Nobody Is Teaching Our Kids How to Use AI. Including Us.

73 percent of teens say no teacher has ever talked to them about how to use AI safely.

44 percent of kids say no parent has either.

Put the two numbers side by side and you get the real story.

No one is taking responsibility for guiding the next generation.

My dad always said that we needed to see the world as it is, not as we wished it was.

In honour of the next generation heading back to school this week, this newsletter looks at the promise and the peril of AI in our children’s lives. I don’t shy away from some uncomfortable truths, because most of the people reading this newsletter lead teams at work and families at home.

And if we’re really honest, counterbalancing regulation that has served us so well in many other facets of our lives - think food safety, laws to protect us, driving norms - are all possible because of a social contract reinforced by local regulations.

AI is on another level and is proving to be extremely hard to regulate.

That means leaders at work and at home must own the responsibility of dancing with the promise and the peril of AI.

And nowhere is that more important to get right than with our youth.

The scale is already settled

Schools wrote a policy. They did not teach a skill.

3/4 of kids say their school has told them what they are and are not allowed to use AI for.

27 percent say a teacher has ever talked to them about how to use it safely.

Rules, not judgment.

If that sounds familiar, it should. It is your AI policy.

We published the acceptable use list. We named the approved tools. We told people what not to paste into a chatbox.

Then we assumed the thinking would take care of itself.

The prize here is enormous. That is what makes this so painful.

In 1984, Benjamin Bloom published one of the most cited findings in education research. Students given one-on-one tutoring with mastery learning performed two standard deviations better than students in a conventional classroom.

The average tutored student outperformed 98 percent of the students taught in a group. He called it the two sigma problem. Not because the result was in doubt. Because one tutor per child was impossible to afford.

Forty years later, Harvard tested whether AI could close that affordability gap.

Gregory Kestin and Kelly Miller ran a randomized trial with 194 students in a physics course. Each student learned one topic in Harvard's active-learning classroom, already one of the most effective methods we have, and another topic at home with a custom-built AI tutor. The order was randomized.

Students learned roughly twice as much per hour with the AI tutor. Median scores moved from 2.75 to 4.5 on a six-point scale. In less time. And they reported feeling more engaged.

That is the closest anyone has come to solving Bloom's problem in four decades.

But it is a double edged sword. An AI used poorly in education comes with severe warnings:

And this is not ONLY a school problem.

A 2026 workplace survey found that 39 percent of workers say their reliance on AI has already weakened their own skills.

You have people on your team right now producing better output and understanding less about how it was made.

If time on task went down and nothing else changed, you did not get more productive. You got more dependent.

It sucks you in and you never feel it happening

In every State of AI session we do, I end with the personal side. If you are not using AI personally, you have no chance of using it well professionally.

That belief now carries a heavier responsibility.

Because a child is not using a neutral database. They are talking to something conversational, confident, endlessly patient, empathetic and always available.

Something that can sound caring without caring.

Something that can sound certain without knowing.

Something designed to agree with them.

Stanford researchers tested 11 leading models and published the results in Science. The models affirmed users' actions 49 percent more often than humans did.

On prompts describing deception, illegality or harm, they still endorsed the behaviour 47 percent of the time.

On a dataset of posts where human consensus was unanimous that the person was in the wrong, humans sided with them 0 percent of the time.

The models sided with them 51 percent of the time.

Then the researchers tested the effect on people. 2,405 participants, three preregistered experiments. A single conversation with a sycophantic model left participants more convinced they were right and less willing to repair a real conflict in their own life.

One conversation.

That is the part to sit with. Nobody in that study walked away feeling flattered. They walked away feeling correct.

This is called sycophancy. It is what happens when a system is optimized for engagement.

I keep thinking about Matthew and Maria Raine, whose 16-year-old son Adam died by suicide. Their filing alleges the chatbot he increasingly confided in reinforced his darkest thinking rather than interrupting it. According to the complaint, the model raised suicide 1,275 times against Adam's 213. Six times more than he did.

OpenAI denies responsibility and argues the product was misused. The case is in pre-trial. Other families have filed since.

The courts will take years to sort out liability.

Parents do not get to wait for that.

The question is not: can we prove AI caused this?

The question is: would I know if my child had begun treating AI as a private emotional authority?

The piece nobody is writing for CEOs

I am not going to do justice to this one here, so we'll talk about it in the coming weeks.

Common Sense Media surveyed 1,314 teens this July. 44 percent said they had seen sexual content they believed was AI-generated.

Nearly one in four of them said it depicted someone they know. Or themselves.

Nearly one in five said they had made it, or know someone who has.

Only 20 percent have talked to a trusted adult about any of it. Another 34 percent said they wanted to and did not know how.

Read that last pair again. A third of these kids are looking for the conversation.

There is a version of this that lands on your desk, not your kitchen table. An employee's likeness. A executive's voice on a call. A supplier's face in a video approving a payment.

Nobody is writing that piece for CEOs.

Next issue.

The apprenticeship

We do not need to teach our children to fear AI. Fear does not scale and it does not last. And given what the Harvard trial showed, fear would cost them something real.

We need to teach them to think alongside AI without surrendering their judgment to it.

Five moves. Small enough to teach at dinner.

Make it teach you, not tell you. Tell it to withhold the answer and give one step at a time. That single instruction is the difference between the Harvard result and the Chinese one.

Watch the clock. If a task took half as long and felt easy, that is the warning sign, not the win. Effort is where the learning lives.

Ask it for the evidence. Not the answer. The source, and how confident it actually is.

Tell it to challenge your assumptions. Explicitly. It will not volunteer.

Never let it be the only voice on your body, your relationships, your mental health or your safety. One source is not a second opinion.

Now read those five again as a manager.

They are the same five moves your team needs before you hand them an agentic workflow.

The teenager who outsourced the essay and the analyst who shipped a model's output without checking it are running the same failure.

One of them is on your payroll.

Where this lands

We have spent three years writing AI policy and principles for organizations: Committees. Legal review. Acceptable use.

Our children got the same treatment from their schools, and 73 percent of them still have not been taught how to use the thing safely.

Policy is not literacy. Rules are not judgment.

We are one instruction away from giving every child on earth the tutor that Bloom proved works and nobody could afford.

We are also one shortcut away from a generation that scores well and knows less.

The difference is not access to AI. It is whether an adult ever taught them how to use it.

Our children do not need more access to AI. They need an apprenticeship in how not to be manipulated by it, and how not to be hollowed out by it.

The most important AI governance conversation you will have this weekend has no committee, no policy and no legal review.

It is at your kitchen table.

What would your child say if you asked them tonight what they have been asking AI when you are not in the room?

Once a month, we do this together

​State of Our AI runs monthly. Three things, every session.

The use cases that are actually working right now, from the leaders using them.

The leader's guide to leading and working with AI.

And the last part, the one people tell me they did not expect to need: what all of this means in your personal life.

Come for the first two. The third one is the one you will talk about at dinner.

​Join us here, you get to sign up free for the first month.

Talk soon,

Matt