The World Economic Forum just described the problem we started with

In June the World Economic Forum published Shaping the Future of Learning: Education Readiness for the Age of AI — a readiness framework for how school systems should adopt AI. Reports like this are written for ministries and school boards, and most of this one is. But its chapters on risk are written in plain language, they are backed by numbers, and they describe — almost sentence for sentence — the problem STEMBuddy was founded on. If you are a parent deciding what role AI should play in how your kid learns, the risk chapters are worth twenty minutes of your time. Here is what they say.

The risk now has an official name

When we started STEMBuddy, our founding argument was a sentence: the cost of getting a confidently-worded wrong answer dropped to zero, and the cost of building the judgment to recognize one didn't. Back then that was a founder's claim. The WEF's report now makes the same argument under a formal name — cognitive atrophy — and states the mechanism plainly: when an AI system does the reasoning, the cognitive work that would have built understanding simply doesn't happen. Learning is not receiving an answer; it is the struggle that the answer-machine removes.

The numbers underneath that claim:

• The share of students using AI for homework help jumped from 48% to 62% during 2025, per RAND research cited in the report — and 67% of students themselves recognize it may be hurting their critical thinking.
• An Anthropic analysis of roughly 575,000 student AI conversations found that close to half were seeking finished answers or ready-made content, with minimal cognitive engagement — some asking directly how to avoid plagiarism detection.
• In one experiment the report describes, students improved while they had access to an AI tool — then performed worse than students who never had it once the tool was removed. The tool wasn't building capability. It was renting it out.

The confident wrong answer, measured

The report's misinformation chapter contains the single statistic I would put on a poster. When AI models hallucinate — generate false information — they are about 34% more likely to use words like "definitely" and "certainly" than when they're right. Confidence and correctness are not just uncorrelated in these systems; at the moment of failure, the confidence goes up.

Now pair that with the report's finding about age: primary-school children are the group least likely to question an authoritative-sounding source, and the most susceptible to forming lasting misconceptions from one. The most confident-sounding wrong answers are being served to the audience least equipped to catch them. That is the exact gap an 8-to-12-year-old needs to be trained for, and it is not trained by lectures about "being careful online." It is trained by practice: making claims, checking evidence, running fair tests, and being wrong in ways that teach.

What the report tells parents to do

Buried on page 21 is the report's most practical section: a set of questions it hands to parents. Paraphrased: Is my child using AI transparently — and can I understand what tools are used and what data is collected? Is my child developing skills and habits, or producing AI-assisted outputs that merely resemble learning? And how do I set informed boundaries at home?

Read those again. They are pointed questions, and they deserve answers in writing — from any product, ours included. So we wrote ours down on a permanent page: How STEMBuddy uses AI — and where it deliberately doesn't. What the AI does, what it will never do, what it sees, and who's in control. If you're evaluating any AI learning tool — including ours — bring the Forum's three questions and see which products can answer them without flinching.

What we're not changing

It would be easy to read a moment like this as vindication and coast. The more useful reading is that the bar just went up. The report's framework asks whether learning experiences are "intentionally designed to enforce desirable difficulty" — whether the productive struggle is protected, not smoothed away. That has been our design bar from day one: the kid writes the hypothesis, designs the fair test, defends the conclusion. The AI asks questions, offers structure, and refuses to do the thinking. None of that changes.

The report is also honest about what no app provides: learning is relational. Kids learn with and from people. We agree, which is why every STEMBuddy investigation ends with the kid explaining what they found to a real person — usually you — and why we don't build kid-to-kid social features at this age. The collaborator in the loop is the parent. That is a limitation we chose on purpose, and we'd rather name it than paper over it.

The report closes with a question for every school system: will AI be shaped to strengthen human learning, or allowed to replace it? Parents don't have to wait for school systems to settle that. The version of the question you can act on today is smaller and sharper: when your kid sits down with an AI tool, who is doing the thinking? Pick tools where the answer is your kid. That's the entire test.