Young users are replacing Google with AI, and the evidence on learning is largely missing

A dual effect, not a verdict

Researchers at the Norwegian University of Science and Technology (NTNU) reviewed 173 international studies on generative AI in schools and education, covering work published from 2022, the year large language models broke into the mainstream, through spring 2025. The review was led by Hilde Storrøsæter, a PhD research fellow at NTNU's Department of Teacher Education, working with colleagues from Czech universities, and was reported by Norwegian SciTech News.

The headline result cuts against both utopian and doom-laden takes: AI on its own neither improves nor degrades critical thinking and problem solving. The same technology can do either, depending on how it is used. When students use AI to challenge their reasoning, pose questions, compare viewpoints and build their own arguments, it supports those skills. When it becomes a shortcut to a finished answer, the risk of overdependence, more superficial thinking and weaker independent problem solving rises.

The underlying numbers lean positive but are mixed. Of 139 studies examining critical thinking, 67 reported only positive effects, 33 only negative effects and 39 both. Of 103 studies on problem-solving ability, 57 were exclusively positive, 23 exclusively negative and 23 mixed.

Google is losing the young

Usage data from Norway shows how quickly habits are shifting while this evidence gap persists. According to the Norwegian Media Authority's survey of media habits, cited by Norwegian SciTech News, Google search use among 9-to-18-year-olds fell from 72% in 2024 to 47% in 2026. Among 11-to-12-year-olds, 39% already use AI.

The Norwegian youth information site ung.no has felt the shift directly: visits dropped from 24 million to 18 million between 2024 and 2025, a decline attributed in part to young people turning to AI tools for answers and in part to Google's own AI-generated summaries reducing referral traffic.

Three gaps in what we know

Storrøsæter's team identified three structural weaknesses in the research literature.

The first concerns who was studied. In 80% of the analysed studies, the participants were university or college students, meaning very little is known about children and teenagers whose abilities to perceive, process and evaluate information are still forming. Storrøsæter cautioned that findings from a 22-year-old student cannot simply be transferred to a 10- or 14-year-old, and that if AI takes over parts of the thinking process before pupils have built their own skills, the consequences could be entirely different than for adults. "It is striking how little solid knowledge we have about how AI use affects basic skills in younger pupils," she said.

The second is methodological. Most of the reviewed articles rely on self-reporting, which is considered less reliable than objective studies of change over time.

The third is geographic. South Asia, Southeast Asia and East Asia, primarily China, dominate with 28 studies, while Spain and the United Kingdom contributed nine each and the United States 15. Africa, South America and Central Asia are described as seriously underrepresented.

The teacher problem

The review pinpoints three factors that decide which way the dual effect tips: how AI tools are actually used, teachers' own understanding of AI, and the demands teachers place on students. Storrøsæter argues that teacher education has a dual job, producing teachers who can both use and critically assess AI themselves and teach children and teenagers to do the same, rather than merely training them on specific tools.

Her position on classroom use is pragmatic. She supports restricting AI with very young pupils, but says that when it is used, it should happen on the teacher's terms. Her concern is uncritical use, which she warns builds no skills and skips learning situations, widening the gap between students who use AI as a tool and those who use it simply to finish tasks faster, with significant challenges for schools and universities if that happens.

Why it matters

The core problem is a mismatch of speeds. Adoption is outrunning the evidence: Norwegian children are abandoning search engines for chatbots at a striking pace, while the research base for that exact age group barely exists, dominated instead by self-reported data on university students from a narrow set of countries. Because the review suggests the effect of AI depends entirely on how it is deployed, the choices teachers and schools make about task design may matter more than the technology itself. Until better research on school-age children exists, educators and policymakers are steering by assumption rather than evidence, during precisely the years when critical thinking skills are formed.