PISA 2025: Daily AI users scored roughly 20 points lower in science
AI-assisted search and Gemini have become standard on school-issued Chromebooks across much of the US. At the same time, the research base on how this affects learning is nearly nonexistent — and some schools are already blocking Google's AI answers from students' machines…
No tech giant has been more aggressive in placing AI in front of schoolchildren than Google, according to a major report originally published in the Wall Street Journal and carried via MSN (aa4b0abe). The company has rolled out AI-assisted search and Gemini as defaults on Chromebooks and across the Google Classroom workflow — that is, on the machines students use every day. The report, built on interviews with more than 60 teachers, school leaders, parents, students and clinicians, concludes that the rollout has contributed to making many students dependent on AI both academically and emotionally.
Landing in the same period are new PISA figures for 2025, showing that American reading scores have fallen 14 points since the previous measurement in 2022, and NPR reporting that schools are adopting AI tools without policy or documented effect (65c76137). That makes this a classic modern dilemma: the technology is distributed to millions of devices before anyone knows whether it does more harm than good.
From B student to failing math
The most concrete story in the report concerns an eighth-grader from Pennsylvania who used the AI-assisted Google search field on his school machine to complete schoolwork. According to the student's own testimony and his transcript, he received mostly Bs in his quarterly assessments — but failed his math exam at the end (aa4b0abe).
The case is anecdotal, not evidentiary: one student, one self-reported usage history and one transcript. But it illustrates the mechanism teachers describe. AI-assisted search makes it possible to get finished answers directly in the search field, without the student needing to formulate their own query, evaluate sources or solve the problem themselves. When the tool is pre-installed on the machine the school has issued, the threshold for use is practically zero — and the threshold for dependence follows.
The evidence: 60 voices, one teacher's exam observation
The report's breadth lies in its interviews. More than 60 sources — teachers, school leaders, parents, children and clinicians — describe the same pattern: students turning to AI for help before trying themselves, and young people increasingly treating chatbots as conversation partners (aa4b0abe).
In Miami-Dade, where the school district has made Gemini available to all high school students, AP Art History teacher Kaitlyn Ruano described unexpectedly high failure rates on the AP exam among students who had used AI throughout the year. She relayed the students' own explanations after the exam: "I had students come to class afterward and say, 'Miss, I had no idea about half the material, I'm not going to lie to you. I've used AI all year'" (aa4b0abe). The school district declined to comment. That means the AP failures are so far one teacher's observation — not verified exam data.
The PISA numbers — and why they don't prove causation
The broadest dataset in the picture is the OECD's 2025 PISA survey. Two findings are highlighted in the report:
- American reading scores fell 14 points from the 2022 survey.
- Daily AI users who used AI for concrete school tasks scored roughly 20 points lower in science than students who never used AI — about one school year, according to the OECD's 2025 figures as relayed in the report (aa4b0abe).
This is correlation, not causation. Students who use AI daily for schoolwork may be students who were already struggling, or students at schools that underperform for other reasons. And the OECD's own analysis points in a different direction on one point: AI users who had received instruction in AI literacy actually scored slightly higher. The difference likely lies in how AI is used — as an extension of one's own thinking or as a replacement for it.
That means the honest picture is this: there are simultaneous signals of falling results and widespread AI use, but no documented causal chain between them. That is precisely why Rodriguez's formulation is apt.
"An experiment without a control group"
"We are running this experiment on an entire generation's cognitive development, with no control group," said Mikel Rodriguez, a former Google researcher who left last year, in the report (aa4b0abe).
His point is methodological, not merely rhetorical. A control-group design would compare students with AI access against students without, over time. When AI is instead rolled out as a default on all school machines at once, there is no natural comparison group — and no way to isolate the tool's effect after the fact. What remains are correlational analyses like PISA, which by definition cannot close the causal question.
Google's side — and Google's own research
Google does not dispute that learning outcomes are the benchmark. A company spokeswoman said teachers around the world use the technology every day because it "solves real classroom challenges and helps improve learning outcomes." She added that Google focuses on encouraging critical thinking and has specific protections for users under 18, including to prevent emotional dependence (aa4b0abe).
These are corporate claims, not documented findings. In the available source material, there is no published data showing that the protections work as described.
At the same time, the report points to something that sharpens the tension: Google's own researchers have written that the technology can pose elevated risks for children — who struggle to distinguish chatbots from human interaction — and has potential for cognitive and emotional harm. This comes through interviews with current and former Google employees and a review of more than a dozen Google research reports (aa4b0abe). An important nuance: the reports are known only through the article's account, not through the primary documents themselves. It is worth noting that a company simultaneously producing research on AI's risks for children is rolling out the same technology to schoolchildren as a default.
The evidence gap: 800 articles, very little knowledge
If the research base on the student side is thin, the one on the pedagogy side is thinner. A recent Stanford review of more than 800 academic articles found that research on how schools can use AI to help students — while avoiding potential harms — is extremely limited (65c76137).
That is the backdrop for NPR's story: school districts across the US are adopting AI tools without policy, without evaluation and without evidence to draw on. The decision to let Google build AI into the school machine itself is thus made in an information vacuum — where the tool is free or nearly free, integrated into existing workflows and immediately available, while the costs (if any) arrive later and are hard to measure.
What schools are actually doing: Fairfax County's plug-in
The most concrete countermeasure in the material comes from Fairfax County Public Schools. The district has developed its own plug-in that blocks Google's AI search results on students' devices, and says it plans to roll it out to student machines soon (aa4b0abe).
Three things are worth noting. First, the plug-in is announced, not deployed — its effect is currently unknown. Second, the measure itself says something: a school district that adopted Google's Chromebook platform is now building its own tools to switch off parts of it. Third, the contrast with Miami-Dade, which declines to comment, shows how differently districts are responding to the same situation. There is no nationwide policy; each school is alone with the choice.
What remains to be known
Three questions remain open, and none of them can be answered with today's data:
Longitudinal data. There is no controlled study of AI use on school machines over time. PISA provides snapshots, not follow-up. Until such data exists, any claim that AI causes learning decline is an overstatement — and any claim that it does not is impossible to substantiate.
The role of instruction. The OECD's finding that AI users with AI literacy instruction scored slightly higher suggests the difference lies not in the tool but in its use. But there is no documentation of which training models actually work, or at what age.
Google's under-18 protections. The company claims specific protections against emotional dependence. None of the available sources document that they work, and several of Google's own researchers have written about precisely this risk.
In the meantime, the rollout continues. Rodriguez's description of an experiment without a control group may be the most precise thing in the entire story: not because the outcome is established to be harmful, but because no one — not Google, not the school districts, not the OECD — has designed the study that would show it.