Study: AI raises homework scores 18%, then sinks exams 20%

A working paper landing in the ed-tech conversation this week has an uncomfortable finding for anyone selling AI into classrooms. In a study of 26,811 Chinese high school students in grades 7-12, AI adoption lifted homework scores 18% and cut homework time 30%, then within six months monthly exam scores dropped 20%, and college entrance exam scores fell 18-24%, with the worst performance showing up after two years, Fortune reports.

The researchers, from Stockholm University and the University of Hong Kong and published through the Centre for Economic Policy Research, attributed roughly 80% of the score decline to students who 'outsourced' their homework rather than working through it. Their line is the one worth pinning above the whiteboard: 'For students, completing these tasks efficiently is not the goal; learning from them is.' The gap only shows up when the AI is taken away and the student has to sit an exam alone.

Why this matters for anyone buying, building, or regulating ed-tech: the metric most product dashboards optimize for, task completion and homework grade, moved in the direction vendors promise. The metric that actually matters for students, exam performance, moved the other way. If the effect generalizes, districts piloting AI tutors on completion data are measuring the wrong thing, and the damage does not surface for a year or two.

Take the specifics as reported, not settled. This is one working paper, from one country's school system, and the reporting does not spell out which AI tools the students used or whether guardrailed 'tutor mode' products behave differently from open chatbots. Neuroscientist Jared Cooney Horvath, quoted in the piece, frames the mechanism bluntly: 'When you use offloading tools...as a novice, as a student, you don't learn the skill. You simply learn dependency.' Jacob Shelley, an associate professor of health law at Western University, describes the results he is seeing in his own classroom as 'anomalous. That just never happened in 20 years of teaching.'

The forward-looking read is that a market opens for products and assessments that can show exam-outcome deltas rather than homework speedups, and for teacher-training pitches that treat AI more like a calculator than a tutor. The buyers most exposed are the districts and platforms that signed multi-year AI tutoring contracts on completion metrics alone.