In 2024, an Elsevier journal published a peer-reviewed paper on lithium-battery separators. The introduction opened with a single sentence: "Certainly, here is a possible introduction for your topic."[1]
Nobody deleted it. Not the three authors, not the peer reviewers, not the editor who signed off. The chemistry underneath may have been perfectly sound. What ended the paper was one line no human wrote on purpose — the seam, the exact spot where a person stopped typing and ChatGPT took over, sitting in plain sight under three names.
It wasn't a fluke. Researchers have since catalogued a pile of published papers still carrying the tell their authors forgot to cut. The line "As an AI language model, I don't have access to real-time information" sits stranded mid-paragraph in the permanent record of science.
Here's what those papers share with your LinkedIn feed. Readers can't always tell you why something smells like a machine. But they catch the spot where you stopped writing and the robot took over, every time.
The numbers behind that instinct are lopsided. Ahrefs scanned 900,000 pages in April 2025: 74.2% now contain AI-generated content, but only 2.5% are pure AI.[2] The gap is the whole story. Almost three-quarters of new pages are a human-AI blend — writers and robots sharing a document and stitching the result together with their eyes closed.
The seam is everywhere
900,000 web pages, April 2025 · who actually wrote them
- A human–AI blend71.7%
- No AI — pure human25.8%
- Pure AI, no human2.5%
And yet the machines are losing where it counts. AI passed humans in raw publishing volume back in November 2024; Graphite clocks more than half of everything new as machine-made. But when Graphite checked what actually ranks, 86% of the articles on Google's first page were written by humans, and 82% of the sources cited by ChatGPT and Perplexity were human too.[3] The machines won the volume war and lost the one that pays.
A minority of what's written. A majority of what wins.
Share written by humans, by stage of the content lifecycle
So everyone's worried about the wrong thing. The seam isn't whether you used AI. It's where you let it show.
I use AI. So does almost everyone making content now. I've also spent two articles pulling apart how detection works and why the whole apparatus is broken, so this isn't a sermon about authenticity you can buy in a bottle. It's a field manual for moving the seam somewhere it can't hurt you. Which starts with admitting that most of the popular advice is already dead.
The Tells That Peel Off First
You know the greatest hits. "Delve." "Tapestry." The em dash. These tells are real. They're also rotting by the month, which makes them a terrible thing to build a strategy on.
"Delve" is the cautionary tale. Its use in formal writing exploded after ChatGPT. One analysis of millions of papers found that AI-preferred words like it now fingerprint at least one in ten 2024 medical abstracts, a bigger vocabulary shift than the COVID pandemic caused.[4] Then the joke went viral. And then it got complicated.
In April 2024, Paul Graham, the Y Combinator co-founder, received a cold-email pitch, spotted the word "delve," and pronounced it ChatGPT. "No one uses it in spoken English," he wrote.[5]
A continent disagreed. Nigerian and other African writers pointed out that "delve" is ordinary formal English where they're from, the Commonwealth English drilled in by the same colonial school system that handed them "expatiate" and "burgeoning." The Guardian's Alex Hern had floated the likely reason months earlier: the low-paid workers who train these models by grading their output sit disproportionately in Nairobi and Lagos, and a model that learns from Nigerian raters learns to write, in Hern's words, "slightly like an African."[6]
So the most famous machine-tell might just be formal human English, mislabeled. And the cost lands on real people. One Nigerian journalist watched her brother get accused of using ChatGPT on his postgrad assignments for writing the way he was taught. Students now water down their own prose on purpose to slip past detectors.
Before you take that story as gospel, though, the tidy version doesn't fully hold either. When the linguists Juzek and Ward actually checked the corpora, "delve" turned out not to be uniquely Nigerian, or uniquely anything.[7] A VC bet his reputation that no human says "delve," got schooled by a continent, and the linguists still can't fully explain where the model picked up the habit.
The em dash tells the same story faster. After a year of people wielding it like a murder weapon, OpenAI announced in late 2025 that it had finally taught ChatGPT to drop the dash. On request. Read the fine print. It obeys only when you ask, and still reaches for the dash by default. Sam Altman called it a "small-but-happy win."[8] The signature you trained yourself to spot is one toggle from vanishing.
That's every surface tell. The instant enough people learn to catch one, the model sands it off, or a famous investor turns it radioactive and takes a few thousand innocent writers down with him. If your entire plan for not sounding like AI is find-and-replace on a banned-word list, you're optimizing against last year's model.
Strip every "delve" and the thing still reads like a machine. Because under the words is a second layer, and it doesn't sand off nearly as fast. The shapes.
The Tells Under the Words
Nobody has mapped those shapes more obsessively than a volunteer army on Wikipedia. WikiProject AI Cleanup, formed in late 2023, has spent two years hunting AI text by hand. They've flagged well over 500 articles and published a 24-point field guide to spotting the machine.[9] They don't run perplexity scores. They read.
And what they found isn't a vocabulary list. It's a set of structural habits: shapes the model reaches for no matter which words you ban.
Their description of the core one is the best sentence anyone has written about AI prose. The machine, they say, keeps "shouting louder and louder that a portrait shows a uniquely important person, while the portrait itself is fading from a sharp photograph into a blurry, generic sketch." A specific fact, "the inventor of the first train-coupling device," comes out the other side as "a revolutionary titan of industry." Less specific, more grand. Every time.
Here are the shapes worth knowing.
Contrastive negation. "Not just X, but Y." "It's not about the beat; it's about the atmosphere." The rhetorical move that used to mark a sharp essayist now marks a language model, which deploys it constantly, unprompted, to manufacture tension that was never in dispute.
Full confession: the thesis at the top of this piece, "it isn't whether you used AI, it's where you let it show," is exactly that construction. It's a good tool. That's precisely why the model stole it. I'll ration them from here.
The trailing "-ing" tail. A fact, followed by a gerund that pretends to analyze it. "...underscoring its importance." "...highlighting broader trends." The Wikipedia editors find this one everywhere, stapled onto sentences that were already finished.
The rule of three. The model loves a triplet. One caught example claimed a technology sparked debate about "authenticity, consent, and the psychological effects of digitally extending personhood." Three, always three, whether the subject has three real edges or not.
Puffery. The mundane inflated into the momentous. In one entry, a small fish native to Lake Malawi picked up a fabricated coda about "sustaining the cultural traditions connected to Hawaii's native flora." The fish is not from Hawaii. The AI needed a grand ending, so it grew one.
The confidence behind these shapes is the frightening part. In January 2023, someone published a 2,000-word Wikipedia article on Amberlihisar, a 15th-century Ottoman fortress, complete with construction dates, a named architect, and a full bibliography. The fortress never existed. Not one citation was real. It passed human review and sat there, unquestioned, until December.[10]
So these shapes run deeper than "delve." They survive a find-and-replace. But watch how fast they die anyway.
In January 2026, a tech founder named Siqi Chen did the obvious thing. He pointed Claude at Wikipedia's 24-point list and told it to build a skill that avoids every item. "It's really handy that wikipedia went and collated a detailed list of 'signs of ai writing,'" he wrote. "So much so that you can just tell your LLM to… not do that." The skill hit 1,600 stars on GitHub in days.[11]
The volunteers who spent two years writing the detection manual had accidentally written the evasion manual. That's the entire lifecycle of a tell, start to finish, in one story. Which leaves the only question that matters: if every nameable tell peels off the moment it's named, why does the machine keep reaching for these shapes in the first place?
Why the Machine Talks Like This
To answer that, start with the machines built to catch it. Detectors measure two things. Perplexity: how predictable your next word is. Burstiness: how much your sentence length jumps around. Humans are unpredictable and lumpy. AI is smooth and even. Score low on both, get flagged.
Here's the catch that should have ended the whole detection industry: good formal writing scores low on both too. Polish reads as guilt.
Stanford researchers led by James Zou proved it in 2023. They ran essays by non-native English speakers, TOEFL exam essays written by Chinese students years before ChatGPT existed, through seven detectors. The tools flagged 61% of them as AI.[12] Essays by U.S.-born eighth graders? Flagged about 5% of the time. Same detectors, same task. The only real difference was that the non-native writers leaned on careful, formal, slightly constrained English, and constraint looks exactly like a machine to a perplexity score.
Then Zou's team ran the cruel, clarifying version. They took those flagged human essays and asked ChatGPT to punch up the vocabulary: "enhance the word choices to sound more like a native speaker." The false-positive rate collapsed from 61% to 11%. Fancier words made real human writing read as human. And when they added flourish to genuine AI text, the detectors waved that through as human too.
Fancier words read as more human
Detector false-positive rate on genuine human essays
Sit with that. The machines reward ornament and punish plainness, which is the exact inverse of good writing. You cannot win this by writing more correctly. Correct is what the model is for.
The reason runs deeper than "it predicts the average," and a paper out of Tsinghua this year finally named it. Training a model on human feedback is mostly teaching it what not to do. The rules for wrongness are clean and finite: don't lie, don't be unsafe, don't contradict yourself. The rules for genuine excellence are infinite and impossible to write down. So the model gets very good at dodging every clear mistake, and, in the paper's words, "the feasible response space narrows monotonically."[13] The safe zone shrinks and shrinks until whatever's left is acceptable to everyone and interesting to no one.
None of that is the model malfunctioning. It's the model doing exactly what safety training rewards, with voice as the collateral. Voice is the willingness to say the one thing the average person wouldn't, and the average is the only place a probability engine can live.
The first time that clicked for me, the flatness stopped reading as a defect to scrub out and started reading as gravity: the pull every real sentence has to push against.
Which tells you what's actually missing from AI writing. Not a bigger vocabulary — the opposite. The detail nobody could guess. The opinion with a cost attached. The number pulled from your own logs. The tangent that shouldn't work but does. AI hands you the rich tapestry. You bring the splinter in your thumb. Everything below is mechanics.
Set the Machine Up Right
Most people run AI backwards. They let it write, then fight the output with a wall of instructions: don't use "delve," don't use passive voice, don't be verbose. Don't, don't, don't.
That wall is making your output worse. There's a real reason, and a 2025 study confirmed it holds inside language models, not just human heads: telling a system not to think about something forces it to think about it. "Don't mention the white bear" lights up the white bear. The researchers watched it happen in the model's own wiring: attention heads amplifying the forbidden word even as it successfully avoided typing it.[14] Negation doesn't remove a concept. It activates the concept and hopes.
So turn every "don't" into a "do." Not "avoid corporate jargon" but "write in plain nouns and active verbs, like you're explaining it to a smart friend." Positive instructions give the model somewhere to go. Prohibitions just wave a flag at the thing you didn't want.
Then feed it you. Not a one-line "write in my voice," which does nothing. The tools that work build a profile from a stack of your real writing: how long your sentences run, how often you use contractions, how many times you reach for "I." Generic in, generic out. Your archive in, something in the neighborhood of your voice out. The machine can't invent your voice. It can only imitate one you hand it.
That's setup, and it happens before the first sentence a reader will ever see.
Then Do the Work It Can't
Now the part where you keep your voice: do the load-bearing work by hand, and let AI carry the rest.
Treat it as a research partner, not a ghostwriter. Gathering sources, condensing a transcript, hunting the counterargument you dodged, listing what you forgot. It's excellent here, and here its averageness is a feature. The moment it starts producing prose a reader will see, take the keys back.
Write the ugly draft first, by hand. Before I open any AI tool, I write the messy version: bad sentences, half-thoughts, the actual point in language I'd never publish. That ugly draft is the spine, and it carries the one thing no model can generate: a reason the piece exists beyond "the topic gets traffic."
Inject what can't be faked. The fastest way to de-machine a paragraph is to drop in something only you could know. When my last experiment's traffic cratered to zero overnight, no model on earth could have written "I refreshed three times thinking it was a bug." A model wasn't there. You were. Specificity is a watermark that doesn't wash out.
Break the structure on purpose. AI loves the four-beat march, setup then problem then solution then takeaway, in every section, forever. So end a section early. Open with the conclusion. Let a paragraph run one line.
Like this.
That unevenness is literally what burstiness measures, and it's the one axis you can move by hand.
You can even let the model clean up after itself: hand it your draft and have it flag its own hedging and limp transitions before you rewrite them. Just be honest about what that is. It's janitorial. It will sweep the floor. It will not decide what the room is for.
That is the division of labor, and it's simple. Hand AI the invisible labor: the research, the transcript, the third rewrite of a clunky sentence, the impatient editor who keeps asking "so what?" Spend your own keystrokes on the parts a reader's trust rides on — the opening line, the argument, the joke, the paragraph that has to land. Let AI do what's true of everyone. Keep what's true only of you.
The Red Flags Worth Hunting
Once you've drafted, go hunting. Not for banned words. For four patterns.
The hedge pile. "It's worth noting that, in many cases, it could potentially be argued…" Forty words, zero claims. AI hedges because it's trained never to be wrong. You're allowed to be wrong. Cut every qualifier that isn't load-bearing.
The fake transition. "Let's dive in." "Here's the thing." Throat-clearing the model does while it works out what to say. You already know what you're saying. Delete the runway.
The empty intensifier. "Robust." "Comprehensive." "Seamless." Words that cosplay as substance. If deleting the adjective changes nothing, it was ballast.
And hunt the structural tells from earlier too: the contrastive negation, the "-ing" tails.
One test beats every checklist. Read it out loud and ask whether you'd actually say this. Not "is it grammatical," but: would these words come out of your mouth, to a real person, without you wincing? If you'd never say "in the ever-evolving landscape of content," don't leave it sitting under your name.
What the Seam Costs
Your byline is the thing you're actually protecting. Ask Sports Illustrated.
In late 2023, the outlet Futurism discovered that the magazine, which once ran William Faulkner, had been running product reviews under authors who did not exist.[15] "Drew Ortiz," keen outdoorsman, had a warm little bio and a face bought off an AI-headshot marketplace. In a review, "he" warned that volleyball "can be a little tricky to get into, especially without an actual ball to practice with." The staff union said it was "horrified." Within two weeks, the parent company's stock had dropped 27% and its CEO was gone.
CNET ran a quieter version and paid the same bill. It published finance explainers under the byline "CNET Money Staff," which sounded like people. One explained compound interest by claiming that $10,000 at 3% would earn you $10,300 in a year.[16] Off by a factor of ten, in an article telling readers what to do with their savings. More than half of the AI-written pieces needed corrections. Wikipedia's editors downgraded CNET from "generally reliable" to "generally unreliable." In 2024 it sold for around $100 million, a fraction of what it reportedly fetched four years earlier.[17]
Notice what actually cost these outfits. It wasn't that they used AI. It was getting caught at the seam: the exact spot where the machine's voice, or its math, or its fake face showed through. Trust is the only thing a publication really owns, and counterfeiting it gets priced in fast.
And you can't buy that trust back by passing a detector, because the detectors are broken too.
The Point
The best detector money can buy, according to a University of Chicago Booth study this year, is a tool called Pangram, with near-zero false positives.[18] The free one most people actually paste into, ZeroGPT, flags roughly one in three genuine human essays as machine-made. It once rated the U.S. Constitution 92% AI. Same text, opposite verdicts, depending on which box you choose. Passing one proves nothing and failing one proves less. The goal was never to beat a detector.
Same text. Opposite verdicts.
How often each detector flags genuine human writing
Every tell in this piece has been a layer, and the layers peel at different speeds. Lexical ones vanish in months. Structural ones peel the moment someone names them. The mechanism underneath is just physics. But the bottom layer never moves: a model trained on the entire internet still wasn't in the room when your thing broke. It doesn't have your specific axe to grind. It can't tell the story only you can tell. The next patch notes won't touch that. It's the one thing that's structurally yours.
So stop trying to make AI sound human. That's the half of the race you lose. Put the human in first, your angle, your specifics, your opinions with teeth, and let the machine build the scaffolding around it. The 72% running the seam straight down the middle of the page will go on sounding like machines in a name tag.
AI can help you build something worth a person's attention faster than ever. It just can't be the reason it's worth reading.
That part's still on you. It always was.
References
- Zhang, M., Wu, L., Yang, T., Zhu, B., & Liu, Y. (2024). Retraction notice to "The three-dimensional porous mesh structure of Cu-based metal-organic-framework - Aramid cellulose separator enhances the electrochemical performance of lithium metal anode batteries". Surfaces and Interfaces, 46, 104550. https://doi.org/10.1016/j.surfin.2024.104550
- Soulo, T., & Guan, X. (2025, May 19). 74% of new webpages include AI content (Study of 900k pages). Ahrefs.
- Smith, E., & Druck, G. (2025, October 14). How does AI-generated content perform in search and answer engines? Graphite.io.
- Kobak, D., Kuznetsova, R., & Varoquaux, G. (2024). Delving into ChatGPT usage in academic writing through excess vocabulary. arXiv. https://arxiv.org/abs/2406.07016
- Graham, P. [@paulg]. (2024, April 7). My point here is not that I dislike "delve," though I do, but that it's a sign that text was written by ChatGPT [Tweet]. X.
- Hern, A. (2024, April 16). TechScape: How cheap, outsourced labour in Africa is shaping AI English. The Guardian. https://www.theguardian.com/technology/2024/apr/16/techscape-ai-english-outsourced-labour-africa
- Juzek, T. S., & Ward, Z. B. (2025). Why does ChatGPT "delve" so much? Exploring the sources of lexical overrepresentation in large language models. Proceedings of the 31st International Conference on Computational Linguistics, 6397–6411.
- Altman, S. [@sama]. (2025, November 14). Small-but-happy win: If you tell ChatGPT not to use em-dashes in your custom instructions, it finally does what it's supposed to do! [Tweet]. X.
- WikiProject AI Cleanup. (2023). Wikipedia:WikiProject AI Cleanup. Wikipedia, The Free Encyclopedia.
- Wikipedia Contributors. (2023, December 4). Wikipedia:List of hoaxes on Wikipedia/Less than one year. Wikipedia, The Free Encyclopedia.
- Chen, S. (2026, January 17). blader/humanizer: Claude Code skill that removes signs of AI-generated writing from text [Source code]. GitHub. https://github.com/blader/humanizer
- Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7). https://doi.org/10.1016/j.patter.2023.100779
- Cheng, Q. (2026, March 17). Via negativa for AI alignment: Why negative constraints are structurally superior to positive preferences. arXiv.
- Mann, L., Saxena, N., Tandon, S., Sun, C., Toteja, S., & Zhu, K. (2025, November 15). Don't think of the white bear: Ironic negation in transformer models under cognitive load. arXiv. https://arxiv.org/abs/2511.12381
- Harrison, M. (2023, November 27). Sports Illustrated published articles by fake, AI-generated writers. Futurism. https://futurism.com/sports-illustrated-ai-generated-writers
- Guglielmo, C. (2023, January 25). What is compound interest? CNET.
- Ziff Davis, Inc. (2024, August 6). Ziff Davis announces acquisition of CNET [Press Release].
- Jabarian, B., & Imas, A. (2025, August 26). Artificial writing and automated detection (BFI Working Paper No. 2025-116). Becker Friedman Institute, University of Chicago.
Mochi Nguyen writes about AI, product development, and where technology meets content strategy.
Comments (0)
No comments yet. Be the first to share your thoughts.