When Lauren Jager, a chemistry undergraduate at Idaho State University, submitted her personal statement for a scholarship, it was written entirely by her. She ran it through an online AI detector anyway, out of curiosity. It came back at nearly 100% AI-generated.
“I started freaking out,” she told Nature.
Jager’s experience is not unusual. The US Declaration of Independence, written by Thomas Jefferson in 1776, also scores 95% to 100% AI-generated when run through ZeroGPT, one of the most popular free detectors. As a Nature investigation published July 6 documents, the tools universities have bet on to catch AI cheating are deeply unreliable, particularly against the writing of students who do not speak English as a first language.
The 61% problem
The most cited study on the issue, led by Stanford researcher Weixin Liang and published in Patterns in 2023, tested seven widely used GPT detectors against 91 TOEFL essays written by Chinese learners of English and 88 essays by US-born eighth graders. The detectors flagged 61.3% of the non-native English writing as AI-generated. By contrast, the false-positive rate for native English writing was 5.1%.
The reason is counterintuitive: non-native English writers use simpler, more predictable language, the same low-perplexity signal that AI detectors use to identify machine-generated text. Running the same TOEFL essays through ChatGPT with an instruction to “enhance word choices to sound more like a native speaker” dropped the false-positive rate from 61.3% to 11.8%.
The bias has been replicated at least four times across independent studies since 2023, with the most recent replication in May 2026 showing GPTZero still producing a 16% false-positive rate on informal ESL writing.
A tool that punishes honest students
The unreliability of AI detectors extends beyond the language-bias problem. Different detectors produce wildly different results for the same text. Short, technical, or formulaic writing, in fields like mathematics, chemistry, and engineering, is routinely flagged as AI-generated. And students who use grammar assistants like Grammarly or QuillBot are more likely to be falsely accused.
The consequences can be severe. Students have been reported for academic misconduct, denied scholarships, or forced to defend their work against opaque accusations. Mike Perkins, a researcher at British University Vietnam in Hanoi who studies AI detection, told Nature: “The short answer is no, they don’t work reliably. The long answer is yes, they can work, but the fact that there are so many concerns about false positives means they shouldn’t really be used when it comes to anything that’s sensitive for a student.”
Some students have responded by deliberately “dumbing down” their writing to avoid triggering detectors. “Less perfect” writing, Jager found, dropped her AI score from 100% to 30%.
An escalating arms race
The detection problem is compounded by the rise of “AI humanizer” tools, services that rewrite AI-generated text to evade detection. Turnitin estimates that these tools now attract 33.9 million monthly visits across more than 150 services. Each improvement in detection leads to better evasion, and vice versa, in a cycle that has no clear endpoint.
“People see a score and trust it,” Perkins said. “Similarity tools worked because they could show you exactly where text matched something else. With AI-detection tools, that evidence just isn’t there.”
Universities push back
At least 25 universities have restricted or disabled AI detection tools, including Vanderbilt, the University of Waterloo, Curtin University in Australia, and several UC campuses. The Australian Catholic University abandoned Turnitin’s AI detection entirely after it generated approximately 6,000 false allegations.
The emerging consensus among educators and researchers is that AI detection scores should not be used as the sole or primary basis for academic integrity decisions. Instead, many propose assessing the writing process, drafts, outlines, oral defenses, rather than relying on the final written product alone.
“There is no AI detector that can ever truly be 100% perfect,” GPTZero acknowledges on its website. “Results should not be used to punish or as the final verdict.”
Turnitin, which serves 16,000 institutions, maintains that its tool has a document-level false-positive rate below 1%, though independent studies have found rates of 4% to 12% on edge-case texts.
The Nature article notes that the fundamental issue may be unsolvable: as large language models improve, AI-generated text becomes increasingly indistinguishable from human writing. “We certainly cannot mass-reject people because of it,” said Marzena Karpinska of Simon Fraser University. For now, the most honest answer universities may have to accept is that the detection problem has no technological fix.
Sources
- McKie, A. “Universities are relying on AI-detection software to catch cheating. How well do the programs work?” Nature 655, 535-537 (2026). DOI: 10.1038/d41586-026-01358-2. https://www.nature.com/articles/d41586-026-01358-2
- Liang, W., et al. “GPT detectors are biased against non-native English writers.” Patterns 4(7), 100779 (2023). DOI: 10.1016/j.patter.2023.100779. https://doi.org/10.1016/j.patter.2023.100779
- Dik, A., et al. “Assessing GPTZero’s Accuracy.” arXiv:2506.23517 (2025). https://arxiv.org/abs/2506.23517