also searched as: are ai detectors accurate · how accurate are ai detectors · are ai checkers accurate · are ai detectors reliable · do ai detectors actually work · do ai detectors work · can ai detectors be wrong
Partly: AI detectors catch most unedited chatbot text, but they also flag real human writing, most often by non-native English speakers. Light paraphrasing gets AI text past most of them. Treat a score as a reason to look closer, never as proof.
Trust it for a first read of long text you suspect came straight from a chatbot.
Double-check it when the writer's English is a second language, the text is short or formulaic, or two detectors disagree.
Don't rely on it if the score would be the only evidence behind a failing grade, a disciplinary case or a firing.
sources checked 2026-10-13 · by dan scutaru
| who tested it | what they found |
|---|---|
| Stanford, Patterns journal, 2023 | 61.3% of non-native essays flagged Seven detectors, 91 TOEFL essays by non-native writers and 88 US eighth-grade essays; the eighth-grade essays were classified almost perfectly. source |
| OpenAI, on its own classifier, 2023 | caught 26%, then withdrawn On OpenAI's own test set it caught 26% of AI text and flagged human text 9% of the time; pulled on 20 July 2023 for low accuracy. source |
| Weber-Wulff and 7 co-authors, 2023 | 74% on raw AI, 42% once tweaked 14 tools on human, ChatGPT and edited ChatGPT text; 96% accuracy on the human text. source |
| Perkins and colleagues, 2024 | 39.5% accuracy on AI text Six detectors, 805 samples from GPT-4, Claude 2 and Bard; accuracy fell to 17.4% after simple evasion edits. source |
| Bloomberg Businessweek, October 2024 | 1% to 2% of human essays flagged GPTZero and Copyleaks on 500 Texas A&M essays from summer 2022, before ChatGPT; some flags claimed near 100% certainty. source |
| University of Chicago, 2025 | near zero to 78% wrong, by tool About 2,000 human passages in six genres; Pangram near zero, GPTZero missed about half of humanized AI text, open-source RoBERTa flagged 30% to 78% of human text. source |
| RAID benchmark, ACL 2024 | fooled by simple attacks 12 detectors, over 6 million texts from 11 models and 11 attacks; current detectors are easily fooled by attacks and unseen models. source |
Right most of the time, which is the problem. What counts is how often a detector flags a human, and independent tests put that anywhere from near zero to more than half, depending on the tool and the writer. Turnitin says fewer than 1% of documents are wrongly flagged when 20% or more of the text is AI. Vanderbilt did the arithmetic before switching Turnitin's detector off in August 2023: 1% of its 75,000 papers a year is about 750 students.
Because it is predictable, not because it is AI. Most detectors measure how easily a language model could have guessed each next word, and low surprise reads as machine-made. The Stanford team traced its result to exactly that: a smaller vocabulary and simpler grammar look like chatbot output. Asking ChatGPT to make the same essays sound more native cut the false-positive rate from 61.3% to 11.6%.
They score style and return a probability, so anything that changes the style changes the score. In the Stanford paper, asking ChatGPT to rewrite its own essays in more literary language pushed detection close to zero. So the people who bother to rewrite their AI essay get through, and the ones left in the net wrote plainly.
On raw chatbot text, mostly yes. On edited, paraphrased, short or second-language writing, much less well.
In a 2025 University of Chicago study of four detectors, Pangram. That is one study on the authors' own texts, and even a 0.5% false-positive rate wrongly flags one essay in 200.
Yes, and the vendors say so. Turnitin's own guidance is to use its score to start a conversation, not to draw a conclusion. If one named tool flagged you, we have pages on whether GPTZero, ZeroGPT, Turnitin and Grammarly's AI detector are accurate.