AINEWS
Stanford found the most popular AI detectors flag 61% of non native English essays as machine generated, yet universities still treat their scores as evidence of cheating.
The U.S. Declaration of Independence is, by detector verdict, a chatbot.
When the 1776 text is run through a popular AI-content detector, the tool returns a confidence score of 95 to 100 percent that a large language model wrote it. The same class of tool, according to Stanford's Institute for Human-Centered AI, classified 61.22 percent of TOEFL essays written by non-native English students as AI-generated, while scoring U.S.-born eighth-graders near-perfectly. Turnitin's plagiarism product, the de facto standard on most U.S. campuses, cannot catch freshly assembled machine text at all because it works by matching submissions against a known corpus. So the university is left, in practice, with a tool that fails the Declaration of Independence and, in independent testing, flags two out of three non-native English speakers, and is treated as the final word on who is honest.
Lauren Jaeger, a chemistry major at Idaho State University applying to PhD programs, ran her personal statement through every detector she could find. Every one of them returned a "nearly 100 percent AI" verdict. She had written it herself.
In reporting for oztalking.com, Jaeger described what she did next. She deliberately rewrote the statement worse, replacing precise vocabulary with simpler words, until the detector scored it 30 percent. She submitted it thinking, "this should do." She was admitted to a PhD program at the University of Utah.
The cost of the current system is not theoretical. It is a graduate applicant learning to write badly so that an opaque classifier will let her through.
Cath Ellis, the academic-integrity lead at Western Sydney University, framed the moment more precisely. The post-ChatGPT era is, in her words, "a fundamental shift in scale," because the volume of at-least-substantially-fabricated submissions has exploded. The integrity problem universities are responding to is real. The chosen tool, on the evidence available, is not yet up to it.
Liang and colleagues at Stanford showed in 2023 that GPT detectors consistently misclassify non-native English writing as AI-generated, and that the same simple prompting tricks that can fool the detector can also be inverted to reduce its bias. The follow-on question was whether length or use case would change the picture. A 2025 study by Dik and co-authors tested GPTZero across short (40 to 100 words), medium (100 to 350), and long (350 to 800) essays. Most AI-generated papers were caught. The unresolved concern is the human side: detectors still do not reliably prove a human wrote the text, and the false-positive rate is the lane the academic-integrity office lives in.
Detector vendors dispute the framing. GPTZero, Turnitin, and others publicly push back on the false-positive thesis, arguing that their products are decision-support, not decision-making, and that thresholds and human review sit above the score. Stanford's evidence and Jaeger's anecdote cut the other way. A score that mislabels the Declaration of Independence and 61 percent of non-native English prose is not a tool the human reviewer can use as a sanity check. It is a tool the human reviewer has to second-guess every time, on every essay, with no ground truth to consult.
That is the institutional trust question. Universities have licensed their honesty function to a classifier that cannot pass a 250-year-old document, and they are scaling it across the most consequential decisions in a young scholar's life. The next admissions cycle is the operational deadline. Either the systems grow a ground-truth layer, a calibrated uncertainty output, and an appeal path that does not require the applicant to write worse, or another cohort of Jaegers will quietly lower their writing to make the math come out right.