In a recent interview for RMF24, Prof. Piotr Sankowski — director of the IDEAS Research Institute, professor of computer science at the University of Warsaw, and a four-time ERC grant winner — said something that has stuck with me for days:

”(…) treat these models as a kind of sparring partner. Ultimately I hope solutions will emerge that we can give our children — ones that, when asked ‘write my homework for me,’ will refuse, and instead hold a discussion with the child (…) about how to write that assignment as well as possible (…) rather than just serving up ready-made answers.”

(My translation from the Polish original.)

It’s the most on-point thing about AI in education I’ve heard in a while. Let me unpack where that hope comes from, where it has a catch, and — most important for us parents — what to do today, before those “refusing” models actually exist.

Where the hope comes from: the two-sigma problem

In 1984, the educational psychologist Benjamin Bloom described a result that still electrifies. The average student working with a one-to-one tutor — with regular testing and feedback — outperformed 98% of students taught in a classroom. In statistics-speak: by two standard deviations (two “sigmas”). In plain terms: a middle-of-the-pack child moved almost to the very top.

It sounds like every parent’s dream. It was — with one catch. A one-to-one tutor for every child can’t be funded or scaled. It was always a luxury for the few. Bloom himself didn’t announce a fix; he posed a problem: how do you get group instruction to approach the effect of individual tutoring without hiring a tutor for everyone?

This is exactly where the AI hope enters. A large language model is cheap, available around the clock, and — in theory — can adapt pace and form to a specific child. If it worked like the tutor in Bloom’s studies, we’d have personalized education for every child, not just the ones who can afford a tutor. That’s a real, beautiful promise.

Where the catch is

Before we buy that promise whole — three things worth keeping in mind so you don’t sell your kid a miracle.

First, nobody has reproduced two sigmas cheaply. The 2σ figure itself is questioned in later research — replication attempts produced much smaller effects (around half a standard deviation), and tutoring meta-analyses typically land at 0.3–0.8. It’s a goal worth chasing — not a guarantee you can check off. Anyone selling you “AI will give your child a tutor’s effect” is oversimplifying.

Second, a model is not a human tutor. Bloom’s effect came from a person who made sure the student mastered the material before moving on. A chatbot that “serves up ready-made answers” to the first question does the exact opposite — it does the work instead of teaching. That’s not a flaw of any one model. The difference is in how you use it and how it’s designed. The same model can be a Socratic guide or a copy-paste machine, depending on how you set it up and what you teach your child to expect from it.

Third, the cultural background isn’t neutral. A student’s assistant carries the assumptions, examples, and language of whoever built it. For non-US families that’s an argument to root for local solutions — built with your own context and values — rather than adopting someone else’s blindly. Sankowski makes the point in the same interview: for some uses we need our own, sovereign models, and to learn how to build something, you actually have to build it.

What it means for a parent

Here’s the core. Sankowski says one more thing worth nailing down: hard things are still worth learning. AI doesn’t excuse a child from thinking — it raises the price of being able to think for yourself. The easier it gets to generate a mediocre answer, the more valuable the person who can judge it, fix it, and push further.

So the question isn’t “should I let my child use AI.” It’s “will I teach them to treat the model as a sparring partner, not a homework vending machine.” The first takes our presence. The second takes only a slogan.

Good news: you don’t have to wait for a model that “refuses.” You can switch on the sparring-partner effect today by setting up three things.

1. Change the question your child asks the model

The whole difference between doing-the-work and teaching lives in the phrasing. Teach your child to ask not for the answer but for guidance. Concretely — give them two opening lines for a chat:

  • Instead of: “Write me an essay about To Kill a Mockingbird.”
  • Try: “I’m in 7th grade and have to write an essay about To Kill a Mockingbird. Don’t write it for me. Ask me questions so I build the thesis and arguments myself. Tell me when something is weakly supported.”

The same works in math (“don’t give the answer, nudge me to the next step”), in science, in a foreign language. One sentence at the start turns a vending machine into a tutor.

2. Talk about what the model produced

Sankowski describes it precisely: the deepest understanding comes not when a child copies the output, but when they discuss it. “Why did the model write it that way? Do you agree? What would you change?” Two minutes of that at the dinner table beats an hour of solo clicking. It’s also the simplest test of whether your child understands what they’re handing in.

3. Teach them to spot when the model is wrong — and when it’s flattering

Models can be, as Sankowski puts it, “sycophantic” — very eager to confirm what we already think. Ask a leading question, get the answer you wanted to hear. A great exercise for an older child: ask the chat the same question from two opposing sides and watch it enthusiastically support both. It’s the cheapest critical-thinking lesson I know — and a vaccine against treating the model as an oracle.

One thing to do today

Sit with your child at the next homework assignment they want to “do with AI” anyway. Don’t ban it. Instead, rewrite the first line of the prompt together — from “write it for me” to “don’t write it for me, guide me.” Watch how the conversation changes. Ten minutes, and your child has a tool, not a stand-in.

Because the stakes here are simple: the same model does one child’s homework and teaches another to do theirs. Which one it’ll be isn’t decided by the technology — only by us.

So one question for tonight: the last time your child sat down with AI — did they come away smarter, or just done faster?