KEY POINTS
-
While the use of generative AI is skyrocketing, student performance began to decline before these tools even emerged.
-
Among the highest-performing education systems, approaches to using AI vary widely.
-
The PISA survey encourages us to distinguish between uses that replace cognitive effort and those that can support it.
The PISA 2025 results are cause for concern. Between 2015 and 2025, the average score for OECD countries fell by 22 points in mathematics—the equivalent of just over one year of learning—and by 28 points in reading. Furthermore, 20% of 15-year-olds now fall below the basic proficiency level in all three assessed areas, compared to 16% in 2022.
France is no exception to this trend. Between 2022 and 2025, it lost 18 points in reading and 16 points in mathematics, while science scores remained relatively stable (-4 points). The OECD notes that these results are now among the lowest ever recorded for the country.
When PISA scores drop in this way—and on such a large scale—there is a strong temptation for families, teachers, and society at large to look for someone to blame. In this case, could this decline be due to artificial intelligence?
But before blaming artificial intelligence (AI), we need to look at the timeline: the decline in scores is part of a trend that began before the COVID-19 pandemic and, therefore, well before the widespread adoption of generative AI.
The use of AI alone is not enough to explain the differences
There are significant disparities between countries. While France ranks around the OECD average, education systems in Singapore, Japan, South Korea, and Estonia achieve significantly higher results in all three assessed areas.
Among the highest-performing education systems, however, strategies for integrating AI vary greatly. In Japan, a particularly high proportion of students report never or almost never using AI chatbots for schoolwork, and Estonia also shows a relatively low frequency of use. Conversely, Singapore demonstrates that more frequent use of AI can coexist with very high academic performance.
The OECD notes in particular that Singapore and South Korea have developed large-scale policies promoting digital learning, with Singapore explicitly integrating AI tools. Despite very different strategies, Japan, Estonia, and Singapore therefore achieve high academic results.
More broadly, the issue of AI is part of the broader question of digital uses and their ability to support—or, conversely, disrupt—learning. Digital distraction is a telling indicator in this regard: while 28% of OECD students report that their classmates are distracted by digital devices during most or all science classes, this proportion is less than 10% in Japan, South Korea, and several Chinese jurisdictions. Preserving spaces for attention and cognitive effort is therefore essential to promoting the meaningful educational use of technology.
Between Moderate Use and Dependence on AI
PISA 2025 shows just how quickly AI has become integrated into learning activities. Only 14% of students in OECD countries report never or almost never having used an AI chatbot for any of the four activities studied. The most common use is “to help me learn,” followed by preliminary information searches, writing texts, and summarizing readings.
Students who report not using AI tend to achieve better results than those who do use it, but the OECD emphasizes that this association does not establish a causal relationship.
Some students may, for example, rely more heavily on AI precisely because they face greater difficulties. Furthermore, when summarizing texts or conducting preliminary research, moderate users—who use AI between once a month and twice a week—achieve better results than very occasional users or those who use it almost daily.
Diverse Uses of AI in Education
While AI allows students to delegate certain cognitive activities necessary for learning—such as summarizing, writing, arguing, or solving problems—it can also serve as a learning support by providing a clue, an explanation, or feedback that enables the student to continue their thinking.
It is therefore essential to distinguish between uses that replace cognitive effort and those that support it and promote more personalized learning. The #ppAI6 model helps distinguish these different levels of use, ranging from excessive reliance on AI to creative and transformative uses.
The PISA data do not allow us to attribute the decline in test scores to AI, since this trend had begun before AI became widespread. However, they reinforce a key point: how AI is used matters. The OECD also emphasizes that AI should contribute to challenging student learning rather than replace their own effort, as understanding requires students to perform the cognitive work themselves.
In a world where obtaining an answer is becoming increasingly easy thanks to AI, learning requires students to continue exerting cognitive effort and to regulate that effort when faced with complex tasks. Among the many current challenges facing the education system, it is therefore necessary to help students regulate their use of AI—both in and out of school—so that this technology supports the effort to learn rather than replacing it.
Margarida Romero, Research Associate, Universitat Internacional de Catalunya; Université Côte d’Azur; Instituto de Investigación en Inteligencia Artificial (IIIA - CSIC)
This article is republished from The Conversation under a Creative Commons license. Read the original article.
