The OECD's PISA 2025 report, released September 8, 2026, gives schools and AI companies a blunt warning: students using chatbots heavily for homework are scoring worse, not better.
The numbers land at an awkward moment for an industry built on the promise that AI makes people smarter. According to the OECD's new PISA results, 15-year-olds who used AI chatbots almost every day for tasks such as summarising assigned reading, drafting text or doing preliminary research scored lower in science than students who didn't use the tools for those jobs. For summarising texts, the gap was nearly 30 points after accounting for socio-economic background.
That's not small.
PISA is not a classroom poll dressed up as research. The 2025 assessment tested more than 760,000 students across 91 countries and economies, representing about 33 million 15-year-olds around the world. It has tracked global student performance since 2000. When an instrument with that scale starts flashing a warning about a specific classroom habit, you should take it seriously.
Here's where it gets interesting for anyone building AI products. The relationship between chatbot use and test scores doesn't move in one straight line down. Students who used AI about weekly to help them learn, rather than to draft or summarise for them, scored around 500 in science, among the highest results reported for any AI-use group. The OECD also found that students who frequently used AI for learning and were taught at school to assess AI-generated information performed slightly better than students who used AI less often or not at all.
Robots Marched On Poland's Digital Ministry To Demand AI Regulation About 30 humanoid and quadruped robots demonstrated outside Poland's Ministry of Digital Affairs in Warsaw on September 7, demanding regulation of the AI and robotics that could replace human workers. Digital Affairs Minister Krzysztof Gawkowski met the organizers on site and called uncontrolled AI models in humanoids "a major problem." - robots protest Poland digital ministry AI regulation - humanoid robots demanding workplace protection laws Warsaw
So the problem isn't the chatbot by itself. It's the work the student hands over to it. A student who asks an AI tool to summarise a reading passage and then copies the answer has skipped the thing the assignment was meant to build. A student who uses the same tool to test their own understanding, and who has been trained to spot when it makes things up, is doing something different.
That distinction matters directly for edtech founders and anyone shipping AI into classrooms right now. Deploying a chatbot and calling it done is not a serious education product. Deploying one with prompts, checks and classroom guidance that force students to question the output at least has a case behind it. Frankly, the PISA data makes the lazy version look dangerous.
The wider decline makes the AI signal harder to ignore #
The AI gap is not happening in isolation. The OECD said PISA 2025 recorded the lowest average performance it has observed so far across science, reading and mathematics. Reading fell 28 points across OECD countries between 2015 and 2025, which the organisation said is about one and a half years of learning. Mathematics fell 22 points. One in five 15-year-olds across the OECD is now a low performer in science, mathematics and reading, up from 16% in 2022.
There are several causes here, and pinning the whole decline on chatbots would be too easy. The OECD points to digital distraction, weaker reading habits, school resource pressures and uneven student engagement. Still, AI now sits inside that broader problem. It arrived in classrooms fast, and most schools are still deciding whether it is a tutor, a shortcut, a cheating tool or some mix of all three.
The industry has spent the past year selling a simple productivity story. OpenAI, Anthropic, Google, you name it: the pitch is that these tools help you write faster and get more done. For an adult at work, that claim may hold. For a 15-year-old learning how to read, explain and reason, the same shortcut can remove the practice that makes the skill stick.
That's the uncomfortable part.
The OECD's own conclusion is narrower than a blanket warning and more useful than one. Students need to learn how to question AI-generated information, not just how to access it. Around six in ten students across OECD countries said they had been asked in school lessons to assess the quality of AI-generated information, but those opportunities were not evenly spread. Advantaged students reported more of them. That means AI may be creating another education divide before most school systems have even named it.
Two Dark Money Groups Outspent Victoria's Political Parties With AI Ads Two anonymous groups, Fix Victoria and Better Victoria, spent nearly $140,000 combined on AI-generated shock ads ahead of Victoria's November 2026 election, outspending several political parties on social media. A registration loophole means neither has to disclose who is funding them. - dark money groups AI political ads Victoria election - anonymous political advertising spending Victoria state election 2026
If you're building classroom AI now, the product spec is no longer just chat, answer, submit, done. It has to include friction that keeps the student thinking: asking for evidence, comparing answers, checking sources, explaining the reasoning in their own words. Otherwise the tool is not helping them learn. It is doing the learning-shaped part for them. Also read: China's OpenBMB Releases MiniCPM5-2B, Beating Every Open Model Under 4B • AI Startups Are Counting Signed Contracts as Real ARR, VCs Admit • SAP's Own Board Says It Needs an AI Breakthrough Within Months
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.