Debunked Fact Technology

AI Detectors Reliably Catch ChatGPT-Written Essays

They flag false positives at around 20 percent and have a particular bias against non-native English writers

Universities and high schools have AI detectors now. GPTZero, Turnitin's AI module, Originality.ai. Paste an essay in and they tell you whether ChatGPT wrote it. Some students have been failed for cheating on the basis of the score. Some students have been told their entire submission was AI when they wrote every word themselves. The detectors are getting better. The arms race is real. The technology is catching up.

AI text detectors are genuinely terrible at the job they are sold for. A 2023 Stanford study showed that all major detectors flagged 50-70 percent of essays written by non-native English speakers as AI-generated, while missing easily-edited GPT outputs. A 2024 University of Maryland review found false positive rates of 20 percent or more across detectors at standard sensitivity settings, meaning one in five honest students gets accused of cheating. OpenAI quietly retired its own detector in mid-2023, citing 'low rate of accuracy.' Turnitin's AI module has been the subject of multiple academic-integrity complaints from students with documented evidence they wrote their own work. The fundamental problem is statistical: an LLM produces text that looks like fluent English, and so does a fluent human, and there is no robust signal to separate the two without watermarking the AI output at the source. Students have been failed on the basis of these tools. Some have sued. The tools should not be used as evidence. Many institutions have started saying so out loud.

Believed 2022–2025
Year Revised 2023
Why Changed New Evidence
Confidence Fully Debunked
Region Worldwide

Reception

8/10
8/10

Sources

Start typing to search 553 wrong facts