On July 20, 2023, OpenAI quietly took its own AI detection tool offline. Six months earlier, the company behind ChatGPT had launched the classifier with the warning printed right on the announcement. It caught 26 percent of AI-written text. It mislabeled 9 percent of human writing as machine-made.
The note they left in its place is just one sentence long, “the classifier is no longer available due to its low rate of accuracy.” Five days after that, on July 25, the company where I was president sent out a press release about our own free AI detector. It had reached 98.3 percent accuracy, the release said.
I ran Content at Scale from the spring of 2023 until October 2024, when I left to start First Movers. The detector was the free thing on our site that pulled people in, and I was the person on the podcasts and the conference stages talking about the company that built it.
Somewhere in that stretch, I started to get pretty curious. I pulled ten articles, some that real people had written and some straight out of the big AI platforms, and ran the lot past the detectors marketers were paying for. Every single one came back flagged as AI. The human ones came back flagged right alongside the machine ones, and that’s about when I stopped being able to say the word accuracy with a straight face.
So, is AI detection a scam? I get asked that on almost every Q&A now, usually from a student or a freelancer holding a screenshot with a red number on it. My answer has changed since 2023, and so have the tools. In 2023 they flagged people. In 2026 they clear machines.
What AI detection measures, and what it can’t
An AI detector works from the text you give it, looking for patterns it has learned to associate with AI writing. Two terms come up a lot when you start reading about how these tools work:
- Perplexity measures how easy it is for a language model to predict the next word. Familiar wording and predictable phrasing tend to produce a lower score, which can make the writing more likely to be flagged.
- Burstiness describes variation across the writing, often explained through sentence length and structure. You might spend a long sentence explaining something, then follow it with a few words. A more uniform pattern can look like AI to a detector.
GPTZero’s homepage still mentions both, although it says its model considers hundreds of factors. So the idea that every detector just checks those two things is too simple.
The percentages need some explaining too, because each tool reports them differently. Turnitin estimates how much of the text it can assess looks AI-generated. If that figure falls between 1 and 19 percent, it shows an asterisk because results in that range are less reliable. GPTZero gives probabilities for whether the document is human, AI or a mixture, alongside sentence highlights. Pangram reports the estimated share of AI-generated or assisted text, with a confidence rating.
That means two results showing “40%” can mean different things. Before I read much into either, I’d want to know which tool produced it and what that percentage refers to.
There was another detail in OpenAI’s classifier launch notes beyond the 26 percent detection rate: it said the tool was very unreliable on text shorter than 1,000 characters. It also struggled with highly predictable writing. That leaves an obvious problem for anyone checking a short, straightforward email. There may be very little in the wording to distinguish something you typed yourself from something a model produced.
We described our Content at Scale detector quite plainly on Product Hunt back in March 2023. The listing said it was checking “how robotic sounding the content is.”
I still think that wording helps explain why these results need care. People can sound robotic too, especially when they’re trying to write something professional. A detector may flag that writing, but the result alone can’t tell you how it was written.
How accurate are AI detectors? The record from 2023 to 2026
In 2023 an independent test of 14 detectors put every one of them under 80 percent accuracy, and the people getting flagged were human. By 2026 that had reversed: the newest test found Turnitin scoring every fully AI-written paper as 0 to 20 percent AI. I wish I’d had this table in July 2023. Each row comes from an independent test, with the date and sources included.
| Date | Who tested | What they found | Which way it erred |
|---|---|---|---|
| Jan 31 to Jul 20, 2023 | OpenAI, on its own classifier | Caught 26% of AI text, flagged 9% of human text. Withdrawn after six months for low accuracy | Both ways |
| Jul 10, 2023 | Stanford, Patterns (Liang et al.) | Seven detectors. 61.3% of non-native English essays flagged as AI. US eighth-grade essays passed | Against people |
| Aug 16, 2023 | Vanderbilt University | Disabled Turnitin’s detector. At the vendor’s 1% false-positive claim, about 750 of the 75,000 papers Vanderbilt submits a year would be wrongly flagged | Against people |
| Dec 2023 | Weber-Wulff et al., International Journal for Educational Integrity | 14 tools, 756 tests. None above 80% accuracy, five above 70%. Roughly 20% of AI text read as human | Both ways, mostly missing AI |
| Oct 18, 2024 | Bloomberg Businessweek | 500 Texas A&M application essays from summer 2022 run on GPTZero and Copyleaks. 1% to 2% flagged, some at near 100% certainty | Against people |
| Apr 18, 2025 | University of Chicago IT Services | GPTZero caught nearly every AI sample and cleared human text 99% of the time. ZeroGPT called an entirely human paragraph 100% AI | Split, depends on the tool |
| Jun 29, 2026 | Vrije Universiteit Brussel, International Journal for Educational Integrity | 160 papers, four tools. No false positives for Pangram, Copyleaks or Turnitin. Turnitin scored 100% of fully AI-written papers as 0 to 20% AI. GPTZero caught 0 of 40 hybrid papers. Pangram caught 97.5% of fully AI papers on the study’s inclusive measure and 92.5% of humanized ones | Missing AI, except Pangram |
Back in 2023, people were getting flagged for work they’d written themselves. Early research found that this was a particular problem for writers whose first language wasn’t English. You can see some of that caution in how results are reported now. Turnitin, for example, hides scores between 1 and 19 percent because false positives are more common in that range. But avoiding false accusations is only part of the job. A detector still needs to recognize AI writing.
The Brussels study, published in 2026, shows how difficult that had become. The team generated its full papers using ChatGPT’s Deep Research feature in May 2025. GPT-5 arrived while they were finishing the study. Of the four detectors tested, only Pangram produced results the researchers considered satisfactory.
Pangram’s training method may help explain why. The company pairs human writing with AI versions matched as closely as possible in topic, tone, style and length. It also finds human examples the detector gets wrong and uses those, alongside their AI counterparts, for further training. That makes it harder for the model to get by on obvious differences in wording or format. The researchers thought this might explain why their humanizing prompt had little effect on Pangram’s results, although they couldn’t establish the cause.
What surprised me was Turnitin’s performance. It put all 40 fully generated papers in the study’s 0–20 percent band, yet reached 60 percent accuracy on papers mixing student writing with AI sections. The authors suspect the model difference mattered: Deep Research produced the full papers, while standard GPT-4o wrote the inserted sections.
The team then scanned 1,163 master’s theses from 2024–2025. Pangram flagged 45.5 percent, mostly at low or moderate levels; very high scores were rare. Those figures describe what Pangram flagged. The researchers didn’t know how the theses had actually been written, and a score couldn’t tell them whether a student had broken a rule.
I’d also include the people who took part in a 2025 ACL study. Researchers paid readers to identify nonfiction articles written by humans and by GPT-4o, Claude and o1. Five participants who frequently used ChatGPT for writing were particularly good at it. Across 300 articles, half human and half AI, their majority vote got 299 right, including tests involving paraphrasing and humanization.
Their explanations were pretty interesting too. They noticed familiar AI vocabulary, but also how formal the writing sounded and whether it had anything original to say. In that study, their combined judgment beat most of the detectors tested. Spending time writing with these tools seems to teach you quite a lot about recognizing what comes out of them.
So, is AI detection a scam?
No, and I have a 98.3 percent press release in my past, so I’m not saying that to be nice. AI detection is real statistics with a real signal inside it. The trouble is the signal is weakest exactly where people want to use it, on one short piece of writing and one person’s reputation. Where the scam creeps in is three places, and I’ve stood in all three.
1. The accuracy claims.
Our 98.3. Turnitin’s under 1 percent false positives at launch. The 99 percent on half the detector homepages today. Every one of those numbers was measured on the vendor’s own test set, and the independent record above has never matched a single one of them.
Weber-Wulff’s team actually named four companies in its 2023 paper that were each claiming to be the best, Content at Scale was one of the four, and then scored every tool it tested under 80 percent. I didn’t enjoy reading that, and it was fair, which stung more.
2. The interface.
A probability shows up as a percentage bar and a color, and a tired reader sees a verdict. That’s how Moira Olmsted got a zero at Central Methodist University in the fall of 2023. Bloomberg Businessweek told her story the following October. She was seven months pregnant, back in school after time off to start a family, and a detector had read her assignment as machine-made.
She told her teacher she is autistic and that her writing runs in a set pattern a tool could easily mistake for a machine’s. She got the grade back after she pushed, along with a warning that the next flag would be treated like plagiarism.
3. The humanizers.
A whole industry now sells the same statistics run backward, a paste box that rewrites your draft until a detector shrugs. Every one of those rewrites pushes toward plainer, safer prose, because plain and safe is what the score rewards, and I said that in the first version of this post.
The Brussels study made it worse. One humanizer prompt got past three of the four tools in June 2026, and the fourth, Pangram, caught it 92.5 percent of the time anyway. None of the nine writing tools I compared this month ships a detector switch, and I’d be suspicious of one that did. The only test I’ve ever seen predict anything is a reader finishing the piece.
Does Google use AI detection to rank content?
No, and Google has said so in writing since February 8, 2023. That post, with Danny Sullivan and Chris Nelson’s names on it, says Google rewards high-quality content however it is produced, and that using automation, AI included, to generate content mainly to manipulate rankings breaks its spam policies.
On March 5, 2024 Google widened that rule into what it calls scaled content abuse, many pages generated mainly to rank, no matter how they were made. Its current page on generative AI content repeats the line and points you to two sections of the quality rater guidelines, 4.6.5 on scaled content and 4.6.6 on main content made with little effort, originality or added value. Read those two sections and notice they never once ask what tool you used.
At Content at Scale, we were putting over 50 million words a month of AI-assisted content onto customer sites in 2023 and 2024, so take the next sentence as a pattern I watched up close and nothing stronger.
The March 2024 update went after volume with nothing inside it, which is what the policy text says it went after. Google is asking one question of a page, would a person be glad they landed on this. A detector can’t answer that, and Google has never asked one to.
How I publish 500+ AI-made videos without hiding one
My YouTube channel passed 300,000+ subscribers on videos I never filmed, which still reads strangely when I type it. The avatar is a HeyGen video clone with an ElevenLabs voice, every script starts in my Claude system and I approve every one from my phone, and each video says on screen and in the script that you’re watching the clone.
That was my decision before it was anybody’s rule, mostly because enough viewers asked if I was a clone that I ended up writing a post with that exact title. The answer, sometimes, is yes.
YouTube made it a rule on March 18, 2024. Creators have to disclose realistic, altered, or synthetic content, and the first example on YouTube’s own list is a synthetic version of a real person’s voice.
The label sits in the description or right on the player. Using AI for scripts, ideas or captions doesn’t need one, which I think is the right line. I switch the disclosure on at upload, and I’d keep doing it if YouTube dropped the rule tomorrow, because the audience has been fine with the clone for exactly as long as I’ve been upfront about it.
Then the law caught up. Article 50 of the EU AI Act has applied since August 2, 2026, and if your content reaches Europe, it applies to you. Anyone deploying AI has to label deepfakes and AI-generated text on subjects of public interest, and the companies building the models have to mark generated audio, video, images and text so a machine can read the mark.
The Commission put out its final guidelines in July, but what struck me about them is what isn’t in there. Nobody is asked to run a detector. The whole direction is a mark at the source plus a label the audience can see, and a clone channel that has labeled itself from the first upload is already standing where the rules are headed.
What to do when AI detection flags your writing
You answer a detection flag with your process, and the evidence is already sitting in your drafts. This is the section I get the emails about. Students, freelancers whose clients run a checker on every draft, people whose own writing came back too tidy for a tool. This is the order I’d work in.
- Ask which tool, which version, and what threshold. Turnitin treats anything under 20 percent as uncertain and won’t print the figure. GPTZero’s probability and Turnitin’s percentage aren’t the same unit. A score with no threshold next to it is a number somebody is nervous about.
- Bring the process. Version history in Google Docs or Word, draft files with timestamps, your research notes, the messages you sent while you were writing it. Detectors only ever see the finished text, and the finished text is the least informative thing you own.
- Bring the record. Vanderbilt’s note from August 2023 explaining why it switched Turnitin’s detector off is short and written for administrators, which is who you’re talking to. Bloomberg’s 2024 test and the Brussels study from June 2026 cover the rest. If English is your second language, add the Stanford paper and underline the 61.3.
- Ask for a human read. Five heavy ChatGPT users voting together got 299 of 300 right. Ask for a reviewer who writes with these tools every day, and ask them to write down why they landed where they did.
- If you run a team, write the policy before you buy the tool. Mine is three lines. Say what the machine did on every piece, name the human editor who read it, and never let a score make a decision on its own.
Readers are the only detector that has ever mattered
Everything with my name on it starts in a Claude system loaded with 100,000+ words of my interviews, transcripts, and frameworks. We call it Voice DNA, and it’s the first thing we build for a client too, before the writer, before the clone, before anything else.
The drafts come out sounding like me on a good day. My team can barely tell which paragraphs I wrote. And then a human editor reads every single piece before it ships, and that editor answers to the reader, nobody else. The words heavy ChatGPT users clock in a second, the even rhythm, the neat little summary sentence bolted onto the end of every paragraph, all of it gets cut, because readers notice, and readers are the only detector I’ve ever cared about.
The method underneath hasn’t changed since the first version of this post. Topic, personal insight, call to action. Start with a topic a real reader has, put one thing in the piece that only you could know, and give them a single next step.
If you want that system inside your business, my team builds it in two months on the consulting page, Voice DNA first, then the writer, the clone, and the rest of the AI staff. If you’d rather learn it yourself, AI Labs is where I teach it, 65+ courses, weekly live sessions and 3,000+ students trained so far. And if the clone is what you came for, the two-day build is Clone Mastermind in Scottsdale.
FAQs about AI detection
Rank Math implementation note. Keep the following six questions and answers as visible on-page copy and paste them into the existing Rank Math FAQ block. Do not add a separate JSON-LD schema script to the article.
What is AI detection?
AI detection is software that estimates the probability that a passage of text, an image or a video was generated with AI. For text, the tools measure how predictable the wording is and how much sentence rhythm varies, then a trained classifier predicts a label. The output is a likelihood. No current tool can identify who typed a passage.
How accurate are AI detectors in 2026?
AI detector accuracy in 2026 depends on the tool and on which mistake you’re measuring. A June 2026 peer-reviewed test of 160 academic papers found no false positives for three of four detectors, while Turnitin scored every fully AI-written paper as 0 to 20 percent AI and GPTZero caught none of the hybrid papers. Pangram caught 97.5 percent of fully AI papers on the study’s inclusive measure and 92.5 percent of humanized ones.
Is AI detection a scam?
No. The statistics are real, and the signal is weak on short texts and on writers with plainer vocabulary. The misleading part is vendor accuracy claims measured on the vendor’s own data, and interfaces that turn a probability into a verdict. Treat a score as a reason to look closer, never as proof.
Does Google penalize AI content?
No. Google’s guidance since February 2023 says it rewards quality however content is produced. It penalizes scaled content abuse, many pages made mainly to rank, whichever way they were made, and its guidance points to the rater guidelines’ sections on low-effort main content, which apply to human and machine drafts alike.
Can AI detection flag human writing as AI?
Yes. Stanford found 61.3 percent of non-native English essays flagged in 2023, Bloomberg found 1 to 2 percent of pre-ChatGPT college essays flagged in 2024, and plain, formula-like writing styles run a higher risk. Rates were near zero in the June 2026 Brussels test, and a flag is still a probability, never a finding.
What should I do if AI detection flags my work?
Ask for the tool, the version and the threshold, bring your drafts and version history, cite the published record, and request a human read. Detectors judge the finished text, and your drafts are the evidence a score can’t argue with.
Where I land on AI detection
AI detection in 2026 is a decent first flag and a terrible judge. The tools that accused people in 2023 now clear machines. The one tool that catches machines still can’t tell you who typed the words. And the platforms and regulators who decide these things have moved to labels at the source, which is where the clone channel has been since the first upload.
Use a detector the way you’d use a smoke alarm, as a reason to walk over and look. Then publish work a person would be proud to sign, say what the machine did, and let readers be the judge.
That’s the whole system I teach inside AI Labs. Come learn it.
