I wrote 100 articles by hand. A machine called me a liar.
This was during my experiment—the one where I tried to rank a website on pure white-hat SEO. Every article original. Every sentence mine. I'd researched, drafted, rewritten, killed my own darlings at 2 a.m. Real human work, the kind that leaves a mark on you.
Then, out of curiosity, I pasted one of my pieces into an AI detector. The ones universities swear by. The ones that decide whether a student keeps their degree.
92% AI-generated.
I tried another article. 78%. Another. 64%. The piece I'd agonized over most—the one with my actual voice bleeding through every line—scored highest of all.
The verdict on my own writing
AI-detector score · articles I wrote by hand
I sat there for a minute, unsettled. Not because anyone was about to expel me. Because the tool wasn't measuring whether AI wrote something. It was measuring whether the writing was clean. And clean is exactly what I'd spent years learning to do.
That's the dirty secret of AI detection. It doesn't catch machines. It catches anyone who writes well.
Everyone wants the same simple answer: "Is this AI or human?" Detector companies promise 99% accuracy. Universities stake academic careers on the verdict. Google claims to filter AI spam from search. The FTC and the EU have gotten involved too.
They were all selling the same lie. Not because they're evil—because the problem they claim to solve is mathematically impossible. Which is a fun thing to build a billion-dollar industry on top of.
So I pulled the thread, expecting a broken product I could route around. I found wreckage.
The Body Count
Start with the people it already destroyed.
Haishan Yang had a PhD in economics and was deep into a second doctorate at the University of Minnesota. International student. Decades of real academic work behind him. Then a detector flagged his exam.
The damning evidence? He used the acronym "PCO"—Primary Care Organization—a standard industry term his professors didn't happen to know, but one that appears in thousands of published papers. The detector itself couldn't keep its story straight: 89% probability on one answer, 19% on another. A coin flip with a confidence interval, treated as expert testimony.
Didn't matter. Yang was expelled. He called it "a death penalty." His student visa was revoked. He's stranded abroad now, suing for over a million dollars.[1]
Yang at least had the standing to fight on his own terms. Orion Newby, enrolled at Adelphi University in a program for students with learning disabilities, got accused on the strength of a single Turnitin score. He produced evidence of working with tutors. Documented 15 to 20 hours on the assignment. The university shrugged and gave him a zero.
His family spent over $100,000 in legal fees fighting it. In January 2026, a New York judge ruled the accusation "without valid basis and devoid of reason."[2] Newby won. Most people can't afford to even start that fight.
And Newby's case points at the cruelest part of this whole machine. He was a student with a learning disability—which is precisely the profile detectors love to flag.
Think about what these tools reward. Varied sentence rhythm. Idiomatic flair. Unpredictable word choices. Now think about how a lot of neurodivergent people are taught to write. Autistic writers often favor literal, hyper-specific, direct language—logic over rhetorical decoration. Students with dyslexia and ADHD get coached by disability services to use rigid outlines and sequential transitions: First. Next. Finally. Someone with ADHD writing in a hyper-focused burst produces eerily uniform sentence lengths.
To the algorithm, every one of those coping strategies reads as a robot. False-positive rates for highly divergent writing styles can run north of 50%. We built a tool that interprets a disability accommodation as a confession.
It gets stupider. In May 2023, an agricultural sciences professor at Texas A&M-Commerce decided to police cheating by pasting his students' final essays into ChatGPT and asking it whether it had written them. ChatGPT, a machine with no memory of past conversations and a deep need to be agreeable, said yes. Of course it did. Agreeing is the whole product.
The professor—who reportedly noted he doesn't "grade AI bull"—failed more than half the class. The university withheld diplomas from students who had already graduated.[3] The man used a hallucinating chatbot as a forensics lab, and an institution let him.
Now multiply that. Turnitin runs across more than 70 million students. The company's own Chief Product Officer eventually conceded the sentence-level false-positive rate sits around 4%—one in every 25 sentences flagged as machine-written is, in fact, human.[4] Do that math against 70 million people and you're guaranteeing millions of innocent students dragged into academic tribunals to defend writing they actually wrote.
This is what AI detection does in the wild. It doesn't catch cheaters. It picks targets, and it picks them from the people least equipped to fight back.
Why The Math Doesn't Work
My 92%. Yang's exam. Newby's zero. Every one of those verdicts fell out of the same two numbers—and once you see how the numbers are built, the whole thing stops looking like science.
Detectors measure two things: "perplexity"—how predictable your word choices are—and "burstiness"—how much your sentence length varies. The theory says AI writes smooth, even, predictable prose. Humans lurch around: a three-word fragment slamming into a forty-word monster of a sentence.
The fatal flaw? Good writing is smooth too.
Academic writing guides train students to do exactly what lowers perplexity. Use standard transitions. Keep terminology consistent. Remove ambiguity. A student who nails formal English produces text that is, statistically, indistinguishable from a language model. Bad writing is safe. Good writing is suspect. The winning move, apparently, is to write worse on purpose.
The em dash—this little horizontal line—is now treated as a confession. Language models, trained on professionally edited text, reach for it constantly. The average undergrad doesn't. So clean punctuation became evidence of a crime. Go read the student forums; they're full of kids coaching each other to strip out em dashes and dumb their prose down to survive software.
One finding should have ended the industry on day one. Stanford researchers ran seven detectors over 91 TOEFL essays—confirmed human writing from non-native English speakers. The average false-positive rate came back at 61.22%.[5] Ninety-seven percent of the essays were flagged by at least one detector. Nearly one in five got unanimously branded as machine-written by all seven. Native-speaking American eighth graders, run through the same tools? Near zero.
Same essays. Opposite verdicts.
Confirmed human writing flagged as AI · 7 detectors
The detail that turns bias into farce: when the researchers asked an AI to rewrite those same essays with "more sophisticated, literary language," the detectors flipped them back to "human." The tools don't detect machines. They detect a limited vocabulary. They punish the person who learned English as an adult and reward the one who can ask a chatbot for fancier words.
A 2026 study from Pindrop, slated for one of the field's top conferences, pushed the knife in further. Across 16 detection systems, false positives clustered hard on English-language learners—and non-White learners got flagged at higher rates than their White peers. The bias is intersectional: non-White, male, English-learning, worst of all. When humans were asked to do the same detection by hand, they showed no statistically significant demographic bias. The machines invented a prejudice the people didn't have.[6]
But bias is the symptom. Here's the disease—the one nobody selling these tools wants you to understand.
This is where I went from angry to resigned. Researchers at the University of Maryland proved—proved, with the math—that reliable detection is impossible. Not hard. Not unsolved. Impossible. Their argument rests on the "total variation distance" between two things: the distribution of human writing and the distribution of machine writing. A detector's best possible accuracy is mathematically capped by how far apart those two distributions sit.[7]
Why it can’t be fixed
As the models improve, the two distributions converge
Now remember what a language model is built to do. Its entire purpose, the literal target of all that training and alignment, is to make machine writing indistinguishable from human writing. Every improvement to the model shrinks the gap. And as the gap shrinks, the ceiling on any detector's accuracy drops with it—toward a coin flip.
So the better AI gets, the worse detection gets, by definition. The two move in lockstep in opposite directions. The popular fantasy that detectors will simply "catch up" has it exactly backwards: the thing detectors chase is the thing that destroys them. That's not an engineering gap you close next quarter. It's a perpetual motion machine.
The Accuracy Death Spiral
You don't have to take the theorem on faith. Watch the numbers fall off a ledge.
2023: detectors hit 95%+ against GPT-3.5. 2024: down to 70-80% against GPT-4. Against today's models, real-world accuracy lives in the teens. A 2024 study by Perkins and colleagues ran six major commercial detectors across 805 samples and found a baseline accuracy on raw, unedited AI text of 39.5%. Then they applied the laziest evasion imaginable—light manual edits, a few typos, a paraphrasing pass—and accuracy collapsed to between 17% and 22%. Turnitin, the industry standard, dropped 42 percentage points the moment anyone tried even slightly.[8]
The accuracy death spiral
Real-world detector accuracy as the models improve
Show the figures
| Model era | Real-world accuracy |
|---|---|
| 2023 (GPT-3.5) | 95% |
| 2024 (GPT-4) | 75% |
| 2026 (today) | 16% |
The most damning admission came from the source. In January 2023, OpenAI—the company that built ChatGPT, the people with the model weights, the training data, every possible advantage—launched its own AI text classifier. If anyone could detect ChatGPT, it was the people who made it.
It correctly identified 26% of AI text. It falsely flagged 9% of human text. It couldn't handle anything under 1,000 characters and fell apart in any language but English. In July 2023, six months later, OpenAI quietly killed it.[9] The creators of the machine could not detect the machine. That should have been the obituary for the entire field.
Meanwhile the other side runs up the score for free. "Humanizer" tools—Undetectable AI, Rephrasy, a dozen others—exist for one purpose: take AI text and launder it past detectors. They work by recursive paraphrasing. Generate the text with one model, then run it through a second that rewrites every sentence, scrambling the statistical fingerprint while keeping the meaning. The signature evaporates. The essay survives. It costs roughly nothing.
And it cuts the other way, which is the part that should scare you. The same trick that hides a machine can frame a human. Researchers have shown you can probe a model enough to learn its statistical signature, then inject that signature into real human writing. Feed the doctored text to a detector and it screams "AI." You can now weaponize an academic-integrity tool against anyone you dislike. The detector can't tell the difference, because there is no difference left to tell.
Someone ran Charles Dickens through the gauntlet. A Christmas Carol, fed through multiple detectors. Verdict: 95% AI-generated. Dickens died in 1870, a century and a half before ChatGPT existed to copy. If the man who wrote "it was the best of times" reads as a robot, the premise was never real.
A growing list of schools has quietly accepted this. Vanderbilt switched off Turnitin's AI detector back in 2023.[10] Waterloo, Cornell, Johns Hopkins, Pittsburgh followed. MIT Sloan issued guidelines against the tools. But most institutions still cling to them, which means most students are still guilty until proven human—a thing you cannot actually prove.
Security theater with a body count.
The Watermarking Mirage
When statistical detection died, the industry didn't admit defeat. It pivoted to watermarking—the seductive idea that you don't need to detect AI if the AI signs its own work at birth. Embed an invisible signature into the text the moment it's generated, and detection becomes a key-check instead of a guess.
Elegant. Also doesn't hold.
The leading method comes from Kirchenbauer and colleagues.[11] At every word, the model secretly splits its vocabulary into a "green list" and a "red list," then nudges itself to pick green words. A human, blind to the split, scatters green and red evenly. AI text leans green. Run the stats, catch the watermark.
Two problems, both fatal. First, the detection paradox: the harder you push the model toward green words, the more you force it to choose worse ones. Crank the watermark up and quality craters—on some constrained tasks, by nearly 97%. You can have a strong watermark or a good model, not both. Second, the whole scheme depends on the exact sequence of words. Swap in synonyms—the same paraphrasing that defeats every other detector—and the signature is gone.
Google DeepMind built the most ambitious version, SynthID, and scaled it across Gemini, its image models, its music models.[12] DeepMind says the mark survives mild edits. It survived about a weekend. The method, again, is regeneration: feed the watermarked text to a second AI and have it rewrite the thing from scratch, drawing from an unwatermarked distribution. Meaning preserved. Watermark erased. For images, the trick is "re-noising"—extract the picture's structure, regenerate it through a fresh diffusion pass, and the original pixel-level signature dissolves while the image looks identical.
This is the unbreakable loop at the center of the problem. Anything one AI can embed, another AI can wash out, and the washing costs less than the embedding. Watermarking isn't a solution. It's the same arms race wearing a cryptography costume.
Google Can't Detect AI Either
Now scale the failure up to the entire web, and watch the biggest company in the room fold in slow motion.
In April 2022, Google's John Mueller called AI-generated content spam, flatly, comparing it to old black-hat article spinning. Auto-generated, against the rules, eligible for penalties. Clear as day.
But the models got good, and Google did the math everyone eventually does. By February 2023, the official guidance had softened to rewarding "high-quality content, however it's produced." E-A-T quietly gained an extra E for Experience—the one thing a machine supposedly couldn't fake.
Then, in September 2023, came the tell. Google's Helpful Content documentation had long promised to reward content "written by people, for people." They deleted two words. It now read "content created for people." The phrase "written by people" simply vanished, because enforcing it would have meant penalizing automated content that's existed since long before ChatGPT—and that they couldn't reliably spot anyway.
E-E-A-T was supposed to be the firewall. AI can't have lived experience, the thinking went. So experience became the next thing to fake. Travel sites spin up AI headshots and invented bios. "Meet Jessica, our senior travel editor with 12 years of experience"—generated this morning, no face behind the face. Google can validate that the authorship data is structured correctly. It cannot validate that Jessica exists. E-E-A-T isn't a truth filter. It's a credential check, and credentials forge beautifully.
We don't have to guess about any of this, because in May 2024, more than 14,000 of Google's internal ranking attributes leaked onto GitHub.[13] The single most important thing in that document is what isn't in it: there is no binary "this is AI" flag. None. The most sophisticated search company on earth did not build the detector everyone assumes it has, because it can't.
What they built instead is more honest about the real problem. There's an attribute called contentEffort—an LLM-based estimate of how much actual work a page took to make. The logic is brutal and correct: if a language model can cheaply reproduce your article, your article is worth nothing, regardless of who typed it. There's originalContentScore, rewarding novelty over length. There's siteAuthority. And there's NavBoost, which watches real clicks from Chrome—good clicks, bad clicks, whether you bounce straight back to the results—and demotes anything people don't actually find useful.
See the shift? Google stopped asking "did a human write this" and started asking "could a machine trivially redo this." That's not a detection question. It's a replaceability question, and it's the only version of this that works.
And even that isn't keeping up. In January 2024, researchers from Weimar and Leipzig published a year-long study across 7,392 product-review queries and concluded what everyone already felt: search is degrading, the top results are drowning in affiliate spam, and Google is losing the cat-and-mouse game.[14] Two months later Google fired its biggest weapon—the March 2024 core update, aiming to cut low-quality content by 45%. It introduced "Scaled Content Abuse," which retired the old "auto-generated content" rule entirely. The new crime isn't using AI. It's mass-producing junk, by any means. Origin doesn't enter into it.
And the punchline, in 2026: Google now generates the slop itself. AI Overviews sit on top of the results, synthesized by Gemini, answering your question so you never click. Zero-click rates around 83%. The company spent two years learning it couldn't detect AI content, then became the largest publisher of it—while still scraping the human sites underneath for the answers it serves. The house that penalized scrapers became the scraper.
The Law Catches Up
Here's how you know an industry's core promise is broken: the regulators start writing it down.
In 2024 and 2025, the FTC ran an enforcement sweep called Operation AI Comply[15], and one of its targets was a detector company called Workado. Workado sold AI detection for $49 a month, advertising 98% accuracy at spotting ChatGPT, Claude, and the rest. The FTC looked under the hood and found the company had taken an open-source model some Norwegian students built to flag AI in academic abstracts, never retrained it for general text, and shipped it. On real-world content, its accuracy was about 53%. The FTC's verdict, in writing: "no better than a coin toss."[16]
Advertised accuracy vs. reality
What detectors claim · what they actually score
The consent order is the part worth screenshotting. Workado is now barred from making accuracy claims without "competent and reliable scientific evidence," has to keep its testing data for 20 years, and has to email its own customers to tell them the tool doesn't work as advertised. A federal agency has formally established that selling AI detection as reliable is deceptive.
Across the Atlantic, the EU took the more interesting route. Rather than pretend detection is coming, Article 50 of the EU AI Act—live as of August 2026—drops the pretense and shoves the burden onto the people making the content. AI output has to be labeled for humans and marked in a machine-readable way, leaning on the C2PA provenance standard that Adobe, Google, and Microsoft have all signed. Free-form AI text over 200 tokens has to carry a watermark.[17] Miss the rules and the fines reach into the millions, or a slice of global revenue.
Read that closely. The most serious AI legislation on the planet has concluded that you cannot detect this stuff after the fact, so the only hope is to tag it at the source. Except—as we've seen—metadata strips off the instant someone screenshots a page, and watermarks wash out the instant someone regenerates the text. The law mandates a guarantee the technology can't make. When both the FTC and the EU are effectively conceding the thing can't be done, the debate is over. They're just negotiating the terms of surrender.
What Still Works (For Now)
So if the detectors are broken, the watermarks wash out, Google gave up, and the law is writing the eulogy—how do you actually spot AI?
You stop trusting software and start trusting your own pattern recognition. It still works. For now. Here's the field guide.
The Delve Index
A study combing through 15 million scientific abstracts found "delve" showing up 28 times more often in post-ChatGPT writing than before.[18] Not 28% more. Twenty-eight times. The rest of the starter pack: "tapestry," "landscape," "multifaceted." If a piece promises to explore "the rich tapestry of factors in today's rapidly evolving landscape," relax—no human did that to you.
The Opening Disease
"In today's rapidly evolving digital landscape, businesses must navigate the ever-changing terrain..."
This is the "Once upon a time" of the corporate internet. It means nothing. It fits any year, any topic, any company. The world has been "rapidly evolving" since the steam engine. The model deploys it to buy a few seconds while it works out what it's actually going to say.
Sarah, The Busiest Woman In AI
"Picture this: Sarah, a marketing manager at a mid-sized tech firm, is drowning in spreadsheets."
Meet Sarah. She is the protagonist of the AI universe. Always a marketing manager. Always at a mid-sized firm. Always drowning in data, always discovers the solution, always emerges efficient and grateful. She appears in roughly 100% of AI-generated case studies and she has never once existed. When a stock human materializes to demonstrate a product, you're reading a machine.
The Hedge Olympics
"It's worth noting that, in many cases, it could potentially be argued that quality might, perhaps, play a somewhat significant role..."
Forty words. Zero information. Models hedge everything because they're trained to be helpful, harmless, and honest—which, in practice, collapses into never saying anything that could be proven wrong. Humans with opinions commit. Machines covering their bases do this.
The Model Dialects
The accents are diverging. You can name the model now. Claude opens with "I'll be direct" and "Here's the thing"—the preacher clearing his throat, performing candor. Gemini says "Let's dive in"—the tech YouTuber hitting the transition. GPT-4o drops "Look," at the front of a sentence—fake-casual, a baseball cap worn over a tuxedo.
The Oops Artifacts
Sometimes the tell is printed, bound, and peer-reviewed. More than 100 published academic papers have been caught containing the phrase "As an AI language model, I cannot..." or the stray "Regenerate response"—the ChatGPT button label, copy-pasted straight into the manuscript.[19] These cleared peer review. They sit in real journals right now. Someone forgot to delete the receipt stapled to the back of their homework, and three reviewers signed off anyway. All that scrutiny aimed at undergrads, and nobody was checking the professors.
The Uncomfortable Truth
Here's the part that ruins my own field guide: every tell I just handed you is already dying.
Models train on the feedback that names these patterns. They're learning to drop "delve" and "tapestry." Learning to vary their sentences. Learning to fake an opinion well enough that you'd swear it had one. The crutches are getting sanded off, one update at a time. That's not a side effect—it's the same convergence the Maryland proof described, happening on your screen.
The window for human pattern recognition is closing, and I can feel it closing while I write this. A year from now this section is obsolete—not because the machines got worse at pretending to be us, but because they got better.
What survives is the inverse skill. Stop hunting for what's wrong in the text. Start noticing what's missing. The jagged edge. The detail too specific to invent. The willingness to be wrong out loud. The smell of the actual pizza shop instead of "a bustling restaurant."
And honestly, that's the wrong fight anyway. The real question was never "did AI write this." It's "does that even matter." A tool that expels Haishan Yang and bankrupts Orion Newby's family while the actual cheaters launder their essays past it for free isn't a solution. It's a liability with a subscription fee.
The sharpest institutions already figured this out and stopped asking the unanswerable question. They moved to oral exams, process portfolios, writing done in the room where someone can watch. And the ones who kept using detection at all redirected it somewhere it actually works: for its 2026 conference, the field's flagship NLP organization stopped scanning submissions for "AI style" and instead pointed a tool at the bibliographies, checking whether the cited papers actually exist. They gave up on "is this human" and started asking "is this true." That's the only detection question with an answer.
The point was never catching the machine. The point was learning, and honesty, and whether the work is real. No detector ever measured any of that. Mine certainly couldn't measure the year I spent writing those 100 articles by hand—it just looked at the polish and called me a fake.
Turns out the machine and I had something in common after all. Neither of us can tell the difference anymore.
References
- Minnesota Court of Appeals. (2026, February 2). Haishan Yang, Relator, v. University of Minnesota, Respondent (A25-1019). Thomson Reuters.
- Supreme Court of the State of New York. (2026, January 28). Matter of Newby v. Adelphi Univ. (Nassau County, CPLR Article 78).
- Texas A&M University-Commerce. (2023, May 17). Texas A&M University-Commerce addresses concerns about ChatGPT in Ag classroom. TAMUC News Desk. https://new.tamuc.edu/news/texas-am-university-commerce-addresses-concerns-about-chatgpt-in-ag-classroom/
- Chechitelli, A. (2023, June 14). Understanding the false positive rate for sentences of our AI writing detection capability. Turnitin. https://www.turnitin.com/blog/understanding-the-false-positive-rate
- 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
- Stowe, K., Afanaseva, S., Raimundo, R., Sun, Y., & Patil, K. (2024). Identifying bias in machine-generated text detection. Pindrop / arXiv. https://arxiv.org/abs/2408.01234
- Sadasivan, V. S., Kumar, A., Balasubramanian, S., Wang, W., & Feizi, S. (2023). Can AI-generated text be reliably detected? arXiv. https://arxiv.org/abs/2303.11156
- Perkins, M., Roe, J., Vu, B. H., & Khuat, H. (2024). Simple techniques to bypass GenAI text detectors: Implications for inclusive education. ResearchGate.
- OpenAI. (2023, January 31). New AI classifier for indicating AI-written text [Updated July 20, 2023]. OpenAI Blog. https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/
- Vanderbilt University. (2023, August 16). Guidance on AI detection and why we're disabling Turnitin's AI detector. Vanderbilt University Brightspace.
- Kirchenbauer, J., Geiping, J., Wen, Y., Katz, J., Miers, I., & Goldstein, T. (2023). A watermark for large language models. Proceedings of the 40th International Conference on Machine Learning (ICML). https://arxiv.org/abs/2301.10226
- DeepMind. (2023). SynthID: Identifying AI-generated content. Google DeepMind. https://deepmind.google/technologies/synthid/
- King, M. (2024, May 27). Secrets from the algorithm: Google Search's internal engineering documentation has leaked. iPullRank. https://ipullrank.com/google-algo-leak
- Bevendorff, J., Wiegmann, M., Kiesel, J., Hagen, M., & Stein, B. (2024). Is Google getting worse? A longitudinal investigation of SEO spam in search engines. Leipzig University & Bauhaus-Universität Weimar. https://downloads.webis.de/publications/papers/bevendorff_2024a.pdf
- Federal Trade Commission. (2024, September 25). Operation AI Comply: Detecting AI-infused frauds and deceptions. United States Federal Trade Commission. https://www.ftc.gov/news-events/news/press-releases/2024/09/ftc-announces-crackdown-deceptive-ai-claims-schemes
- Federal Trade Commission. (2025, April 24). FTC order requires Workado to back up artificial intelligence detection claims. United States Federal Trade Commission.
- European Parliament and Council of the European Union. (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union.
- Kobak, D., Kuznetsova, R., & Varoquaux, G. (2024). Delving into ChatGPT usage in academic writing through excess vocabulary. arXiv. https://arxiv.org/abs/2406.07016
- Cabanac, G. (2024). Tortured phrases and ChatGPT tells in academic papers [Dataset and PubPeer independent tracking]. PubPeer Foundation.
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.