Academic Integrity & Assessment

Assessment is how education certifies what a student has learned, and generative AI has shaken its foundation. When a chatbot can write a passable essay in seconds, the take-home written assignment — a century-old default — no longer reliably measures a student's own work. Schools and universities have scrambled to respond, and the response has divided into two very different approaches.

The first is detection: software that tries to tell whether text was written by AI. The second, and increasingly favored, is redesign: rethinking assessment so that it is resilient to AI in the first place — authentic tasks, supervised or oral components, and assessment of the process rather than only the final product. The honest truth of 2026 is that detection is on shaky ground, while assessment redesign is the harder but more durable path.

A third response sits between those two, and 2026 delivered its most expensive test. Remote proctoring tries to preserve the traditional exam by watching the test-taker — a locked-down browser plus algorithms monitoring the webcam for wandering eyes, extra people in the room, or background conversation. In mid-2026 the National Autonomous University of Mexico, known as UNAM, ran its entrance exam fully online for the first time on exactly that stack. The scores came back unlike any in the university's history: about 16 percent of candidates cleared 100 correct answers out of 120, against roughly 3 percent across the preceding five years. A technical commission concluded the results could not be trusted, and roughly 58,000 applicants — about 36 percent of the field — were called back to sit an in-person control exam under human supervision.

The lesson generalizes past any one vendor. Surveillance-based proctoring watches the room, but the AI a candidate might use does not have to be in the room, and the failure is invisible until aggregate scores make it obvious. Note also what actually caught the problem: not the proctoring software, but statistics — a score distribution that did not match five years of history. That is the practical takeaway for institutions holding the line on traditional assessment. If you cannot detect misuse during the exam, you can still detect it afterwards in the data, and you should be looking.

On the detection side, Turnitin is the dominant academic-integrity platform and GPTZero and Copyleaks are widely used AI-writing detectors — but all three sit in a genuinely contested category. On the redesign side, Packback builds AI-resilient assessment and AI-assisted grading around curiosity and credibility, and Cadmus provides an authentic-writing environment plus AI-supported oral assessment. Gradescope speeds grading of student work at scale. The credible institutional answer increasingly leans toward redesign over detection.

AI is forcing a genuine rethink of how learning is measured, and the honest story here is unusually important. AI-writing detectors are unreliable: independent testing shows meaningful false-positive rates, they disproportionately misflag non-native English writers, and their accuracy collapses on lightly edited text — in 2026 a major university publicly disabled AI detection over these concerns. Treating a detector's output as proof of cheating is dangerous and has harmed innocent students. The more durable response is redesigning assessment to be authentic and AI-resilient, which shifts educators' work from policing toward better task design. The role of human judgment — in grading, in interpreting a detector's signal, and in designing assessment — becomes more central, not less.