GlobeskoolKarpathy Wants to Build AI Education. Here Is What Parents Should Actually Demand It Looks Like.
Andrej Karpathy wants AI to reshape how kids learn. Most AI tutoring tools just hand kids faster answers. Here is the difference that actually matters, and what to ask before trusting any tool with your child.
The Ambition Is Real. The Problem Is What Gets Built Next.
Andrej Karpathy spent roughly two years building Eureka Labs, an AI-native education company, before departing to join Anthropic. Before he left, he said he remains deeply passionate about education and plans to return to it. The edtech world took that seriously. Parents probably should too.
But there is a version of AI education that makes kids sharper, and a version that just makes them dependent faster. Right now, most of what is being built and marketed to families is the second kind, dressed up to look like the first.
The tell is simple: does the AI ask your kid a question before it answers one? If not, it is not a tutor. It is an answer machine with better branding.
What Most AI Tutoring Tools Are Actually Doing
The dominant loop in AI tutoring right now goes like this: kid types a homework question, AI returns a clean correct answer, kid moves on. That is not education. That is autocomplete for assignments.
The thing it removes is friction. And friction is not the enemy of learning. In many cases, friction is learning. The moment a kid sits with a problem they cannot immediately solve, and chooses to push through rather than paste it somewhere else, that is where actual skill gets built. The judgment to frame a problem well. The tolerance to sit with uncertainty. The ability to notice when an answer feels wrong.
AI tools optimized for speed and satisfaction scores are quietly eroding all three.
Karpathy himself put it plainly: the goal should be students who are proficient with AI but also able to exist without it. That second clause is not a footnote. It is the whole challenge. And almost no tool being built today is optimized for it.
Four Things Worth Demanding From Any AI Education Tool
Before you hand your child a subscription to anything calling itself AI-powered learning, here are four concrete things worth evaluating.
Does the tool ask the kid a question before it answers one? A tutor who only delivers answers is not building judgment. It is building the habit of waiting for answers. That is the opposite of a Future Skill.
Does it surface confusion rather than paper over it? The goal is for your kid to understand why they were wrong, not just to see the right answer appear. If the tool corrects without explaining, or explains without checking whether the kid actually followed, it is managing the appearance of learning, not the thing itself.
Does it make itself progressively less necessary? A good tutor works toward their own obsolescence. Most apps are designed for daily retention and engagement metrics. Those two goals are not the same as building an independent thinker. Ask whether the tool has any mechanism for pulling back its support as a kid’s skill develops. Few do.
Can your kid explain what they learned once the session ends? Close the app and ask. If the answer is a shrug, the tool did the thinking. Your kid watched.
The Skill That Does Not Show Up on a Report Card
There is a capability that sits underneath all of this, and schools do not grade it: knowing what question to ask first.
When a kid encounters a problem they have never seen before, the first move is not to solve it. The first move is to understand it well enough to frame it. That is a learnable skill. It takes practice, repetition, and yes, friction. Kids who get good at it can use AI as a genuinely powerful tool, because they know what to ask it, how to evaluate what it returns, and when not to trust it.
Kids who have been trained to paste questions and accept answers have none of that. They have faster homework. They do not have a skill.
What Good AI Education Actually Looks Like
It is not anti-AI to demand more from AI education tools. The goal is not to keep kids away from these tools. The goal is to make sure the kid is in charge of the tool, not the other way around.
That means real projects with real outputs, not worksheets dressed up in an interface. It means the AI prompting the kid to think before it thinks for them. It means measuring progress in judgment built, not problems completed. And it means core subjects sitting alongside the skills that make those subjects matter: critical thinking, problem framing, communication, and the ability to work with technology rather than defer to it.
Globeskool is built for kids aged 8 to 16, and this tension sits at the center of everything it does. If you want to see where your child actually stands before committing to anything, the free assessment takes about 5 minutes and gives you a real picture to work from.