Karpathy: AI Homework Detectors Are Doomed

Former OpenAI researcher Andrej Karpathy argues that efforts to detect AI‑generated homework are futile, calling current detectors “doomed to fail,” saying they “don’t work” and that educators will “never be able to detect” all AI use. He urges a rethink of grading and testing in schools now that powerful models are widely available to students.

Karpathy points to demos like Google’s Nano Banana Pro — which can solve exam problems while even mimicking students’ handwriting — to illustrate how easily AI can bypass traditional checks. His proposal: shift high-stakes assessment back into supervised, in‑school settings and treat AI as a permitted learning companion outside class. The educational aim, he says, should be for students to be “proficient in the use of AI” while still able to “exist without it.”

Why it matters: AI capabilities have outpaced school policies and detection tools. Karpathy’s position reframes the issue from policing (cat‑and‑mouse detection) to redesigning curricula, assessment formats, and learning goals so students learn to use AI responsibly and demonstrate skills under supervision.