The idea underneath all of this
Bloom's two sigma problem
In 1984 Benjamin Bloom compared students taught in an ordinary classroom against students taught one-to-one with mastery methods. The tutored students scored two standard deviations higher. In plain terms: the average tutored student outperformed 98% of the classroom students, and about 90% of them reached a level only the top fifth of the class reached.
Bloom's own conclusion wasn't "everyone should get a tutor." It was a problem statement, because one-to-one tutoring is out of financial reach for essentially every school system on earth. Forty years later that question is still the central one in education research.
This is the honest frame for the whole argument. Not "technology is exciting." The question is whether something that behaves a little like a patient personal tutor — available at 9 p.m., at near-zero marginal cost — closes any part of a gap we have never been able to close.
The early evidence says: a real part of it. Not two sigma. But a real part.
Before the reasons
The decision in front of you isn't the one it feels like
It feels like: should my child use AI? It is actually: should my child use AI with me, or without me?
Over half of American teens already use chatbots for schoolwork. Barely a third of parents know their teen uses AI at all. A household ban doesn't move the first number. It moves the second one down.
Teen figures: Pew Research Center, 1,458 U.S. teens 13–17, fielded Sept–Oct 2025, ±3.3 pts. Parent awareness: Common Sense Media. Graduating seniors: Handshake, Class of 2026.
Teens were asked where AI beats a human — hiring, driving, diagnosis, songwriting. They said worse on all of them, with one exception. Teaching a skill was the only task where more teens thought AI would win.
The case
Six reasons, each with its receipt
Every claim carries the study it rests on and the caveat that limits it. If a reason doesn't apply to your kid, it doesn't apply — take the ones that do.
It is the closest thing to a personal tutor most families will ever have
A randomized controlled trial in a Harvard physics course put a purpose-built AI tutor against in-class active learning — already one of the strongest teaching methods there is. Students with the tutor learned more, in less time on task, and reported higher motivation.
Caveat that matters: that tutor was engineered around learning science — it withheld answers, asked for the student's reasoning first, and moved in small steps. A raw chatbot pointed at homework is not the thing that was tested.
It works where there is nothing else, at a cost that isn't comparable
The World Bank ran a six-week after-school program in nine public schools in Benin City, Nigeria, using GPT-4 through Microsoft Copilot as a tutor. It produced measurable gains across English and digital skills at $48 per student — among the most cost-effective education interventions ever measured.
Caveat: the widely repeated "two years of schooling in six weeks" line is the authors' extrapolation from a six-week result, not something they measured over two years. The six-week number is the solid one — and still remarkable for the price.
It helps the kid who is behind more than the kid who is ahead
This is the finding that should matter most to anyone who worries AI will widen gaps. In Nigeria the strongest effects were among girls, narrowing a pre-existing achievement gap. In a Stanford trial where 1,000 elementary students were tutored with or without an AI assistant advising the tutor, mastery rose four points overall — but up to nine points for students assigned the least experienced, lowest-rated tutors.
Meanwhile private tutoring in the U.S. runs roughly $40–120 an hour, and $100–200 in New York or Los Angeles. Academic help has been rationed by household income for as long as it has existed. This is the first tool that doesn't ration that way.
It answers the question your kid would never raise their hand to ask
Every teacher knows the student who is lost in week two and says nothing until the exam. The barrier is almost never the difficulty of the question — it's the social cost of asking it in front of thirty people, for the fourth time.
A chatbot has no memory of the previous three times, no facial expression, and no other students waiting. For an anxious kid, or one who has quietly decided they are "bad at maths," that is the whole difference between asking and not asking.
No study proves this one cleanly — it's a mechanism, not a measurement. But it's consistent with the engagement findings across the trials above, and it's the reason teenagers themselves give.
Kids who learn differently get scaffolding they were previously denied
A 2025 systematic review of AI interventions for students with learning disabilities — 11 studies, 3,033 participants — found positive outcomes in every study, with dyslexia the most researched. The practical wins are unglamorous and large: dense text summarized into plain sentences, an assignment broken into ordered steps, a writing process that keeps a student with ADHD moving instead of stalling at the blank page.
Caveat: it's a small literature, and the same review is explicit that results were strongest when the tool sat inside a structured plan — not handed over as standalone self-help.
Using it well is now a literacy, and it is being priced into their future
U.S. job postings requiring AI skills more than doubled in a year, and roughly half are outside IT entirely. What employers describe wanting is not coding — it's the ability to use these tools, judge whether the output is any good, and know when not to use them at all. That last skill is one a kid can only build under supervision, with someone asking "how do you know that's right?"
Yes, but —
The four things parents say next
"It's just cheating with extra steps."
Sometimes it is. 59% of teens say AI cheating happens at their school at least somewhat often. That's real and you shouldn't pretend otherwise.
But the same tool that writes the essay also runs the practice test, marks the answer, and refuses to hand over the solution when told not to. A calculator was a cheating device until schools decided which problems it was allowed in — and then it became infrastructure.
The practical answer isn't vigilance, it's a rule your kid can state: if you can't explain it out loud with the screen closed, you're not finished.
"It'll make them mentally lazy."
This one has evidence behind it. An MIT Media Lab study wired 54 people to EEG while they wrote essays with an LLM, with a search engine, or unaided. The LLM group showed the weakest brain connectivity, and after the first session more than 80% couldn't quote a line from the essay they had just written. The authors call it cognitive debt.
So take it seriously — and notice what it measured. People outsourcing the writing. The Harvard trial measured people being made to reason. Same technology, opposite instruction, opposite result.
The risk is real and specific: it lives in generation, not in explanation. That's a rule you can set, not a reason to abstain.
"It gets things wrong. Confidently."
It does, and it will keep doing so. For a student that's a genuine cost — and, handled right, a genuine curriculum.
A child who has personally caught an AI inventing a citation has learned something about authority and verification no assembly can teach. Chase the source; notice that fluent isn't the same as correct.
"My kid is too young for any of this."
Very possibly, and this case doesn't argue otherwise. Most major services set 13 as their floor and UNESCO recommends the same threshold for education tools. Pediatric guidance is that the youngest children shouldn't interact with chatbots at all.
Everything here is an argument about school-age and teenage students. For a six-year-old, the correct amount is none.
Where the case stops
What this argument does not claim
- Not that any chatbot does this. The results came from tools built or instructed to teach. Point a default chatbot at homework and you're running a different experiment — closer to the MIT one.
- Not that more use is better. Nothing here supports unlimited or unsupervised use, and some of it argues directly against it.
- Not that companions are included. None of this extends to character or companion bots. Common Sense Media recommends no one under 18 use AI companions, and that stands separately.
- Not that the studies are settled. One course at one university, six weeks in nine schools, eleven studies in a young literature. They point in a consistent direction. They are not a finished science.
- Not that it replaces a teacher. The Stanford result is clearest on this: the win came from AI making a human tutor better.
Next
The whole case rests on one condition
Every benefit above holds when the student's brain does the work, and evaporates when it doesn't. That's not a caveat on the argument — it is the argument. So the useful question was never whether to allow it. It's how to set it up so the good version happens by default.
Get Part Two and Part Three
The rest of the guide is free. Enter your email and it opens straight away — no confirmation step, no waiting.
- Part Two — How to set it up. The parental controls that actually exist and what they can't show you, guidance by age, the platforms' real age limits, prompts that make AI teach instead of answer, and a printable family agreement.
- Part Three — What to do at home. The twenty-minute first sitting, a playbook for each kind of homework, scripts for when you catch them, and the two-minute check.
Where this comes from
Sources
- — Benjamin Bloom, "The 2 Sigma Problem," Educational Researcher (1984).
- — Kestin et al., AI tutoring outperforms in-class active learning: an RCT, Scientific Reports (2025).
- — World Bank, From Chalkboards to Chatbots, Policy Research Working Paper 11125 (2025).
- — Stanford NSSA, Tutor CoPilot (2024).
- — AI-Based Interventions for Students with Learning Disabilities: A Systematic Review (2025).
- — Kosmyna et al., Your Brain on ChatGPT, MIT Media Lab (2025). Preprint; n=54.
- — Pew Research Center, How Teens Use and View AI (Feb 2026).
- — Common Sense Media, teen AI companion survey and risk assessment.
- — UNESCO, AI Competency Framework for Students (2024).