AI Will Be Part of Childhood. The Question Is Who Controls the Experience.
TL;DR
Daniel Susskind makes a strong case against trying to protect children from AI by simply banning it. Children need foundational skills, but they also need to learn how to work with AI because it will be embedded in the world they inherit. His “teach both, test both” framework—teach students to work both with and without AI—is particularly compelling. (The Guardian)
There is another piece we need alongside AI literacy: AI designed specifically for children cannot behave like an unrestricted general-purpose chatbot. A child can learn to question AI, but adults still have responsibility for what the AI is permitted to do, how it responds to emotional situations, when it challenges rather than agrees, and when it hands responsibility back to a human.
That makes education a powerful VERN OS use case. The goal should be to give children access to increasingly capable intelligence inside behavioral boundaries established by parents, educators and institutions.
Blog
Our Kids Need to Learn AI. AI Also Needs Rules for Our Kids.
Daniel Susskind asks a question I suspect millions of parents are quietly asking: What should my children do about AI?
He has an unusual perspective on it. Susskind has studied the impact of artificial intelligence on work and society for 15 years, but he is also raising three young children. In The Guardian, he argues that trying to shield children from AI misunderstands the world we’re preparing them to enter. Schools can prohibit the technology today, but these children are going to spend their adult lives surrounded by systems vastly more capable than those available now. (The Guardian)
I agree with his basic premise. We should teach children to use AI.
But there is another half of this conversation that deserves just as much attention: What should we require of the AI we’re asking our children to use?
Teach Both, Test Both
One of Susskind’s strongest ideas comes from an earlier technological disruption: the calculator.
When calculators became widely available, educators worried that students would lose fundamental mathematical skills. The eventual solution wasn’t prohibition. Schools taught students to calculate both with and without them. Susskind proposes applying the same principle to AI across education: “teach both, test both.” (The Guardian)
That makes enormous sense.
Children still need to read, write, calculate, reason and develop their own judgment. Those capabilities become even more valuable when AI can produce convincing answers almost instantly, because students need enough independent knowledge to recognize when the machine is wrong.
At the same time, refusing to teach them how to use one of the most consequential technologies of their generation would leave them equally unprepared.
The skill we’re really trying to develop is judgment: When should I use AI? What should I delegate? How do I know whether the answer is good? When should I challenge it? When should I ignore it entirely?
That requires experience.
But Children Shouldn’t Be Responsible for AI Safety
There is a danger in taking the AI-literacy argument too far.
We cannot teach a 10-year-old critical thinking and then transfer responsibility for the behavior of a probabilistic AI system onto the child.
Recent research has documented children forming emotional relationships with chatbots, encountering inappropriate conversations and relying on AI for increasingly personal guidance. General-purpose models can hallucinate, become sycophantic, misread emotional situations and behave differently even when given similar inputs.
Children can and should learn skepticism. The adults deploying these systems still have an obligation to establish the boundaries.
An AI tutor should be allowed to explain fractions in 20 different ways if that’s what helps a child understand them. It should not be equally free to determine how it handles a conversation about self-harm, bullying, abuse, sexuality, loneliness or a request to keep something secret from a parent.
Those are fundamentally different kinds of interactions.
The Best AI Tutor May Know When Not to Answer
Susskind sees enormous potential in AI tutoring, and he’s right. A human tutor cannot be available to every student at every moment. AI can patiently explain something again, change the explanation, adjust the difficulty and personalize instruction at a scale that would be economically impossible with human tutors alone. (The Guardian)
But personalization creates responsibility too.
Imagine an AI tutor working with a student who repeatedly fails a problem. The words alone might suggest another explanation is needed. But if frustration is escalating, continuing to push harder may be exactly the wrong response.
Perhaps the AI should simplify the problem. Perhaps it should encourage the student. Perhaps it should suggest a break. Perhaps it should alert a teacher. The appropriate behavior depends on what is happening with the human being, not simply whether the last answer was correct.
This is where emotional intelligence becomes particularly important in educational AI.
A system that can recognize emotional signals throughout the interaction can respond differently as those conditions change. But even recognition isn’t enough. The institution still needs control over what the AI is permitted to do with that information.
AI for Children Needs an Operating System
This is one of the reasons we built VERN OS.
The underlying model can provide extraordinary intelligence: answering questions, generating lessons, adapting explanations, creating simulations and helping students explore ideas that previously would have required individual instruction.
VERN OS provides deterministic runtime control around that intelligence.
A school, educational technology company or parent can establish behavioral requirements governing how the AI operates. Those requirements can remain consistent even as the underlying models change.
That separation becomes especially important in education because the most capable model and the most appropriate behavior for a child are two different engineering questions.
We should be able to give children access to extraordinary intelligence without giving that intelligence unrestricted discretion over the relationship.
Prepare Children to Supervise AI
Susskind argues that trying to “future-proof” children by identifying particular skills AI will never master is probably futile. Coding was once presented as precisely that kind of future-proof skill. Now AI is exceptionally good at coding. (The Guardian)
There may be a more durable capability worth teaching.
Children need to learn how to supervise machines.
They should know how to interrogate an answer, compare sources, recognize uncertainty, establish an objective, evaluate whether AI achieved it and retain the confidence to overrule the machine when their own judgment tells them something is wrong.
Those skills will matter whether today’s leading model comes from OpenAI, Anthropic, Google or a company that hasn’t been founded yet.
And there is an important symmetry here.
We should teach children how to govern their use of AI while simultaneously building AI systems whose behavior adults can govern.
That creates a much healthier relationship than either extreme: banning a technology children will inevitably encounter or handing them an extraordinarily persuasive probabilistic system and hoping they know what to do with it.
Susskind writes that we’re letting children down if we prepare them for the world we grew up in rather than the one they will actually inhabit. (The Guardian)
He’s right.
Preparing them for that world means teaching them to use powerful AI. Our responsibility is making sure the AI they encounter is worthy of that trust.
