As artificial intelligence tools become ubiquitous in classrooms worldwide, a fundamental question emerges: do these technologies enhance learning, or do they replace it? New findings from the OECD's PISA 2025 study suggest that students who use AI tools daily for homework and studying achieve lower average scores than those who use them sparingly or not at all.
The data comes from the Programme for International Student Assessment (PISA), which evaluates the academic performance of 15-year-olds across different countries. The 2025 cycle included more than 760,000 students from 91 nations and represents the first assessment conducted after generative AI tools became widely accessible to the general public.
The most striking gap appeared among students asked about using artificial intelligence to draft written assignments. Those who reported "never or almost never" using AI for this purpose scored an average of 509 on PISA's science assessment. In contrast, students who said they do so daily or nearly daily averaged 481 points.
That 28-point difference persisted even after adjusting for socioeconomic background. According to the study's estimates, the gap is equivalent to approximately one and a half years of instruction.
Similar patterns emerged across other uses: conducting preliminary research for assignments, summarizing assigned reading material, and obtaining ready-made answers rather than engaging independently with content. However, the researchers caution that this represents a statistical correlation rather than definitive proof that AI use directly causes lower scores. It's possible, for instance, that struggling students turn to these tools more frequently in the first place.
Most students are already using AI for schoolwork, with only 14% reporting no use at all. But adoption rates vary dramatically by country: in Japan, nearly 40% of students said they don't use AI for academic work, compared to just 4% in Vietnam.
The OECD warns against uncontrolled integration of these tools in classrooms and emphasizes the need for clear guidelines. In one striking comparison, the organization likened the unregulated rollout of AI in education systems to "giving children an untested vaccine."
Yet the findings present a more nuanced picture than simple prohibition. Students who used AI tools approximately once or twice weekly tended to perform better than both those who used them rarely and those who used them daily.
Among students who reported using AI "to help me learn," weekly use correlated with higher achievement, even compared to students who didn't use the tools at all. In other words, the critical factor isn't whether students use artificial intelligence, but how they use it.
The data also highlights a significant role for teachers. Among students who used AI daily, science scores were 13 points higher when classroom instruction regularly required them to examine, cross-check, and evaluate information produced by AI. That gap represents more than half a year of schooling.
Andreas Schleicher, who directs the OECD's education and skills division, framed it this way: artificial intelligence should serve as "scaffolding, not a crutch." Just as physical fitness cannot be developed by merely watching sports, he noted, learning doesn't occur through content consumption alone, it requires active mental effort and engagement with new material.
The AI findings come against a backdrop of broader declines in student performance across OECD countries. Reading scores dropped 28 points among member nations between 2015 and 2025, while mathematics scores fell 22 points.
The report attributes these declines to deep-seated problems in education systems, including digital distractions and declining interest in reading books. Meanwhile, policy responses to AI vary widely: New York City, for example, has decided to ban artificial intelligence tools in elementary and middle schools for the upcoming academic year.
The message emerging from the data isn't necessarily a call to banish artificial intelligence from schools entirely. Rather, it underscores that frequent, unmonitored use can turn the tool into a shortcut that undermines the practice of thinking, writing, and comprehension. Measured, guided, and critical use, where students are required to verify answers, identify errors, and explain their reasoning process, may transform AI into a genuine learning aid rather than a substitute for learning itself.
