Google Dropped From 72% to 47% in Two Years: Inside the NTNU Review That Says We Know Almost Nothing About AI and Kids' Thinking

On September 22, 2026, Norwegian SciTech News published the findings of a systematic literature review that should unsettle anyone building education policy on top of generative AI. Researchers led by Hilde Storrøsæter, a PhD research fellow at NTNU’s Department of Teacher Education, analysed 173 international studies on the use of generative AI in schools and education — and their central conclusion is not that AI is good or bad for learning. It is that we know remarkably little about the group whose learning is changing fastest: children.

The review, conducted with colleagues from Czech universities (Tomáš Bendl, Lenka Krajňáková, Emil Drápela and Martin Hanus of Masaryk University), appears in the Journal of Information Technology Education: Research, Volume 25, 2026, under the title “Dual Nature of GenAI: Catalyst or Barrier for Critical Thinking and Problem-Solving Skills? A Systematic Review” (DOI: 10.28945/5834). It covers research from 2022 — the year large language models swept onto the world stage — through the spring of 2025.

The headline number: Google is losing an age cohort

The single most striking statistic in the NTNU write-up does not come from the review itself but from the Norwegian Media Authority’s survey of media habits among 9-to-18-year-olds: the share of Norwegian kids using Google search fell from 72% in 2024 to 47% in 2026. Two years, a 25-percentage-point collapse, in the age group that most needs to learn how to find, compare and evaluate information.

The same survey found that 39% of Norwegian 11-to-12-year-olds now use AI. These are pupils two years away from their teens, outsourcing questions they would previously have typed into a search box — and, crucially, skipping the entire skill stack that search used to force: formulating queries, scanning snippets, distinguishing an advertisement from a result, choosing between conflicting sources.

The traffic data corroborates the survey. ung.no, the Norwegian government’s official information channel for 13-to-20-year-olds, saw visits fall from 24 million to 18 million between 2024 and 2025 — a 25% decline in a single year. The NTNU article attributes part of that drop to Google’s own AI summaries, which intercept questions before they ever become a click. When both the search engine and the government’s youth portal are losing an entire generation’s questions to chatbots, the infrastructure of “looking things up” is being quietly rewired.

The core finding: AI is a mirror, not a motor

Strip away the adoption statistics and the review’s scientific finding is more nuanced — and in some ways more interesting — than either the techno-optimist or the doom camp would like.

“The most important finding is that AI in and of itself does not make us any better or worse at thinking,” Storrøsæter told Norwegian SciTech News. “It’s easy to believe that the tools will either strengthen or weaken learning. We found a clear dual effect. The same technology can do both.”

The numbers behind that “dual effect” are worth sitting with. Of 139 studies that examined AI’s effect on critical thinking, 67 reported only positive effects, 33 only negative, and 39 both. Of 103 studies on problem-solving ability, 57 were exclusively positive, 23 negative, and 23 mixed. The evidence tilts positive — but not overwhelmingly, and not in a way that survives contact with how the tools are actually used.

The mechanism matters more than the count. AI tools support critical thinking when used to challenge, ask questions, compare viewpoints and build one’s own line of reasoning. The risk flips when “AI becomes a shortcut to a ready-made answer” — at which point, Storrøsæter says, “the chance of overdependence, more superficial thinking and weakened independent problem solving increases.”

In other words, the technology is a mirror of pedagogy. The same chatbot is a Socratic sparring partner in one classroom and a homework-vending machine in the next. That is an uncomfortable finding for anyone hoping the tool itself would settle the debate.

The three knowledge gaps

The review identifies three structural blind spots in the research literature, and the first is the one that should matter most to parents and policymakers.

Gap 1: Children are barely studied. In 80% of the analysed studies, the participants were university students — overwhelmingly adults whose critical-thinking habits, for better or worse, are already formed. “We can’t simply assume that what applies to a 22-year-old university student also applies to a 10- or 14-year-old,” Storrøsæter argues. The timing is the problem: children are still building the very capacities — perception, information processing, evaluation — that adults take for granted. “If AI takes over the thinking process before pupils have had time to develop their own skills for critical thinking and problem solving, the consequences could be completely different for them than for adults.”

Gap 2: The evidence is largely self-reported. Most studies rely on participants describing their own abilities and usage — a method the researchers flag as less reliable than objective, longitudinal measurement of change over time. A literature built on self-report is a literature that cannot confidently detect gradual cognitive offloading, precisely the effect everyone fears.

Gap 3: The geography is skewed. South, Southeast and East Asia — primarily China — dominate with 28 studies. Spain and the United Kingdom lead Europe with nine each; the entire United States contributed 15. Africa, South America and Central Asia are, in the researchers’ words, seriously underrepresented. Any claim that “research shows AI is safe for classrooms” is quietly a claim about Chinese, Spanish and British undergraduates.

Context: Norway is the canary, and it knows it

The review lands in a country that has been unusually deliberate about AI in schools. In June 2026, Norway imposed a near-ban on generative AI for elementary school pupils, restricting use by 6-to-13-year-olds and allowing 14-to-16-year-olds access only under teacher supervision. The policy followed Norway’s earlier smartphone-ban playbook, which its own research linked to reduced bullying and better grades.

Against that backdrop, Storrøsæter’s position is notably not prohibitionist. “I support being restrictive in terms of not using AI with very young pupils. But when AI is used, it has to happen on the teacher’s terms,” she says. Her real worry is distributional: uncritical use builds no skills and burns learning opportunities, which “will only increase disparities between those who use AI critically and as a tool, and those who use it uncritically simply to complete a task faster.” An AI divide layered on top of existing educational inequality — with the differences compounding year after year — is a far more plausible failure mode than a uniform collapse in cognition.

The prescription that follows is teacher-first. “Teacher education has a dual task. We have to educate teachers who can use and relate critically to AI themselves. At the same time, our education students should also be able to teach children and young people how to use AI.” And the design question she poses is the one every school will eventually have to answer: which parts of the thinking process are important for pupils to still do themselves, and how do we design tasks where AI is a tool rather than a replacement for thinking?

Why this matters beyond Norway

The rest of the world is running the experiment in reverse order. UNICEF’s Innocenti office estimated in June 2026 that up to 50% of surveyed children already use AI, much of it for homework. Pew reported in February 2026 that just over half of US teens have used chatbots for schoolwork. Policymakers in most countries are deciding whether to encourage, tolerate or restrict classroom AI on an evidence base that — as this review documents — is 80% adult, mostly self-reported, and geographically concentrated in three regions.

Norway’s Media Authority data gives the debate its first clean leading indicator: when 9-to-18-year-olds abandon search faster than any other cohort, the behavioural shift has already happened, regardless of what the schools decide. The NTNU review’s quiet contribution is to establish that the research needed to evaluate that shift has not been done — and to specify exactly where the gaps are, so that the next 173 studies can aim at children, use objective measures, and sample beyond three regions.

“The education sector is already strongly influenced and shaped by AI,” Storrøsæter concludes. “The question is whether we can shape the use of AI so that it supports students’ thinking instead of taking it over.” Two years into the fastest information-behaviour change on record, that question is now simultaneously urgent and — by the best available evidence — unanswered.