AI study tools are tanking student test scores

Giving a student an AI tutor sounds like an instant academic upgrade, but the cold hard data tells a very different story. Kids leaning on AI for schoolwork are actually doing worse.

A new OECD PISA study, drawing on data from over 760,000 15-year-old students across 91 countries, tracked academic performance in science, math, and reading. Crucially, this is the first PISA assessment carried out since generative AI went completely mainstream.

Even after controlling for socioeconomic background, students who never use AI consistently outperformed AI users in science.

Outsourcing the struggle

The gap largely comes down to what students actually ask the machine to do.

Students who used AI for shortcut tasks, like summarizing texts they were supposed to read or drafting essays, saw the steepest drop in scores. Those using it for exploratory research or general conceptual help took far less of a hit.

Frequency shows an odd curve too. Students who use AI daily and those who only use it once or twice a year performed the worst. The sweet spot was moderate, intentional use, roughly once a week or a few times a month.

OECD director of learning Andreas Schleicher put it bluntly: you do not get fit by watching sports. Real learning requires a productive cognitive struggle with difficult material. When an AI tool short-circuits that mental effort, it quietly sabotages the student's own intellectual muscle.

Adoption also varies wildly by region. Over 95 percent of students in Vietnam report using AI tools, compared to just 60 percent in Japan. The study also noted that daily AI users showed the highest levels of natural curiosity, but that spark is not translating to higher grades.

The critical thinking dividend

AI is not an automatic downgrade, provided students are taught to treat it with suspicion.

Students who used AI weekly for general learning saw a slight edge over nonusers. That advantage widened significantly when students were trained to actively critique and verify the quality of AI-generated answers. With that critical training in place, regular users comfortably beat nonusers.

Why it matters

This study punctures the lazy narrative that simply putting generative AI in classrooms will automatically elevate learning. In practice, AI often acts as a cognitive crutch that eliminates the exact friction required to retain knowledge.

For students and parents, the takeaway is straightforward: use AI for preliminary brainstorming and questioning, but never for reading and writing on your behalf. Skipping the messy middle of synthesis is directly linked to worse retention and lower scores.

For edtech developers, this should be a wake-up call. Building tools that spit out polished answers is actively harming users. The real opportunity lies in software that challenges students, introduces deliberate friction, and forces critical evaluation rather than passive consumption.