Higher education is quietly retreating. Institutions that once treated AI detection software as the answer to academic dishonesty are now stepping back from it. According to RTL Today, reporting on AFP, Yale and Cornell have banned or advised faculty against using detection tools as primary evidence of cheating.
The retreat is not an admission that AI use is acceptable. It is an admission that the evidence is not reliable. A detection score is a probability, not a proof. When a student’s degree or admission hinges on that score, the difference matters. Alternative assessment methods, such as oral examinations, have their own limitations, often favoring more extroverted students while disadvantaging those who think more slowly and reflectively.
Markus Goldbach has spent 15 years in academic integrity technology. He is the CEO of Mentafy, a platform built on a different premise: that the writing process itself is the evidence. Goldbach argues that the industry has been asking the wrong question. He says, “AI detection software only evaluates silent grammar patterns and word frequencies. It produces a probability, a likelihood. There is no underlying proof whatsoever.”
The technical limits are not a matter of opinion. TBS Graduates, citing AFP, reported that the University of Wisconsin-Madison warns in its faculty guidelines that detection tools are imperfect at best, carry the risk of false positives, have been shown to be biased against non-native English speakers, and will not prevent students from using these tools.
The institutional response is already shifting. As per Inside Higher Ed, a growing number of universities have prohibited AI detectors, arguing they are unreliable indicators of cheating. For example, Yale, Vanderbilt, Johns Hopkins, and Indiana have all enacted policies banning or discouraging reliance on detection tools as sole evidence, and at least a dozen universities have disabled Turnitin’s detection software.
Goldbach says, “Detection tools look for patterns in grammar, word choices, and style that AI tends to produce. But a student who happens to write in a similar style gets flagged, even if they never touched AI.”
An IEEE Xplore benchmarking study published in June 2026 tested five widely used detectors against 180 academic writing samples. It found false positive rates ranging from 15.6% to 30.6%. Accuracy dropped sharply on hybrid texts that mix human and AI writing to between 54.2% and 70.8%. If detection cannot reliably identify the work it is designed to catch, it is not a safeguard. It is a coin toss with consequences.
Goldbach says, “Honest students are getting caught in the middle. They are accused of something they did not do, and they have no way to prove otherwise. The false negative rate is also substantial, and we’re not even accounting for that here.”
The alternative is not more sophisticated detection. It is a shift in where institutions look. Mentafy documents the drafting journey from the first blank page, tracking revisions, edits, and the evolution of ideas. The goal is to make AI use transparent rather than concealed. A student who uses AI for grammar polishing, a practice most universities already permit, leaves a different trail than a student who generates an essay from a prompt. The writing process shows which is which.
Goldbach is not anti-AI. He describes the modern equivalent of asking a parent or peer to proofread a paper, or asking a librarian for a useful source. The problem is that detectors cannot distinguish that permitted use from wholesale generation. When they flag a polished sentence, they catch the student who followed the rules and miss the one who did not. ABC posted an article where hundreds of students at an Australian university were wrongly accused of submitting AI-generated papers. After months of waiting, the institution concluded they were innocent. The technology has not become more reliable since.
The structural solution Goldbach proposes is documentation. Universities should require the writing process to be recorded digitally from day one, not as surveillance but as a record of intellectual work. “The final PDF is not the learning,” he says. “The learning is the revision, the false starts, the moment a student realizes an argument does not hold.” That record is what a detector cannot see and what a teacher needs.
“If you do not have blood, sweat, and tears during your learning process, you will not learn that much. Your brain is a muscle; if you let AI do the heavy lifting, it will never get strong,” Goldbach says. The warning is not rhetorical. Deep academic work builds problem-solving stamina that no detection score can measure.
The long-term consequences of inaction are not abstract. If degrees become rubber-stamped AI prompts, employers will stop valuing them. Goldbach warns that the damage will not necessarily appear in five years but in the near future, when a generation of graduates enters the workforce without the capacity to think independently. For countries without natural resources, like Germany, the loss of innovation capacity is an economic threat, not just an educational one.
The path forward is not a return to handwritten exams. Oral examinations have their place and justification, but also their own bias, favoring extroverts and penalizing students who think slowly and deeply. The answer is to preserve the long-form work while making the process visible. Universities do not need to fear AI. They need to stop expecting an algorithm to solve a problem that requires a change in how they evaluate learning.
The real value of an education was never the final sheet. It was the intellectual work it took to get there. The institutions that understand this will be the ones that emerge from the AI transition with their credibility intact. Those that keep chasing detection scores risk presiding over the decline of an entire generation of graduates.
