What kids get right about AI that most engineers don't

How kids use AI turns out to be far more disciplined than how most adult teams use it. MIT Technology Review interviewed children aged 10 to 18 across the US and Canada and published their answers on 13 August 2026: How kids feel about AI, in their own words.

I opened it expecting a cheating story. What I found instead was a set of delegation rules cleaner than anything I have seen in a company AI policy, including my own.

---

📊 The data says boring, not apocalyptic

The reporters anchor their interviews to a Pew Research Center survey of US teens published in February 2026. The usage split is not what the panic coverage suggests:

What US teens use chatbots for	Share
Searching for information	57%
Help with schoolwork	54%
Entertainment	47%
Emotional support	12%

The reporters' own summary of that spread: teens are over four times more likely to be using AI in innocuous ways than in potentially harmful ones.

Two things follow from that table. First, the dominant use is search, not essay generation. Second, the emotional-support number that drives most of the alarmed op-eds is the smallest column on the chart.

Key takeaway: The loudest risk in the AI-and-children debate is the rarest behaviour in the data. Policy written against the 12% will fail the 57%.

---

🧭 They drew a line adults still haven't

The interviews are where it gets interesting. The kids quoted are not uniformly enthusiastic and not uniformly against. They have specific, defensible boundaries.

Sylvia, 10, on maths: "I might not need to take a math test by myself. Like AI won't tell me the answer, but it will help me." That is the distinction between a tutor and an answer key, articulated by a ten-year-old, and it is a distinction plenty of adult "AI productivity" pitches still refuse to make.

Wesley, 14: "I want to have my creativity be my own and not have AI influence it." Not a ban. A carve-out.

Evelyn, 13, on medicine: "I never want it to be my doctor."

Winter, 17, is blunter: "AI isn't the solution to our problems. I'm afraid it's going to be the end of creativity and critical thinking." Hazel, 17, described how AI makes her feel in one word: "Angry." At least one of the kids profiled rejects the technology outright on environmental grounds.

Set against that, Remy, 16, offers the most useful framing in the piece: "I think AI right now is sort of like the first car or the first airplane. It's interesting but crude."

Sort the quotes and a policy falls out on its own:

Task	Kids' verdict	Underlying rule
Explaining a maths problem	Delegate	Learning is the output, not the answer
Writing fiction, drawing, music	Keep	Ownership of the work is the point
Diagnosing an illness	Keep	Stakes too high for a guessed answer
Looking things up	Delegate	Low stakes, verifiable
Deciding what to build	Keep	Judgement is the skill being trained

That last row is Krishiv, 17, who reads the money rather than the models: "With billions of dollars of investment being poured into agentic AI, it's clear that initiative, judgment, and self-direction are becoming some of the most important skills you can build." If agents handle execution, the scarce input becomes knowing what is worth executing. That is a sharper read on the labour market than most career advice I have seen this year.

---

🇱🇰 Why this matters for a Sri Lankan classroom

Every Sri Lankan school, university department and small dev shop is currently writing some version of an AI rule, usually informally and usually as a ban. The Pew split is a warning about how those rules get written.

• A blanket ban targets the 12% and punishes the 57%.
• Detection-first policies push students toward hiding usage, which removes the teacher's only chance to correct bad usage.
• A disclosure rule costs nothing to enforce and produces information you can actually act on.

There is also a local access point that the US-and-Canada framing misses. For a student here, the free tier is not a convenience, it is the entire product. Nobody in an O/L or A/L class is paying USD 20 a month. That means:

1. The model matters less than the habit. A free-tier model that a student interrogates is worth more than a paid one they copy from.
2. Rate limits are a feature. Hitting a daily cap forces a student to decide which question is worth asking. That is Krishiv's point in practice.
3. Offline-first still wins. Anything you can do without a token round-trip is faster, cheaper and available during a power cut.

If you are choosing which free tier to standardise on for a class or a small team, our AI model comparison tool lays out context windows and pricing side by side, and the AI token counter tells you how far a free quota actually stretches before you commit to one.

---

🛠️ Writing the boundary down

The kids in this piece did something most teams skip: they decided the boundary before the deadline pressure hit. Here is the version I would put on a whiteboard, taken directly from their reasoning rather than invented for this post.

1. Name the output you are being graded on. If it is the artefact, AI can help. If it is your ability to produce the artefact, it cannot.
2. Never delegate a decision you cannot check. Evelyn's rule. Ask it about symptoms, do not let it be the doctor.
3. Protect one thing entirely. Wesley's rule. Pick the work whose value is that it came from you, and keep AI out of it completely.
4. Treat every output as a first draft from a crude machine. Remy's rule. Early aeroplanes flew. They also crashed.
5. Write the rule down before you need it. A boundary you improvise at 2am under deadline is not a boundary.

Bottom line: Nothing on that list requires a new tool, a subscription, or a school-wide detection system. It requires deciding in advance what you are unwilling to hand over.

---

💡 What this means for you

If you teach, stop writing your AI policy against the cheating scenario and start writing it against the search scenario, because that is what 57% of your students are actually doing. Ask for disclosure, not abstinence, and spend your effort on teaching verification.

If you are a student, the useful takeaway is that the kids in this piece with the strongest opinions were not the ones who used AI most or least. They were the ones who had thought about where the line goes. Copy that habit before you copy anyone's prompt.

If you build things for a living, Krishiv's observation is the one to sit with: when execution gets cheap, the premium moves to judgement about what to execute. It took a seventeen-year-old to say that plainly.

I went in braced for a story about kids gaming the system. I came out with a group of ten to eighteen year olds holding clearer rules than the industry selling to them. Worth reading in full at the source.