Is there a tool that can detect ChatGPT-written text?
Dozens exist — GPTZero, Turnitin, Copyleaks, Originality.ai — but none is reliable enough to accuse someone. OpenAI killed its own detector on July 20, 2023 for low accuracy: it caught just 26% of AI text while falsely flagging 9% of human writing. Peer-reviewed testing of 14 tools called them "neither accurate nor reliable."
Why — the first-principles explanation
The tools are real. The question is what a result from one is worth, and the answer follows from what the detector can actually see. It doesn't see history. It never watched you type. All it gets is finished text, and it measures statistical properties of that text — mainly how predictable each word is given the words before it, plus how much that predictability varies from sentence to sentence. Language models pick high-probability words, so their output tends to sit in a smooth, low-surprise band. Humans wander.
That's the entire signal, and it has three structural problems that no amount of engineering removes.
Problem one: plenty of humans write predictably. A careful, plain, well-organized writer — exactly what students are trained to be — produces low-surprise text. So does anyone writing in a second language, drawing on a smaller vocabulary and safer sentence patterns. This isn't hypothetical. Liang et al. (Patterns, 2023) ran seven detectors over 91 TOEFL essays written by non-native English speakers: 61.3% were flagged as AI-generated, 97% were flagged by at least one detector, and all seven wrongly agreed on 19% of them. The same tools were near-perfect on essays by U.S.-born eighth-graders. The tool isn't detecting AI. It's detecting predictable prose — and it punishes the people least able to defend themselves.
Problem two: the signal is trivially destroyed. Reword a few sentences, run it through a paraphraser, or just ask the model for more unusual phrasing, and the statistical fingerprint disappears. Weber-Wulff et al. (International Journal for Educational Integrity, 2023) tested 14 tools — 12 public plus Turnitin and PlagiarismCheck — and found that content obfuscation techniques significantly worsened performance. So the tool fails hardest on the people who are actually trying to cheat, and works best on people who aren't.
Problem three: base rates. Even a genuinely good detector produces mostly-wrong accusations when cheating is uncommon. If 5% of a class used AI and your detector is 98% accurate, then out of 1,000 papers you catch about 49 real cases and wrongly flag about 19 honest students — roughly one in four accusations is wrong, with the tool working exactly as advertised.
OpenAI, which knows more than anyone about how its own model writes, ran into this and quit: it withdrew its AI Text Classifier on July 20, 2023 "due to its low rate of accuracy." That's the single most informative fact in this whole debate.
An example that makes it click
Imagine a tool that identifies professional photographs by measuring how level the horizon is.
It works, sort of. Pros use tripods, so their horizons are level. But it will flag every careful amateur who braced against a fence — and it will clear every pro who tilts the camera five degrees on purpose. You'd never let that tool get someone fired for lying about their credentials, because it isn't detecting professionalism. It's detecting levelness, which merely correlates with professionalism until someone has a reason to break the correlation.
AI detectors measure predictability the same way. Predictable text correlates with AI — right up until a careful human writes plainly, or a cheater adds one round of rewording. The correlation snaps precisely when there's a motive to snap it.
How to do it
- Understand what you're getting: a probability score derived from how predictable the text is — not evidence of authorship.
- If you're an educator, do not treat a score as proof. Every serious source, including the peer-reviewed testing, says these tools cannot substitute for a finding of misconduct.
- Look for real evidence instead: document version history in Google Docs or Word, draft timestamps, and whether the writing matches the student's known level.
- Ask the writer to discuss the work. Someone who wrote a paper can explain their choices; someone who generated it usually cannot.
- If you're accused wrongly, ask which tool produced the score and what its stated false-positive rate is, then present your version history and drafts. Cite the OpenAI withdrawal and the Liang and Weber-Wulff studies — an accuracy claim without a false-positive rate is not a claim at all.
- If you're writing and worried about false flags, keep your version history switched on before you need it. It's the only artifact that proves process rather than style.
Key facts
- OpenAI discontinued its AI Text Classifier on July 20, 2023, stating it was no longer available "due to its low rate of accuracy."
- OpenAI's own classifier correctly identified only 26% of AI-written text as "likely AI-written" (true positives) while incorrectly labeling human-written text as AI-written 9% of the time.
- Weber-Wulff et al. (International Journal for Educational Integrity, vol. 19, art. 26, 2023) tested 14 detection tools — 12 public plus Turnitin and PlagiarismCheck — and concluded they "are neither accurate nor reliable," with a main bias toward classifying output as human-written.
- The same study found content obfuscation techniques significantly worsen tool performance — meaning detection fails most against deliberate evasion.
- Liang et al. (Patterns, vol. 4 issue 7, 2023) found seven GPT detectors flagged 61.3% of 91 TOEFL essays by non-native English speakers as AI-generated; 97% were flagged by at least one detector and all seven agreed on 19%.
- The same seven detectors were near-perfect on essays written by U.S.-born eighth-graders, showing the error is concentrated by population rather than spread evenly.
OpenAI's conversational assistant — the most-used AI chatbot in the world.
▶ The 60-second explainer (script)
Is there a tool that can detect ChatGPT writing? Dozens. GPTZero, Turnitin, Copyleaks, Originality. The real question is whether a result from one means anything. Start with the most telling fact. OpenAI built a detector for its own model — nobody knows better how GPT writes — and killed it in July 2023, quote, due to its low rate of accuracy. It caught twenty-six percent of AI text and falsely flagged nine percent of human writing. Here's why. A detector never watched you type. All it sees is finished text, and it measures one thing: how predictable each word is. AI picks likely words, so its output sits in a smooth, low-surprise band. Humans wander. That's the whole signal — and it breaks three ways. One: lots of humans write predictably. Stanford researchers ran seven detectors over ninety-one TOEFL essays by non-native English speakers. Sixty-one percent were flagged as AI. The same tools were near-perfect on American eighth graders. It's not detecting AI. It's detecting plain prose. Two: the signal is trivially destroyed. Peer-reviewed testing of fourteen tools found obfuscation significantly worsens performance — so it fails hardest on people actually cheating. Three: base rates. If five percent of a class cheats and your detector is ninety-eight percent accurate, about one in four of the students it flags is innocent. That's the tool working perfectly. So: the tools exist. They are not evidence. Use version history instead.
What authoritative sources say
People also ask
Which detector is the most accurate?
There's no independent, current, peer-reviewed ranking that holds up on real edited text. Published comparisons mostly test clean, unmodified model output — the case that almost never occurs in practice.
Vendors advertise 99% accuracy. Are they lying?
Usually not — they're reporting a real measurement on a favorable test set of unedited AI versus clean human text. The number is honest; the framing hides the false-positive rate and who bears it.
Can a school punish a student based on a detector score?
Policies vary, but the peer-reviewed evidence doesn't support it. Turnitin itself positions its score as an indicator requiring human review, not proof of misconduct.
Does rewording AI text beat the detectors?
Largely yes, which is the core problem. Weber-Wulff et al. found obfuscation significantly degrades performance, so detectors work worst against deliberate cheating and best against honest plain writers.
Why not just watermark AI output?
It's technically possible and OpenAI said it was researching better provenance techniques after retiring its classifier. But watermarks only cover participating vendors, and survive neither paraphrasing nor open-source models.