Log inSign up

A letter from the founder

The fix for school has existed for forty years. It cost $200 an hour. Not anymore.

The average American 13-year-old reads no better today than one did in 1971.

That sounds like hyperbole. It isn’t. It’s the government’s own test, the Nation’s Report Card, given to 13-year-olds since 1971. In the fifty years since, we’ve rewritten the curriculum a dozen times, put a computer on every desk, and wired every classroom to the internet. Whatever school is optimized for, it isn’t learning. Mostly it’s optimized for time: sit through the lectures, fill out the worksheets, move up in June whether anything stuck or not.

The strange part is that none of this is a mystery. We’ve known what works for forty years.

98th

An education researcher named Benjamin Bloom published this finding in 1984. Forty years later, it still hasn’t been fixed.

He found that As bigger, more careful studies came out, that gap shrank. What never changed is the crown: one-on-one tutoring still beats practically everything else anyone has tried. In 2020 researchers pooled Academics don’t use words like that unless they mean it. Nobody argues about whether tutoring works. They argue about the price. A good tutor costs $50 to $200 an hour.

You can see what happens when families just pay it. There are now private schools built around this kind of individual attention, where kids work at their own pace, get through the academic part of the day in a couple of hours, and end up with some of the best test scores in the country. I think those schools are basically right. But they cost as much as $60,000 a year, in a handful of cities. So the fix for school already exists. It’s just rationed, by zip code and income.

For the first time, the tutoring that cost $200 an hour can cost almost nothing.

Voice AI finally got good enough to hold a real conversation: one you can interrupt, one that follows your reasoning and pushes back. That was the last missing piece. And the first careful trials of tutors engineered this way are starting to land. At Harvard, In Sierra Leone, in classrooms with almost no textbooks.

StudyHall is the start of the online school engineered for learning: the tutoring rich kids have always had, for every kid.

But here’s the part the AI industry won’t say out loud: most AI is making kids worse at school.

Handing a teenager a chatbot that answers homework questions isn’t tutoring. It’s outsourcing. In the best experiment we have, A study tracking thousands of students in China found Assisted performance is not learning. It just feels like it.

This is the whole design problem, and it’s why our tutor argues with you instead of answering you. Every principle below exists to keep the thinking in your head, where it counts.

Core principles

You learn in the argument, not the answer.

Most apps grade what you picked. That’s the weakest feedback there is. You actually learn one level down, at the exact step in your reasoning that went sideways. Researchers have tested this directly: So here at StudyHall, we never just say you’re wrong. We find the step that broke and make you look at it.

You have to say it out loud.

Explaining your own reasoning is one of the most reliable results in all of learning science: across 64 studies, It’s also the only way a tutor can see where you broke. If you never say why, no one, human or AI, can fix it.

A good tutor almost never says you’re wrong.

Researchers at Stanford spent years taping the best human tutors to find out what set them apart. The answer: That sounds like a style choice. It isn’t. It protects the one thing that keeps a teenager learning: the belief that they can. Kill that and no method on earth works.

A careless mistake is never careless.

Miss a minus sign and everyone calls it sloppy. It usually isn’t. Your working memory holds only a few things at once, and The fix isn’t “be more careful.” It’s making the basics automatic, the way knowing your times tables cold frees your head for the actual problem. So we track those misses as their own category and drill the block, not the blame.

Big skills are small blocks, stacked.

Nobody fails “algebra.” They’re missing one block, usually from two years earlier, that everything since has wobbled on. School can’t run that way because the calendar outranks the kid. A tutor can, because the kid is the calendar.

The right question sits just past your reach.

There’s a band between what you can do alone and what you can’t do at all. A psychologist named Lev Vygotsky called it and it’s where learning happens. Too easy teaches nothing. Too hard teaches panic. And here’s the twist from memory research: so left alone, everyone drifts toward practice that feels good and does nothing. A tutor’s job is to hold you in the band, question after question.

You forget. Your tutor doesn’t.

and even though it feels messier. The hard part is the bookkeeping: tracking every question you’ve missed and catching each one right before it fades. No human tutor can hold all that. An AI never loses any of it. Yours can reach back to a question you missed in March, or the way you explained it, and hand it back the week it matters, shuffled in with new ones. It feels harder. That’s the point: the difficulty is what makes it stick.

Your mistakes should teach the next kid.

A wrong answer isn’t random. Across thousands of students, the same few wrong turns show up again and again, and each one has a name and a fix. One UK platform labeled every wrong answer choice with the specific misconception behind it, and Every session on StudyHall feeds a map like that: what tripped you, what untangled it, what worked for the last thousand kids who broke in the same place. The tutor gets smarter about a mistake before you make it.

A tutor worth the name checks in first.

School help waits for you to raise your hand, which is exactly what a struggling teenager never does. When a state university sent students That’s where our tutor is headed: it remembers your last session, notices what you’re about to forget, and messages you the plan: a review built from Tuesday’s mistakes, a quiz timed for the day the memory fades. Over time it learns what you’re into, and we’re careful with that one: wrapping a math problem in a soccer jersey can hide the math. But a tutor who knows you is the whole point of a tutor.

The adults

None of this removes the adults. It promotes them. A teacher standing at the front of a room is spending the day on the one thing software can now do: broadcasting the same explanation at thirty different kids. When software handles the explaining and the drilling, the teacher becomes what the best teachers always wanted to be, the coach in the room: noticing the kid who won’t start, holding the standard, being the person who cares whether you showed up. We’ve already seen a preview:

Parents, same deal. “Did you do your practice” is a fight nobody wins, and the motivation research is blunt about why: So let us hold the standard. You get to be the support. That division is the best trade we can offer a family.

And the hours this saves shouldn’t be poured back into more school. Get the academics tight so there’s time left for the season, the job, the thing you’re building in the garage. School should be the launchpad for a life, not the whole life.

Why the SAT

None of this is really about the SAT. But the SAT is a good place to prove it works, because it ends in a number you can’t argue with. Reading & Writing especially, where every question is why is this the answer?, which is exactly the question a tutor should make you answer out loud.

It also keeps us honest, because the bar is known: which is real but modest. When a prep company promises you 200 points, someone is rounding up. We’d rather run the experiment.

The SAT is the wedge, not the destination. Start with one test. Then every subject. Then everything school never got to: the college essay, the extracurriculars, personal finance, whatever a kid actually wants to learn.

A real tutor for every kid, one they actually want to talk to.

Ryan

Founder, StudyHall

Sources

  1. 1The Nation's Report Card · NCES, federal long-term trend assessment, age 13
  2. 2The 2 Sigma Problem (1984) · Benjamin Bloom, Educational Researcher
  3. 3The Impressive Effects of Tutoring (2020) · Nickow, Oreopoulos & Quan, NBER; meta-analysis of 96 randomized trials
  4. 4AI tutoring vs. active learning (2025) · Kestin et al., Scientific Reports; randomized crossover, 194 Harvard physics students
  5. 5LearnLM in Sierra Leone (2026) · Google DeepMind & partners; preregistered classroom trial, 1,763 students
  6. 6Generative AI can harm learning (2025) · Bastani et al., PNAS; field experiment, ~1,000 high school students
  7. 7The learning penalty (2026) · CEPR Discussion Paper 21577; longitudinal panel of Chinese secondary students (observational)
  8. 8The relative effectiveness of tutoring (2011) · Kurt VanLehn; review across human and computer tutoring studies
  9. 9Inducing self-explanation (2018) · Bisra et al., Educational Psychology Review; meta-analysis of 64 studies
  10. 10What expert tutors do (2002) · Lepper & Woolverton, Stanford; the INSPIRE synthesis of expert human tutors
  11. 11Cognitive load during problem solving (1988) · John Sweller, Cognitive Science
  12. 12Effectiveness of mastery learning (1990) · Kulik, Kulik & Bangert-Drowns; meta-analysis of 108 controlled evaluations
  13. 13Mind in Society (1978) · Lev Vygotsky, Harvard University Press
  14. 14Learning versus performance (2015) · Soderstrom & Bjork, Perspectives on Psychological Science
  15. 15Test-enhanced learning (2006) · Roediger & Karpicke, Psychological Science
  16. 16The science of effective learning (2022) · Carpenter, Pan & Butler, Nature Reviews Psychology
  17. 17Interleaved retrieval practice (2022) · Sana et al.; classroom experiment
  18. 18Expert-supervised AI tutoring (2026) · LearnLM × Eedi; randomized trial, 165 UK students, every AI message expert-reviewed
  19. 19Let's Chat (2026) · Meyer, Page et al., NBER; two preregistered randomized trials, 2,483 students
  20. 20Tutor CoPilot (2024) · Wang et al.; preregistered field trial, 900 tutors, 1,800 students
  21. 21Motivation interventions that work (2024) · Wang et al.; meta-analysis of 137 effects, 9,433 participants
  22. 22Test prep, measured honestly (2025) · Hao et al., Review of Educational Research; meta-analysis of experimental test-prep studies
Try a session→