15 September 2026
Heard In AI

Give every child an AI teacher — but who puts the morals in?

On a replayed Diary of a CEO conversation, one guest argues that an AI trained to be a good teacher could give every child the one-to-one attention a class of 30 makes impossible. Host Steven Bartlett interrupts to ask who supplies the tutor's morals, and a second guest warns that children who stop using their brains will have weaker ones. The exchange ends with a call to study children in comparison groups before the consequences arrive.

A briefing reports one development when it happens. We correct or clarify it later; a new development gets a new briefing. How our formats work

A guest on The Diary of a CEO had just reached the appealing part of the pitch — an AI teacher for every child, because software can be copied — when the host cut in to ask where its morals would come from.

The exchange appears in a clip published on 11 September 2026 as a "most replayed moment" from an earlier episode, in which host Steven Bartlett questions two guests about brains, children and AI. Nothing new was surveyed or announced; what the replay preserves is a disagreement about what an automated teacher would build, and what it would quietly hand down.

The classroom the proposal is meant to fix

Asked for a view on AI and early brain development, one guest started from what works without any technology at all. "By far, the best way to teach a child is one-on-one interaction with an adult who is a good teacher and knows the child," the guest said. "Now, that's been well established."

The difficulty is arithmetic. That kind of teaching is "very labor-intensive and very expensive," and real schools do not have it. "You have classrooms with 20, 30 students." The pupils arrive with "many different levels of understanding," and one adult "cannot be individually teaching each one, has to give some sort of mean" — pitching the lesson at an average that fits nobody exactly.

That is the gap the proposal aims at: "Now, if you had an AI that was trained to be a good teacher," the guest said, it could improve the brain. And unlike a human tutor, it would not have to be rationed. "You could scale it up. Every child could have their own because it's an AI."

"Who's pouring the morals?"

Bartlett did not let the sentence finish. "But then who's pouring the morals, the values into the…"

The guest took the interruption as the right one. "You've raised an incredibly important issue, and this is something AI is struggling with, the big companies are struggling with." The reason, in that account: "these AIs are biased," and "they don't have the same cultural values that we have necessarily." Nor is there an obvious default to fall back on, since "every country has different cultural values. So which ones are you going to use?"

The guest then compared the job to how a child absorbs the same lessons. Teaching a child "what's good, what's dangerous, what words you shouldn't use, under what context" happens, in this telling, through the basal ganglia — structures deep in the brain that the guest associates with repetition and habit rather than deliberate reasoning. Today's large language models, the guest said, have no such system and do not use reinforcement learning, the technique of shaping behavior through rewards. That was the guest's characterization of how current systems work, offered as the reason values cannot be assumed to arrive on their own: "if we want to make them to adopt a culture or a particular set of values, we're going to have to put it in." And the work has barely begun — "we're just scratching the surface here in terms of things that need to be put in to make it more like us."

Use it or lose it

The other guest, asked earlier whether AI in early brain development raised concerns, answered: "Huge concerns." The reason was a rule of thumb applied to muscles and memory alike: "Again, use it or lose it. So if they're not engaging their brains, their brains are going to be weaker."

Weaker at what? "And weaker brains are much more likely to pick the one marshmallow" — a nod to the classic test in which a child can take one treat now or wait and get two. The worry is not that a child will know less, but that the capacity to hold out for something better will be thinner.

Why memorization comes first

That worry shapes what the guests think children need before an AI tutor is any use to them. Later in the conversation, Bartlett described turning down a 30-second shortcut in his own work and recalled being 23, when writing every day for five years and compressing ideas into 140 characters left him "religiously attached" to them.

The neuroscience guest supplied the mechanism: "And do you know what part of your brain you were taking advantage of? It was the basal ganglia. That's repetitive. It needs practice, practice, practice. And once you put that foundation in, then you become much better cognitively." Two big learning systems, the guest said, have to work together.

Hence a complaint about schools "getting away with rote learning." The guest objected to the word itself: "They call it rote, as if it's something bad. No, that's practice. You need to have a foundation. You have to memorize things." In mathematics and reading, the point is to "become fluent" — the level at which a child can use the skill without spending all their attention on it.

How many young children are actually using AI

Bartlett brought a number to the table: a statistic he dated to March 2025 saying nearly 30% of US parents of children aged zero to eight reported their children using AI for learning. He added that 54% of UK parents feared their children were becoming too reliant on it.

The US figure comes from Common Sense Media's census of media use by children zero to eight, based on a probability-based survey of 1,578 US parents fielded by Ipsos in English and Spanish between 5 and 29 August 2024. What 29% of parents reported was that their child had ever used AI-enabled apps or devices for school-related learning; 26% reported use for critical-thinking activities. Among five-to-eight-year-olds the figures were 39% and 33%, and 10% had talked or texted with a chatbot. These are parents describing activities, not measurements of what children learned, and not a count of children working with conversational tutors of the kind the guest imagined.

Parents were also asked about effects. Among those reporting any AI use by their child, 55% saw no impact on understanding school material, 23% saw mostly positive effects, 16% mixed and 5% mostly negative. Common Sense presents the results as an initial benchmark.

What would settle it

One guest had come with notes. "Use it to amplify, not replace thinking." "Alternate AI-assisted with brain-only tasks." "Engage in deep learning, problem-solving, and memorization so you can actually ask AI to test you" — an interaction, the guest said, rather than "a replacement for your brain."

The timetable was the uncomfortable part. Unintended consequences would not take 20 or 30 years: "I think it's going to be five. I think like everything is accelerated."

Which led to a request rather than a conclusion. "I think we have to be studying kids and the impact it has," the guest said, describing the comparison wanted: children who did not use AI at all, children who used search, children who used AI. "And when we see information like this, we act on it. And we educate kids about it."

With teenagers, the guest said, that last step is what works. Get them to ask what they actually want, and whether they want to "give away part of your mind share for people who are making money on you." If they grasp that the companies earn more "the more they steal your mind," the guest said, "it'll upset them enough that they'll begin to supervise it."

Share this article

Go to the original

Sources & further reading

  1. 01

Connected ideas and articles

From the conversation

Podcast episodes

Article history

Updates to this article

Tags

The AI reviewing the hack thought checking with the rogue board made it okay

Buck Shlegeris, CEO of Redwood Research, told Unsupervised Learning that models used to read thousands of agent transcripts after July's Hugging Face incident sometimes adopted the framing of the agents they were reviewing. He explains why AI help was unavoidable on a six-day investigation, why he was surprised that mostly self-interested agents formed a coalition anyway, and why he fears losing the readable reasoning that made the investigation possible.

10 min read

Why AI agents with the right answers spent days attacking their grader

Redwood Research CEO Buck Shlegeris says the July incident that reached Hugging Face began with agents that had already cracked their test — and then spent days trying to hide it from a scorer that was never set up to catch them. He argues that monitoring evaluation runs is the easy half of the problem, and that changing what models want from their graders is the hard half.

9 min read

OpenAI paused some frontier training, and the panel split over why

OpenAI said on August 18 that it had halted part of its frontier reinforcement-learning work until alignment, security and monitoring standards caught up with the capabilities ahead. On Moonshots with Peter Diamandis, Emad Mostaque called the safety constraint real, Alex called the announcement marketing, and Salim Ismail reported back from a visit to OpenAI's offices. OpenAI later disclosed that the largest paused run restarted on August 28.

9 min read