AI-writing detectors are becoming gatekeepers in classrooms, newsrooms and literary prizes, but the evidence says they cannot reliably identify machine authorship. Style-based systems measure how predictable a text looks—a proxy that misfires on plain prose, translation and second-language writing. The result is a false-positive problem with real victims.
- False positives are systemic: Seven popular detectors flagged an average of 61.22% of TOEFL essays by non-native English speakers as AI-generated in a Stanford University study.
- Commercial tools fall short: A test of 14 detectors, including Turnitin and PlagiarismCheck, found none above 80% accuracy; only five exceeded 70%.
- Watermarks are partial: Google DeepMind’s SynthID Text can mark output, but detection depends on the generator and its configuration. EU labeling rules start in 2026.
Why “AI-like” often means “non-native”
The Stanford experiment compared 91 TOEFL essays from a Chinese education forum with 88 essays by US eighth graders. Seven detectors agreed on labeling 18 of the 91 non-native essays as AI-written (19.78%), and at least one flagged 89 of 91 (97.80%). Native-speaker essays were mostly classified correctly.
The inverse test exposed the mechanism. When researchers asked ChatGPT to enrich the TOEFL essays’ vocabulary, the average false-positive rate fell to 11.77%. When they simplified native essays to sound non-native, it jumped from 5.19% to 56.65%. Limited vocabulary and simple sentences look “too predictable” to the algorithm—yet they are also normal features of second-language writing.
Commercial detectors and the literary fallout
A separate study of 14 tools—including Turnitin and PlagiarismCheck—found none exceeded 80% accuracy, and only five topped 70%. About 20% of AI-generated texts were attributed to humans, rising to roughly 50% when hand-edited and higher after paraphrasing. False positives ranged from 0% for Turnitin to 50% for GPT Zero. Translation through Google Translate or DeepL also increased false positives.
OpenAI’s own retreat reinforced the pattern. It launched the AI Text Classifier on January 31, 2023, and withdrew it on July 20, 2023, citing low accuracy. On its test set, it caught 26% of AI-written text and wrongly flagged human writing 9% of the time. At launch, OpenAI had warned that reliably detecting all AI-written text was “impossible.”
Publishing is already feeling the consequences. After Simbarashe Kundizeza won the Caine Prize for African Writing for “The Blue Fighter,” a detector flagged almost the entire story as AI-generated. He denied using generative AI.
“For the avoidance of doubt, I wrote ‘The Blue Fighter.’ I did not use generative AI to write it. It is entirely my own work,” Kundizeza wrote in a statement.
Several universities—including Vanderbilt, Georgetown and Waterloo—have disabled or refused to treat Turnitin’s AI indicator as proof, citing opaque scoring, false positives and bias against second-language writers.
Watermarks, regulation and the burden of proof
The most technically grounded alternative is watermarking. SynthID Text, from Google DeepMind, makes pseudo-random adjustments to token probabilities during generation without significantly affecting quality. Google released it as open source, and Hugging Face’s Transformers library has had a production-ready implementation since version 4.46.0. A Nature paper published on October 23, 2024, described tests on nearly 20 million Gemini responses with no appreciable quality loss.
But watermarking is not a universal scanner. It detects only text from models that apply the mark. Confidence drops after heavy rewriting or translation, and performance is weaker on factual responses. Detection is probabilistic—marked, unmarked or uncertain—and tied to a specific watermark configuration; models sharing a tokenizer may share a mark, but there is no cross-model standard.
Regulation will help, slowly. Article 50 of the EU AI Act requires providers, from August 2, 2026, to mark synthetic text, image, audio and video in machine-readable form. Systems already on the market have until December 2, 2026. The rules do not reach models running outside that jurisdiction.
For students and teachers, the answer is procedural. A flagged student should ask in writing which tool was used and on what basis, then supply dated drafts, notes and sources. Educators should treat a detector result as a question, not a verdict: discuss the work, ask how conclusions were reached, and assess the process alongside the product. The Stanford findings and translation evidence are serious counterarguments to automatic penalties.
The deeper problem is structural. A style classifier can only measure features that many humans also share. Until a universal, robust watermark exists—and it does not—the burden of proof cannot rest on an algorithm that confuses simple writing with machine writing. Institutions need fewer automated accusations and more visible processes for authorship.