AI Detectors Dropped as Colleges Shift to In-Person Assessment

Colleges are giving up on detection

For three academic years, colleges answered generative AI mostly with software that claimed it could tell a student's writing apart from a chatbot's. That approach is now being abandoned on a large scale. Yale, Vanderbilt, Johns Hopkins and Indiana universities have adopted academic integrity policies that ban or discourage faculty from using AI-detection output as the only evidence of cheating. At least a dozen institutions, including Northwestern, Georgetown and New York University, have switched off Turnitin's AI detection feature.21 The Faculty AI Playbook at Indiana's Kelley School of Business goes furthest. It bans detectors outright and tells instructors to design assignments around process and reasoning instead of trying to catch AI use.21

The US is not alone. In Singapore, the Singapore University of Social Sciences turned off its campus-wide Turnitin AI detector on August 11, and Nanyang Technological University has told faculty it will stop using its detector from 2027.23 In Australia, the New South Wales Education Standards Authority changed its rules in September so that schools cannot rely on detection software when deciding whether a student cheated.23 In South Africa, the University of Cape Town disabled its detector in October 2025, Stellenbosch University ended Turnitin's AI text detection at the end of 2025, and the University of the Free State stopped using detection software on July 1.14 In the UK, universities are restricting or dropping tools such as Turnitin, GPTZero and Copyleaks over false positives and bias.11

The trend reached the most prominent American campuses this fall. In an email to students, Harvard College Dean David Deming said everyone would benefit from "getting out of the AI-detection business," because it damages trust between faculty and students.25 Cornell's guidance for instructors says detection technology is unlikely to offer a workable solution.13

Why the tools lost the argument

The accounts largely agree on the reasons. Studies have repeatedly found that detectors give inconsistent results, flag human writing as AI-generated, and are biased against writers whose first language is not English. Some students who were wrongly flagged have sued their universities.21 The best-known evidence comes from a 2023 study in which seven detectors labelled TOEFL essays by non-native English speakers as AI-written at an average rate of about 61%, while scoring US eighth-graders' essays almost perfectly.23 The same research produced an awkward follow-up result: when the non-native essays were polished with a chatbot, the false-positive rate fell from about 61% to under 12%.13 In other words, using AI made the writing look more human to the software.

The tools also miss a lot of real AI text. One 2023 test of seven popular detectors found average accuracy of 39.5% on unedited AI text, falling to about 22% once the text had been reworked to disguise it.23 A 2026 study in the International Journal for Educational Integrity put overall accuracy at 61–69%, with performance on mixed human-AI writing close to zero.26 Yale's teaching center has said that "humanizer" programs are very good at lowering or removing the share of AI text that detectors catch.21

Faculty seem to have reached the same view. In a survey of 1,057 US faculty by the American Association of Colleges and Universities and Elon University, only 31% said they use detectors, although 34% said their institution pays for one. Only 23% rated the tools effective, and 55% said they were not very effective or not effective at all.13 The survey's authors caution that the sample is not scientific.13 The Washington State University figures are concrete: a third of its 2023–2025 AI misconduct cases that rested only on detector evidence ended in "not responsible" findings.16

Not everyone agrees, and the technology keeps changing

The shift away from detection is real, but it is not unanimous. A September 2025 working paper by Brian Jabarian and Alex Imas found that one detector, Pangram, had false-positive and false-negative rates close to zero, even on disguised text.23 On September 29, Dartmouth's interim dean of faculty let professors use Pangram, and only Pangram, provided they tell students in advance and talk with a student before a score affects a grade.23 In South Africa, North-West University still says its tool is accurate enough to be a useful first indicator, and the University of Johannesburg uses a hybrid model.14 Newer research tools such as IdeaLens try to trace whether AI supplied a document's ideas rather than its wording. Their developers report catching 91% of fully AI-written work with a 0.6% false-positive rate, though they say more testing is needed.28

The practical question, then, is less whether detection technology can ever work. It is whether any score can carry the weight of a misconduct finding. Most institutions now treat a detector flag as a reason to talk to the student, not as proof, and Dartmouth's conditions follow that model too.23 Counts of how many universities have formally dropped detectors vary widely, from "at least a dozen" US institutions in Inside Higher Ed's tally to claims of more than 50 worldwide from less rigorous outlets.2119 The direction is clear even if the exact number is not.

What replaces detection costs more

What comes next is clear too, and it will be expensive. Instructors are pairing take-home essays with oral exams, grading drafts and notes alongside final work, moving writing into class, and bringing back proctored exams.25 Deming has argued that take-home exams and problem sets can no longer reliably show mastery.25 The University of Sydney's "two-lane" model is the clearest version: secure, in-person tasks test what students can do unaided, and in open tasks any relevant tool is allowed.23 Tricia Bertram Gallant of UC San Diego summed up the problem behind all of this: higher education still treats unsupervised writing as evidence of learning, and that no longer works.21

Kevin Yee of the University of Central Florida points out a scale problem. Instructors with fewer than 30 students may be able to keep cheating down through personal relationships. Large lecture courses, he argues, need assessments designed to withstand AI, such as blue-book exams or in-person oral exams.21 By his estimate, about half of faculty want to change as little as possible.21 Several observers make the same warning: these approaches take more staff time and are harder to scale than a software check.14 One analysis adds that oral and in-person formats bring fairness risks of their own, so institutions should track how they affect second-language students as closely as they once examined detectors.23 That point deserves attention. Ending a biased tool does not guarantee that what replaces it will be fair.

Why the pressure is growing

Usage figures explain the urgency. Pew found that 54% of US teens aged 13 to 17 have used chatbots for schoolwork, up from 26% who said they used ChatGPT for school a year earlier, and 10% say they do all or most of their schoolwork with AI help.1 Common Sense Media's August survey put teen schoolwork use at 70%.5 The two surveys differ, but both point to majority use. In Pew's survey, 59% of teens said AI cheating happens regularly at their school.7 Among UK undergraduates, 94% use generative AI for assessed work, and the share who paste AI-generated text directly into assessed work rose from 3% in 2024 to 12% in 2026.10 An Edinburgh Napier survey of more than 6,600 students found that 32% admitted some AI use that was not permitted.11

Policy has not kept up, especially in K-12. Only 44% of teens in the Common Sense survey said they had received guidance on when AI is appropriate for coursework.8 Many districts block public chatbots and allow approved products such as Khanmigo instead.2 Some universities are going the other way: UC Berkeley School of Law has reportedly banned AI in coursework from fall 2026.19

The tutoring research complicates the case for AI

The argument for redesigning assessment rather than banning AI often rests on the idea that AI can help students learn. The research on that is mixed. In a Harvard randomized trial, physics students using a purpose-built AI tutor learned more than twice as much, in less time, as students in an in-class active-learning session.36 Stanford researchers have described trials in which Google's LearnLM matched human tutors at fixing mistakes, and in which the Tutor CoPilot tool raised topic mastery by four percentage points, with the largest gains for students of less experienced tutors.31

A new randomized trial with 2,379 undergraduates at a large US public university points the other way. Students given access to a course-integrated AI tutor earned final grades up to 0.37 standard deviations lower, and their participation in course activities fell sharply.34 First-generation students lost more than twice as much as their peers.34 The authors cite other work showing that general-purpose AI can raise performance on assisted tasks while hurting later independent work, and that design guardrails can reduce that harm.34

Taken together, the studies suggest that results depend on how a tool is built and how students use it, not on AI in general. That is the reasoning behind the move away from detectors. The tools can no longer show who did the work, and AI help is not guaranteed to produce learning. The only dependable evidence of what a student knows comes from tasks they complete in person or in conversation with an instructor, and the institutions that will cope best are the ones willing to pay for the staff time those tasks require.