A five-year-old is lying in the backyard grass, looking up. “Why is the sky blue?” Then: “Why isn’t it blue at night?” Then: “Where does the blue go?”
Every parent knows that run of questions, and what a good answer is worth. Not a fact read off a phone, but a conversation: an answer at the child’s level, a question back, another try in different words, until the child understands and asks the next thing.
When that conversation happens one-to-one, on purpose, we call it tutoring. For almost all of history only a few children got it. Aristotle tutored the young Alexander. Royal households, and later wealthy ones, hired tutors to sit beside their children. Everyone else got a classroom: one teacher, many children, one pace.
In 1984 the educational psychologist Benjamin Bloom put a number on the difference. In the experiments he summarized, students taught by a tutor, alone or in twos and threes, scored about two standard deviations above students taught in a conventional class: the average tutored student did better than 98% of the class.1
One tutor, one student
Bloom’s experiments were small and short, and the randomized trials since have measured a smaller effect. It is still large by the standards of education research: a 2024 review of randomized trials of tutoring found an average gain of about 0.29 standard deviations.2 Bloom called it the two-sigma problem. We knew what worked, and we couldn’t afford to give it to every child, because a tutor was a person and a person’s hour is expensive.
That is no longer true. An AI can now hold the back-and-forth that tutoring is made of: listening, asking, and explaining again a different way. In a World Bank randomized trial in Nigeria, six weeks of after-school sessions with an AI tutor raised secondary students’ English scores by 0.24 standard deviations, and their regular end-of-term English exam by 0.21.3 Children have already found it: American parents say 29% of children aged eight and under, and 39% of those aged five to eight, have used AI to learn something for school.4
And the cost has collapsed. The price of an AI model performing at GPT-3.5’s level fell more than 280-fold between November 2022, when ChatGPT launched, and October 2024.5
The price of GPT-3.5-level AI, per million tokens
By our estimates, the AI and mobile data for ten minutes of conversation a day cost $7–14 a month per child at our first thousand devices, and $4–9 at a hundred thousand.6 Tutoring has never been that cheap.
So we believe that within ten years, tutoring will do what computing did: move from the few to everyone, as an AI per child. The open question isn’t whether every child will have one. It’s what form it takes, and that will be decided by something that doesn’t change.
Chapter 1
What Never Changes
Parents will always want the best for their child’s development
The technology around children changes every generation: radio, television, the games console, the smartphone. What parents want for their children hasn’t changed at all. They want them to grow up healthy, capable and safe, and they want a say in what shapes them. No parent has ever said they don’t mind how their child develops.
That truth has three consequences for any AI a child grows up with.
Children develop in conversation and in the real world
In a 2018 MIT brain-imaging study of 36 children aged four to six, the children who had more back-and-forth conversation with adults had stronger language skills and more activity in Broca’s area, a key language region of the brain. The sheer number of words they heard wasn’t linked to that brain activity, and once family background was taken into account it didn’t predict their language skills either.7 What builds language is the turn-taking.
Screens crowd those turns out. A 2024 study recorded full days of home audio from 220 Australian families as their children grew from one to three. At age three, each extra minute of screen time went with 6.6 fewer words from adults, 4.9 fewer vocalizations from the child and 1.1 fewer back-and-forth exchanges.8
Larger studies point the same way. Among 7,097 Japanese children, those with four or more hours of screen time a day at age one had 4.8 times the odds of a communication delay on a screening test at age two, and the risk rose step by step with screen time.9 Among 2,441 Canadian children, more screen time at two and three predicted lower developmental screening scores at three and five. The effects were small, but they ran one way: lower scores didn’t predict more screen time later.10
Children’s days look nothing like what experts advise. American children aged eight and under average about two and a half hours of screens a day, rising to nearly three and a half hours at five to eight.4 The World Health Organization recommends none for babies and one-year-olds, and no more than an hour a day from two to four, with less being better.11
Average screen time a day, US children
The real world shapes the body too. In a randomized trial in 12 schools in Guangzhou, adding one 40-minute class outdoors to each school day cut new cases of nearsightedness among 1,903 six-year-olds over three years from 39.5% to 30.4%.12 By 2050, about half the world’s population, nearly 4.8 billion people, is projected to be nearsighted.13
Safety is non-negotiable
Attachment researchers call what lets a young child explore a secure base. Children venture out more confidently when a trusted adult is close by, and come back to them when something goes wrong.14
When parents can’t be sure their child is safe, childhood moves indoors. In 1969, 48% of American children aged 5 to 14 usually walked or biked to school. By 2009, 13% did. Even among children living within a mile of school, the share fell from 89% to 35%.15
Walking or biking to school, US children aged 5–14
Distance and traffic are part of it.16 So is worry. In a 2023 national poll, the top reason parents gave for not letting 9- to 11-year-olds spend time without an adult was fear that someone might scare or follow them (54%), even though only 17% thought their neighborhood was unsafe for children on their own.17
Knowing where a child is, and that they can reach you, is what lets a parent say yes to the park, the walk to school and the bike ride. For a young child, safety isn’t the opposite of independence. It’s what makes independence possible.
Parents stay in charge
Parents decide what shapes their child: the school, the friends, the books on the shelf. AI that talks to children is new, and its first versions have given parents reason to be wary. In September 2025 the Federal Trade Commission ordered seven companies, among them Alphabet, Meta, OpenAI, Snap, xAI and Character.AI’s maker, to explain how they measure and monitor the effects of their AI companion chatbots on children and teens.18 Two months later, testers for the consumer group PIRG found that an AI teddy bear, running on GPT-4o by default, would tell them where to find knives, pills and matches, and would drift into sexually explicit conversation. Its maker suspended sales, and OpenAI suspended the developer.19
An AI that a parent lets near their child has to answer to the parent. It should show them what it talked about, follow their instructions, and have nothing to sell.
Chapter 2
What Follows
The future and the truth together decide the shape of the product
Put the inevitable future next to the truth that never changes, and an AI per child can only take one shape. Each consequence of the truth becomes a requirement.
| Because | An AI per child must |
|---|---|
| Tutoring works best one-to-one | Know the child, remember what they’ve learned, and grow with them. |
| Children develop in conversation | Talk and listen. No screen. |
| Children develop in the real world | Go where the child goes: worn, not left on a table. |
| Safety is non-negotiable | Show parents where their child is, carry voice messages both ways, and alert them if the child falls. |
| Parents stay in charge | Listen only when pressed, and glow while listening. Show parents what it taught, and follow their instructions. No ads, no feed, and no recordings kept. |
Chapter 3
The Three Paths
There are three ways to put an AI in front of a child
The Tablet
The obvious path is an AI tutor in an app. It’s easy to build, and the device is already in the house: 72% of the families on our waitlist told us their children have a tablet of their own.20 But a tablet is a screen. It keeps a child indoors, sitting down and looking at glass, instead of talking with the people around them and moving through the world. However good the tutor inside, it fails the first consequence.
The Toy
A talking toy or a smart speaker has no screen, so it passes the first test. But it stays at home on a shelf. It can’t answer the question in the backyard, and it can do nothing for a child on the walk to school. The first wave of AI toys has also shown what happens when a general-purpose chatbot is sewn into a teddy bear.
An AI Per Child
The third path is the simplest. Give each child a small device they wear on the wrist, with no screen, that talks. It goes where they go. It listens only when they press it, and glows while it does. It tells their parents what it taught, follows their instructions, and lets them see where their child is. It has nothing to sell.
It happens to be what we’re building. We call it Halo.
- Talks, with no screenAnswers out loud and asks back, from about age four.
- Listens only when pressedIts light is on whenever it’s listening.
- Goes where the child goesIts own battery and mobile data, with Wi-Fi at home.
- Answers to the parentThe parent app shows what Halo taught and takes the parent’s instructions.
- Keeps them safeLocation a parent can turn on, voice messages and fall alerts.
- Nothing to sellNo ads, no feed, and no recordings of the child’s voice kept.
Chapter 4
Where We Start
One job in one place
Things that end up everywhere start somewhere small. Amazon began with books; Facebook began with one college. The narrowest version of an AI per child is this:
Ours is one job in one place: the weekly spelling list, at the kitchen table, after school.
In the early grades, many children bring home a list of words to practice each week, and someone has to quiz them. Today that someone is a tired parent at five o’clock, or an app on a tablet.
With Halo, a parent adds this week’s words in the Halo app. After school the child presses Halo and asks to practice. Halo says each word and uses it in a sentence, listens as the child spells it out loud, and checks it letter by letter. The words they miss come back tomorrow, and the parent can see which ones are still tricky. Reading aloud and math facts work the same way: out loud, at the table, without a tablet.
Why start here
- It’s what parents asked for first. Of the 758 parents who ranked what they wanted Halo to do, 61% put help with school in their top three, more than any other feature.21
- They already pay for it. A quarter of the families who told us their budget already spend $50 or more a month on their children’s learning.22
- It happens every week, and it’s measured. The Friday spelling test is a score parents already see.
- It’s done out loud anyway. Spelling practice has always been a conversation.
From the kitchen table, Halo follows the child outside: to the “why?” questions in the backyard, then to reading, math and a second language, and then to the walk to school, with location and voice messages for the parent.
Chapter 5
Early Demand
What parents told us in five weeks
We opened a waitlist for Halo on October 7, 2026, and tested it with a small ad campaign on Instagram and Facebook. In five weeks, 1,403 parents joined, 98.7% of them in the United States, at under $2 per signup.23
Parents on the Halo waitlist
805 of those parents then told us about their families, and 758 ranked what they most wanted Halo to do. The children they described were most often between five and nine.
“Which three matter most to you?”
And it works
Halo is a working prototype today. It runs on its own battery, connects over Wi-Fi or mobile data, and starts answering 1.1 to 1.4 seconds after you stop talking.24 Our next build is the first sized for children.
Chapter 6
Growing Up With Halo
From the kitchen table to the walk to school, as the child grows
More to come.
Chapter 7
The Business
A device, a subscription, and costs that fall every year
More to come.
Endnotes
Why tutoring will diffuse through society as an AI per child
- Benjamin S. Bloom, “The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring,” Educational Researcher 13, no. 6 (1984): 4–16, doi:10.3102/0013189X013006004. Bloom summarized two University of Chicago dissertations: four samples in grades 4, 5 and 8, taught probability or cartography over about three weeks. Conventional classes had about 30 students per teacher; “tutoring” meant a tutor for one, two or three students, with regular feedback and corrective work. About 90% of tutored students reached the level of the top 20% of the conventional class. The experiments were small and short and have not been replicated at that size: see Paul T. von Hippel, “Two-Sigma Tutoring: Separating Science Fiction from Science Fact,” Education Next 24, no. 2 (2024): 22–31, educationnext.org. ↩ ↩
- Andre Nickow, Philip Oreopoulos and Vincent Quan, “The Promise of Tutoring for PreK–12 Learning: A Systematic Review and Meta-Analysis of the Experimental Evidence,” American Educational Research Journal 61, no. 1 (2024): 74–107, doi:10.3102/00028312231208687. Pooled effect 0.288 standard deviations (SE 0.029). Effects were largest with teacher or paraprofessional tutors, in earlier grades, at three or more sessions a week, and during the school day. The 2020 working paper, covering 96 randomized trials, reported 0.37 (NBER w27476). ↩
- M. De Simone, F. Tiberti, M. Barron Rodriguez, F. Manolio, W. Mosuro and E. J. Dikoru, “From Chalkboards to Chatbots: Evaluating the Impact of Generative AI on Learning Outcomes in Nigeria,” World Bank Policy Research Working Paper 11125 (2025), hdl.handle.net/10986/43212. First-year senior secondary students (about 15 years old) in nine public schools in Benin City, Edo State; up to twelve 90-minute after-school sessions over six weeks in June and July 2024, using Microsoft Copilot (GPT-4) with a teacher’s guidance. 1,328 students were randomized and 759 sat the final assessment; the results hold after reweighting for uneven attrition. Effects: 0.31 standard deviations on the end-of-program assessment (English, AI knowledge and digital skills), 0.24 on its English section, and 0.21 on the regular end-of-term English exam. ↩
- S. Mann, A. Calvin, A. Lenhart and M. B. Robb, The Common Sense Census: Media Use by Kids Zero to Eight (Common Sense Media, February 2025), commonsensemedia.org. A probability-based survey of 1,578 US parents of children aged 0 to 8, fielded in August 2024. Parents reported that 29% of children aged 0–8, and 39% of those aged 5–8, had used AI to learn school-related material. Average daily screen time: 2 h 27 min for all children aged 0–8; 1 h 03 min under age 2; 2 h 08 min at 2–4; 3 h 28 min at 5–8. ↩ ↩ ↩
- N. Maslej et al., The AI Index 2025 Annual Report (Stanford Institute for Human-Centered AI, April 2025), hai.stanford.edu: the price per million tokens of the cheapest model scoring at least 64.8 on MMLU fell from $20 (November 2022) to $0.07 (Gemini 1.5 Flash-8B, October 2024). Epoch AI finds prices at a fixed capability falling 9x to 900x a year depending on the task, with a median of 50x: B. Cottier, B. Snodin, D. Owen and T. Adamczewski, “LLM inference prices have fallen rapidly but unequally across tasks,” Epoch AI, March 12, 2025, epoch.ai. ↩ ↩
- Our estimate, October 2026, from published prices and supplier quotes: speech recognition, the language model, Halo’s voice, hosting and mobile data for about ten minutes of conversation a day, per child per month, at 1,000 and at 100,000 devices. ↩
- Rachel R. Romeo et al., “Beyond the 30-Million-Word Gap: Children’s Conversational Exposure Is Associated With Language-Related Brain Function,” Psychological Science 29, no. 5 (2018): 700–710, doi:10.1177/0956797617742725. Words and conversational turns were counted from home recordings, and brain activity was measured with fMRI while the children listened to stories. Adult words correlated with verbal scores (r = .36), but turns correlated more strongly (r = .51), and only turns still predicted scores once parents’ education and income were controlled. Turns, not adult words, were linked to activation in Broca’s area, which accounted for about half of the link between turns and verbal scores. The study is cross-sectional. ↩
- Mary E. Brushe et al., “Screen Time and Parent-Child Talk When Children Are Aged 12 to 36 Months,” JAMA Pediatrics 178, no. 4 (2024): 369–375, doi:10.1001/jamapediatrics.2023.6790. 220 families recorded on an average 16-hour day at 12, 18, 24, 30 and 36 months, between 2018 and 2021. Screen time is audible screen sound coded from the recordings. Adjusted estimates at 36 months, with 95% intervals: −6.6 adult words (−11.7 to −1.5), −4.9 child vocalizations (−6.1 to −3.7) and −1.1 turns (−1.4 to −0.8) per minute of screen time. Mean screen time at 36 months was 172 minutes a day. ↩
- I. Takahashi et al., “Screen Time at Age 1 Year and Communication and Problem-Solving Developmental Delay at 2 and 4 Years,” JAMA Pediatrics 177, no. 10 (2023): 1039–1046, doi:10.1001/jamapediatrics.2023.3057. 7,097 mother–child pairs in Japan’s Tohoku Medical Megabank cohort. A delay is an ASQ-3 screening score more than two standard deviations below the mean, not a diagnosis. Adjusted odds ratios for communication delay at age 2, against under an hour a day: 1.61 (1–2 hours), 2.04 (2–4 hours) and 4.78 (4 hours or more; 95% CI 3.24–7.06). At age 4, the ratio for 4 hours or more was 2.68. ↩
- Sheri Madigan et al., “Association Between Screen Time and Children’s Performance on a Developmental Screening Test,” JAMA Pediatrics 173, no. 3 (2019): 244–250, corrected in 173, no. 5: 501–502, doi:10.1001/jamapediatrics.2018.5056. 2,441 mother–child pairs in Calgary, in a random-intercept cross-lagged panel model. Screen time at 24 months predicted ASQ-3 scores at 36 months (β = −0.06), and screen time at 36 months predicted scores at 60 months (β = −0.08). “The obverse association was not observed.” ↩
- World Health Organization, Guidelines on Physical Activity, Sedentary Behaviour and Sleep for Children under 5 Years of Age (Geneva: WHO, 2019), iris.who.int, and its press release of April 24, 2019. Under 1: screen time is not recommended. For 1-year-olds, sedentary screen time is not recommended; at 2, and at 3 to 4, no more than an hour, and less is better. ↩ ↩
- Mingguang He et al., “Effect of Time Spent Outdoors at School on the Development of Myopia Among Children in China: A Randomized Clinical Trial,” JAMA 314, no. 11 (2015): 1142–1148, doi:10.1001/jama.2015.10803. A cluster-randomized trial in 12 primary schools (six intervention, six control), 2010–2013; 1,903 first-graders with a mean age of 6.6. Parents in the intervention schools were also encouraged to take their children outdoors after school and at weekends. Three-year incidence among children not nearsighted at the start: 30.4% (259 of 853) against 39.5% (287 of 726). ↩
- Brien A. Holden et al., “Global Prevalence of Myopia and High Myopia and Temporal Trends from 2000 through 2050,” Ophthalmology 123, no. 5 (2016): 1036–1042, doi:10.1016/j.ophtha.2016.01.006. A projection from 145 studies of 2.1 million people: 4,758 million people (49.8%) nearsighted by 2050, up from 1,406 million (22.9%) in 2000. ↩
- Mary D. Salter Ainsworth, Mary C. Blehar, Everett Waters and Sally Wall, Patterns of Attachment: A Psychological Study of the Strange Situation (Hillsdale, NJ: Lawrence Erlbaum, 1978); John Bowlby, A Secure Base: Parent-Child Attachment and Healthy Human Development (New York: Basic Books, 1988). ↩
- N. C. McDonald, A. L. Brown, L. M. Marchetti and M. S. Pedroso, “U.S. School Travel, 2009: An Assessment of Trends,” American Journal of Preventive Medicine 41, no. 2 (2011): 146–151, doi:10.1016/j.amepre.2011.04.006: 47.7% of students aged 5 to 14 usually walked or biked to school in 1969 and 12.7% in 2009, whatever the distance. Within a mile: National Center for Safe Routes to School, How Children Get to School: School Travel Patterns from 1969 to 2009 (November 2011), pedbikeinfo.org. ↩ ↩
- Centers for Disease Control and Prevention, “Barriers to Children Walking to or from School — United States, 2004,” MMWR 54, no. 38 (2005): 949–952, cdc.gov. Parents of 1,588 children aged 5 to 18 named distance (61.5%), traffic danger (30.4%), weather (18.6%) and crime (11.7%). ↩
- C.S. Mott Children’s Hospital National Poll on Children’s Health, “Promoting Children’s Independence: What Parents Say vs. Do,” Mott Poll Report 44, no. 2 (University of Michigan, October 16, 2023), mottpoll.org. A nationally representative sample of 1,044 parents of children aged 5 to 11, surveyed in August 2023. ↩
- Federal Trade Commission, “FTC Launches Inquiry into AI Chatbots Acting as Companions,” press release, September 11, 2025, ftc.gov. Orders under Section 6(b), a study rather than an enforcement action, went to Alphabet, Character Technologies, Instagram, Meta Platforms, OpenAI, Snap and X.AI. ↩
- U.S. PIRG Education Fund, Trouble in Toyland 2025 (November 13, 2025), pirg.org. FoloToy’s Kumma ran on OpenAI’s GPT-4o by default. On November 14, PIRG reported that FoloToy had temporarily suspended sales of all its products for a safety audit, and quoted OpenAI: “we’ve suspended this developer for violating our policies” (pirg.org). Kumma later returned to sale. ↩
- The Halo waitlist’s family survey, October 7 to November 10, 2026: 573 of the 796 families who answered “Do your kids have a device of their own today?” chose Tablet. ↩
- The Halo waitlist’s family survey, October 7 to November 10, 2026: 758 parents chose the three of nine features that mattered most to them, worded as shown in Fig. 7. 466 put help with school in their top three, and 169 (22%) ranked it first. ↩ ↩
- The same survey: 158 of the 629 families who answered “About how much do you spend each month on your kids’ learning today, like apps, classes or tutoring?” chose $50–99 or $100 or more. ↩
- Signups to the waitlist at saint.app, October 7 to November 10, 2026: 1,403 in all, 1,385 of them in the United States. Almost all came from a small ad test on Instagram and Facebook; the cost per signup across the test was under $2. ↩ ↩
- Measured on our bench in October 2026, over Wi-Fi: 1.1 to 1.4 seconds from the end of the speaker’s words to Halo’s first audio. ↩