Kids Got Two Years of Free AI Tutoring and Mostly Asked It for Pizza Jokes

Something shifted this month, and it lands right on your desk if you build anything for learning or families. The tools got good enough to do the homework, and the kids mostly shrugged. Parents are cloning themselves. MIT is thinking about tearing up how it teaches. Let me catch you up.

The deep cut

• Access is not engagement. Khanmigo reached nearly every kid in 18 Tennessee schools, and thin use followed.
• Motivation is the feature you cannot ship. Sal Khan learned the tutor works only when a keen student already wants it.
• The best AI knows when to shut up. Quest and Peanut both won by pointing users back to people, not answers.

The homework problem became an institution problem

MIT's ad hoc AI committee put a stamp on what teachers already knew. AI can produce credible solutions to almost any written assignment in the undergrad curriculum, from essays to proofs to code. That is not a cheating story. It is a curriculum story.

The knock-on effects are what should catch your eye. In under three years, the report tracks fewer students at office hours, less online discussion, and fewer study groups in dorms and libraries. When the tool does the task, the social fabric around learning frays. Princeton dropped its century-old Honor Code after a cheating scandal. Some MIT professors are back to oral exams and handwritten essays.

If your product removes friction from a learning task, ask what social behavior that friction was holding up. Sometimes the effort was the point.

The tutor kids refused to use

Here is the result that should reset your assumptions. Scientists gave middle schoolers two years of access to an AI tutor, and access was nearly universal but engagement was thin. Kids sent off-topic prompts. They asked for jokes and essays about pizza. They tried to trick it into handing over answers.

"How much motivation matters," Sal Khan told Chalkbeat, was the thing the study put in a spotlight. Stanford landed in the same place: the strongest results came from AI built for human tutors, not for students working alone. One researcher named the trap plainly. The same tool that makes a good tutor also makes your life easier, and easier means less effort, less learning.

You cannot ship motivation. If your onboarding assumes a curious, driven user, build for the ones who are not.

Parents are patching a gap, not buying efficiency

The family use case is stranger and more human than the pitch decks suggest. A neuroscientist built an AI clone of herself while traveling, trained on texts and family rules, so her 15-year-old had someone to ask. Her son promptly used it to build a case for a coffee shop hangout. "I was manipulated by my own reasoning," she said.

Peanut, the app for moms, noticed something you can steal. Users kept double-checking AI parenting advice against real community members. So Peanut pointed its tool back at the community. Its president, Michelle Battersby, put it flat: moms are not using AI to make parenting more efficient, they come to it to navigate uncertainty. That is a different product than a faster answer machine.

Restraint is becoming the design edge

The instinct to hold attention is losing its shine. Mat Honan, whose own kids traded flip phones slowly for iPhones, describes parents treating tech skepticism as a raging flood. Australia banned under-16 social media. Schools are swapping iPads for books. Even Zuckerberg keeps his kids' faces offline.

Contrast that with Quest, a concept nature guide for kids six to eleven. Its whole trick is that the intelligence knows when to step aside. No dominant screen. It asks a child to look closer before it explains anything. That is a design position, and it is the opposite of engagement-maximizing.

Bill Gates, who now debates sodium batteries with Claude and ChatGPT at 3 a.m., still warns the industry is crossing the safety lines it set for itself. His point for you is quieter than the headlines: enthusiasm and concern can live in the same person, and pretending otherwise is bad PR dressed as strategy.

Three questions for your team

• Where does our product remove effort that was actually doing the learning, and can we tell the difference?
• If our best outcomes need a motivated user or a human in the loop, are we designing for that, or hoping for it?
• What is our stated position when a parent, teacher, or teen asks why they should trust this with a kid, and can we say it in one sentence?