AI detection no longer works: teachers are redesigning exams to demand the process, not the result

By Zendoric · September 8, 2026.

Rachael Zeleny, head of the writing program at the University of Baltimore, had long been recommending Brisk, a tool that logs every revision, edit and pasted fragment in a Google Docs document in order to reconstruct how an essay was written. As The Washington Post reported, she ended up abandoning that approach: she told her own teaching staff last year, "this is a Band-Aid, not a solution." In its place, she introduced checkpoints throughout the process: handwritten outlines, recordings in which the student explains their decisions, and transcripts of the conversation with the chatbot they used.

Zeleny has gone further and brought AI directly into the classroom instead of trying to shut it out: with BoodleBox she created artificial intelligence characters that argue opposing positions in a simulated debate about book banning on a fictional school board, and the student has to respond to their arguments. The point is not to find out whether they used AI, but to check whether they can weigh information and sustain an argument. Andrew Zeiser, a professor of supply chain management at John Carroll University, describes the same shift with a line that sums up the underlying change: "the big worry is no longer whether they're using AI, it's whether they're still learning." His students make decisions in a business simulation, Supply Chain Wars, and afterwards have to justify the reasoning behind each choice.

The context that explains this rush to redesign comes with numbers: according to a 2025 survey by the Walton Family Foundation and Gallup cited by The Washington Post in September, around 60% of teachers already use AI to grade, prepare classes or communicate with families, and a Common Sense Media study puts at close to 70% the share of teenagers who say they use it, with schoolwork among their main uses. AI is not an anomaly to be fought in the classroom: it is already everyday infrastructure on both sides of the desk.

The problem, warns Anna Mills, an English instructor at the College of Marin, is that the next wave of AI makes even this redesign irrelevant. AI agents no longer merely generate text: they can operate directly on online learning platforms and complete exams or assignments from the student's own account, leaving no text to analyze. A plagiarism detector, however reliable, has nothing to detect if the assignment solved itself inside the system. Meanwhile, companies such as OpenAI and Anthropic are offering free access to teachers and school districts, and Google has expanded Gemini within Google Classroom, while New York and the Los Angeles school district have imposed temporary restrictions on generative AI: the sector is advancing without a common policy, caught between the tech companies' active promotion and the authorities' piecemeal brakes.

Our reading is that this episode is a textbook case of something we have already seen in other fields: when AI capability grows faster than the tool designed to police it, the detection approach stops making sense and the entire system has to be redesigned, not patched. In cybersecurity it happens with benchmarks that saturate; in education it happens with plagiarism detectors. The answer these teachers are improvising — assessing the process of thinking rather than the final artifact — is not just a tactical fix: it is a preview of how any institution that certifies knowledge or competence, from a school to a professional exam, will have to adapt as AI agents stop being an occasional aid and start operating on the user's behalf.

This fits a thesis we have been defending about AI's impact on teaching jobs: the winner is not the teacher who polices the most, it is the one who orchestrates AI in the classroom instead of fighting it, as Zeleny does by turning the chatbot into debate material. In the short term the cost is real and should be stated plainly: more workload for teachers, who now have to design checkpoints, recordings and justifications where an essay used to be enough, and an inequality from the outset between schools with the resources to reinvent assessment and those without them. But the underlying direction — shifting value from the result to the reasoning behind it — is exactly the kind of institutional adaptation we need if AI is going to take on more and more mechanical tasks: what is left to assess, and to teach, will be the human judgment that no tool can supply for us.