Education Is Broken. AI Won't Fix It Alone
"It does feel like education is fundamentally broken," Burke Holland wrote after Sam Altman said he was disappointed by how little education has changed since AI took off.
They're both right. Education is broken, and it'll get worse if schools keep teaching and grading as if AI doesn't exist. Students will outsource more of their thinking, and teachers will spend their time policing cheating instead of teaching.
But I think Sam was too optimistic about how quickly schools would change. Most of what's broken in education has little to do with technology, so AI alone won't fix it.
Schools reward passing
In theory, the education system exists to help people learn. In practice, it rewards passing, whether it's a class, the SAT, or the bar exam. In Brazil, where I live, the big ones are ENEM (our national university entrance exam) and concursos (exams for government jobs).
Once a test matters enough, everything bends around it. Governments judge schools by their scores, schools advertise their results, and parents pay for test prep. Students figure this out early, and their main question becomes "Will this be on the test?"
They memorize physics formulas, pass the exam, and forget most of it a month later. They got the grade, but they can't use physics to think through a real problem. I was one of those students. I had good grades and never felt like I was learning anything.
From the outside, it's easy to say "just let people learn with AI" or "just build a better learning app." But most kids don't teach themselves. They go to school, where grades and exams decide what counts. If you want to fix learning, you have to deal with schools too.
Everyone is stuck
People in tech often blame the people inside schools, as if teachers and principals were incompetent or resistant to change. I don't think that's true. Most teachers know the system is broken. So do many principals, parents, and students.
Knowing isn't the same as being able to change it. A teacher can't redesign education from one classroom. A principal has a fixed budget, regulations, and parents to answer to. Parents want their kids to learn, but they also want them to get into a good university, and that depends on exams.
So everyone does the rational thing and focuses on the test, even when they know it isn't what matters.
Why EdTech doesn't fix it
Many EdTech startups say they want to change education, but the money is in helping people succeed inside the current system: exam prep for students, compliance tools for schools, and apps that generate quizzes and lesson plans for teachers.
Those products can be useful, and some are great businesses. They just make the current system more efficient without changing what it rewards.
Y Combinator tells founders to "make something people want." In education, that's exactly why nothing changes. People want to pass the next exam. What they need is to actually learn, and it's much harder to build a business around that.
Tech also tends to treat schools like any other market: build a better product, make it cheaper, add AI, grow fast. But if a productivity app fails, you switch apps. If an education experiment fails, a kid can lose a year of preparation for an exam that shapes their future. It makes sense that parents and schools are careful.
A better way to learn isn't enough. It also has to work for a teacher with 35 students, a required curriculum, exam pressure, and no extra budget.
Zoonk isn't immune
I've been making learning tools for years, and now I'm building Zoonk, so I'm not saying any of this from the outside.
My long-term vision looks more like a game than test prep. Something like an RPG, or No Man's Sky for learning, where you learn physics because you're building something, history because you're trying to understand a conflict, and math because the world gives you constraints you have to reason through.
When I talk to users, though, they rarely ask for that. They ask for help passing ENEM, a concurso, or a professional exam. If you have an exam in three months, you don't care about my philosophy of education.
They're not wrong. I've had to step back from my original vision and build more of what people want now, so I can build what I think they need later.
Why AI alone won't fix it
AI can be a great tutor. It can explain things at any level and help students build projects that used to be out of reach. But it amplifies whatever the system already rewards. When the assignment is an essay for a grade, students use AI to write the essay. When schools are judged by test scores, they use AI to raise test scores.
Eventually, I think AI will force schools to change. If AI can do a traditional assignment in seconds, that assignment stops meaning anything.
It just won't happen in one or two years, like Sam thought. It may take decades. Education is tied to childhood, families, public policy, and people's chances in life, which makes it one of the hardest systems in the world to change.
What might work
I don't have a simple answer, and I'd be suspicious of anyone who says they do. But I do think we teach things in the wrong order.
Most classes work like this: here are ten concepts, memorize them, and maybe one day you'll see why they matter. I'd rather start with a problem and ask what you need to understand to solve it. That makes learning concrete, and it makes thinking much harder to fake, even with AI.
We can't get there by ignoring the system we have, though. People buy what they already understand: grades, certificates, and "this will help your child pass." Learning how to think matters more, but next month's exam is more urgent.
The realistic path, I think, is to build tools that help people with the system as it is today while slowly bringing better learning into it. For example, exam prep that makes students explain and apply ideas instead of memorizing them, or tools that save teachers time instead of giving them more work.
That takes humility from people in tech, me included. We need to spend time with teachers, principals, students, parents, and policymakers, and understand their constraints before deciding what they need.
This will be slower than any of us would like, but I don't see another way.