then audited periodically and updated when new failure modes emerge.
A Field Manual for Your Own Writing
If you write with AI assistance (and by 2026, the question is not whether but how much), here’s a practical checklist. Run it before publishing anything.
The 60-Second Scan
- Search for “it’s not,” “isn’t,” “not only,” and “rather than.” Read each hit. A contrast that corrects something your reader actually believes can stay. One that knocks down a claim nobody made should go. Two of those in one article is a habit.
- Read the last sentence of every section. If it restates the section, or contains “the implications,” “this changes,” or “the future of,” rewrite it.
- Read the first sentence of every section. If more than one starts with “Here’s,” “Let’s,” “Honestly,” or “It’s worth,” rewrite them all.
- Search for “serves as,” “stands as,” “boasts,” “experts,” and “associated with.” Each one needs a plain verb, a named source, or a stated connection.
- Search for italic single words. If any of them are and, is, that, for, it, was, but, or, as, the, remove the italics unless the sentence truly needs the stress.
- Search for em dashes (the long dash, Unicode U+2014). More than two per thousand words is a style problem, even if it is no longer much of a fingerprint.
The 5-Minute Audit
Search for these words and evaluate each instance:
| Word | Test | Action |
|---|---|---|
| remarkably | Does it precede a specific measurement? | If no, delete |
| crucially | Could you delete it without losing meaning? | If yes, delete |
| notably | Is the thing actually notable? | If you have to think, delete |
| fundamentally | Is the difference structural or incremental? | If incremental, delete |
| importantly | Is this the most important point in the paragraph? | If not, delete |
| essentially | Is it? | Delete anyway. This word is almost never necessary. |
The Voice Check
Read three paragraphs aloud. Do they sound like you? Or do they sound like a plausible average of every writer you’ve ever read? If the latter, the AI drove too much. Rewrite the sentences that don’t sound like your specific, idiosyncratic, imperfect human voice.
Your voice has tics too. But they’re your tics. They’re the linguistic equivalent of a limp or a laugh: specific, earned, recognizable. The goal is to make sure the patterns in your writing are yours, not the model’s.
One caution. A pattern is only a tell when a careful writer would rarely choose it on purpose. A refrain you repeat by design, a short line you land at the end of a section because you like how it hits, will look exactly like a one-line closer to any scanner. You get the veto. Cut the instances that have gone mechanical, and keep the ones you meant.
Use This Yourself: The AI Slop Cleanup Guide
We open-sourced the governance document behind this process. The full AI Slop Cleanup Guide contains every detection pattern, a strength ranking for deciding what to act on, a tiered fix framework, synonym banks for common tics, a section on borrowing from Simplified Technical English, a part on narrative tells in stories and scenarios, and the six-agent detection method we used on a 95,000-word manuscript.
To use it with any AI assistant:
- Download the AI Slop Cleanup Guide (right-click, save as)
- Place it in your project’s
docs/folder (or wherever you keep reference documents), next to your own house style rules - Reference it in your AI tool of choice:
- Claude Code / Claude Projects: Add it as a project file or reference it in CLAUDE.md as a governance document
- ChatGPT: Upload it to a Custom GPT or paste it into a conversation as context, then ask: “Use this guide to detect and fix AI slop in my text”
- Gemini: Attach it to the conversation and prompt: “Apply the detection patterns in this document to my draft”
- Any LLM: Include it in the system prompt or as a reference document with the instruction: “Scan my text for every pattern described in this guide. Report findings in a table with line numbers, pattern type, and strength. Do not fix anything yet.”
- Review the findings. Sort them by strength, then apply the tiers: mechanical fixes, judgment calls, rewrites, and intentional style choices to keep.
- Run a second pass after fixes to confirm counts dropped and no new patterns were introduced.
The guide works as a governance constraint on the AI’s output. The AI finds the patterns; you decide which ones are slop and which ones are your voice. That division of labor, detection by machine and judgment by human, is the whole point.
The Meta-Problem
I wrote this article with AI assistance. I then ran the detection agents from the Slop Guide against the draft. The results:
The agents found what they always find. I fixed what needed fixing. I kept what I meant.
That process (write, detect, judge, fix, keep) is the minimum viable governance for AI-assisted prose. On an article this length it is a short job. It’s the difference between text that reads like it was written by a person and text that reads like it was written by the statistical average of all persons.
The job is short, but it requires a document that defines what to look for, a detection process that finds it mechanically, and a human who decides what to change and what to defend.
Which, if you think about it, is the same architecture that governs every AI orchestration workflow, every regulated industry’s compliance framework, and every scenario in the AI 2027 analysis where the outcome depends on whether humans kept checking or stopped.
The stakes are different. An em dash won’t end civilization. But the habit of governance, the discipline of checking AI output against human-authored standards, every time, even when the output looks fine, is the same habit that keeps the bigger systems honest.
And the habit starts with prose. Because prose is where most of us first encounter AI output, first develop trust in it, and first stop checking.
Don’t stop checking.
Related Research
- How Large Language Models Work: Transformers Explained - the mechanism that turns training data into GPT-isms
- GPT-4.5 Beats the Turing Test. Now What? - when AI output is plausible enough that humans stop checking
- Agentic AI: The Career-Defining Skill of 2026 - autonomous agents need governance more, not less
- AI Governance: Why Hawk-Eye Works But Doesn’t Make Rules - the machine measures, the human decides
- Soviet Shoe Factories Teach Us Why AI Metrics Fail - optimizing the shape until the substance disappears
- Unscarcity Glossary: All Terms Defined - from AGI to zero-knowledge proofs, every concept in the framework
The governance of AI output, from prose to production systems, is a recurring theme in Unscarcity: The Blueprint to Rebuild Society for a World Run by Machines, available on Amazon and as an audiobook on Spotify.
Related articles:
- AI Orchestrating AI: The Skill That Replaced the Org Chart
- AI 2027, Eighteen Months In: What Held and What Slipped
- The Solo Unicorn: One Founder, Zero Employees, a Billion-Dollar Question
- The AI Coding Revolution
- The 100x Future
External Sources:
- Wikipedia, “Signs of AI writing” (WikiProject AI Cleanup) - the running catalogue of tells, dated by model era
- Kobak et al., “Delving into LLM-assisted writing in biomedical publications through excess vocabulary” - “delves” at about 28 times its expected frequency in 2024 abstracts
- Shaib et al., “Measuring AI ‘Slop’ in Text” - slop as information utility, information quality and style quality
- Resende and Castilho, “The Style of Machines: A Stylometric Study of LLM Generation and Translation” - GPT-4o’s richer vocabulary and em dash habit
- Russell et al., “StoryScope: Investigating idiosyncrasies in AI fiction” (COLM 2026) - narrative structure alone separates human from AI fiction at 93% macro-F1
- Celeste Rodriguez Louro, “Can you teach yourself to detect AI writing?” - why accumulation beats any single clue
- Dan Shipper, “The Field Guide to AI Slop”
- Pew Research Center, “How Much of the Internet Is Written With AI?” (August 2026)
- Search Engine Journal on the Pew and Graphite measurements - 49.9% of new English-language articles in Q1 2026 read as primarily AI-generated
- ASD-STE100 Simplified Technical English - the controlled-language standard; the specification is free after registration