Beyond the Answer
As AI becomes both coder and tutor, teachers ask what students still need to learn and what role they should play.
According to Satya Nadella, artificial intelligence (AI) writes 20 to 30 percent of the code in Microsoft’s repositories, with some already passing the halfway point. Sundar Pichai puts Google’s figure at roughly a quarter to a third. At my own startup with 15 engineers, no one handwrites code anymore. Two years ago, any of these claims would have sounded like science fiction.
Linus Torvalds, who created Linux, the operating system that powers every Android phone and most of the world’s servers, has pointed out that compilers write 100 percent of everyone’s code and nobody boasts about that. He is right that this has happened before. The difference is that a compiler is reliably correct, and a large language model is not.
If professional software is increasingly written by machines, what do schools have left to teach?
I put that question to four people from very different disciplines who do this for a living: in schools, in coding boot camps, in research, and in the creation of Japanese school textbooks. Surprisingly, there was a good deal of consensus.
That Which Survives
Simon Guest, a computer science educator and former chief technology officer of Code.org, the nonprofit behind the Hour of Code and the largest provider of free computer science curriculum in US schools, teaches AI at both the high school and undergraduate levels. He explains what he thinks endures using house construction as a metaphor.
“The student could ask an AI to ‘build me a house’ in one shot and simply accept whatever it produces,” he said. “Or they could design the house first, deciding on the structure and what each room needs to accomplish, and then use AI to help build each part. Using the second approach, AI may still do much of the construction, but the student is the ‘builder’ of the system.”
Stefania Druga holds a PhD in AI tools for creative coding, created the Cognimates platform, and is currently a staff research scientist at Sakana AI in Tokyo. She arrived at the same place after a decade of studying children learning alongside these tools. When agents write more of the code, she said, “knowing how to decompose a problem, specify what you want, and evaluate what comes back matters more, not less.”
Kani Munidasa cofounded Code Chrysalis in Tokyo in 2017, initially helping people transition into software engineering, and today works primarily with companies to develop their in-house software development and organizational capabilities. He put it in the language of hiring.
“As building software becomes faster and cheaper, the scarce capability is increasingly not simply, ‘Can you code?’ but ‘Can you decide what to build, and can you build it effectively as a team?’ That’s an organizational capability,” he said.
This is not new. I have hired senior engineers for 15 years by asking why they built the thing they are describing—what prompted the migration, who decided, and whether they understood the reasoning. If somebody simply told them to do it, the interview is largely over. What’s new is that I now ask the same of juniors. The judgment that used to earn a promotion is becoming the price of entry.
AI as Teacher
Let’s do a little philosophical exercise. AI could be the perfect teacher. It is infinitely patient. It never tires of the same question asked a fourth time. It explains at whatever level you need, in whatever language you speak, and it will do it at two in the morning.
Munidasa does not hedge: “I think AI will eventually do a lot of what we traditionally think of as teaching, and probably do some of it better than we can,” he said. “It can already explain concepts, answer questions endlessly, adapt to the individual, review code, generate exercises, and give immediate feedback. We shouldn’t pretend those capabilities are somehow protected from AI.”
Of the years Code Chrysalis spent running consumer boot camps, he said: “People weren’t paying us for access to information. Even when we started, you could learn JavaScript, algorithms, databases, or almost anything else online for free. What they were paying for was an environment designed to transform how they learned and worked.”
The evidence for that predates AI. When Harvard put CS50, its introductory computer science course, online for free, 150,349 people registered in the first year. Just 1,388 of them finished. On campus that same year, 703 of 706 students completed the identical course, taught by the same instructor from the same material.
Guest does not think the difference is pedagogical.
“When a student signs up for an in-person course, they’re accountable to another person,” he said. “If they don’t show up or complete the work, there’s a teacher or professor who notices.” Online, most of that disappears. Anyone who has joined a gym in January and stopped going by spring already knows how that story ends.
“I am much more optimistic about AI as a ‘teaching assistant’ rather than a teacher—an AI that can provide immediate, personalized help, while a human teacher remains responsible for the relationship, motivation, and accountability,” he said.
“AI may become an extraordinary teacher, but education still has to develop the human being.”
Who It Fails
“The people who can benefit most from self-directed AI learning are often the people who have already learned how to learn without it,” Munidasa noted.
“The danger,” he said, “is when you don’t yet know what you don’t know. AI can give you an answer that feels complete, and even produce something that works, without you understanding why it works.”
Druga explained it this way: “An answer machine optimizes time to answer. A teaching tool optimizes what you can do after it leaves the room.”
In her most recent study, North American participants already had AI at home, while several from other regions had never used it. “The tool amplifies whatever head start a child already has,” she said. “That is why access has to be designed, not assumed.”
The Adult in the Room
Japan has already tested whether access on its own is enough. In 2018, Japanese students spent less classroom time on digital devices than students in any other country in the Organisation for Economic Co-operation and Development. The government’s response, announced the following December, was the Global and Innovation Gateway for All, or GIGA, School Program, which put a networked device in the hands of every public school student, with 1,769 local governments, or 97.6 percent, having completed planned delivery by the end of March 2021.
Distribution was never the hard part.
The education ministry’s own reviews now describe uneven use between regions and between schools.
The constraint is the adult at the front of the room. A survey by the Ministry of Education, Culture, Sports, Science and Technology (MEXT) for FY2022 found that 23 percent of junior high school teachers responsible for technology were teaching under a temporary license or an out-of-field arrangement.
That is easy to read as a competence problem, but I do not think it is one. “The gap is teacher confidence, and tools built for learning rather than for answers,” Druga said.
Ayana Murakami, who works on the digital content in Japan’s authorized junior high school textbooks and is a software engineer and PhD candidate at Ochanomizu University, describes a second structural problem. “The national curriculum is revised roughly once a decade, and junior high school textbooks follow a four-year cycle, while AI can change dramatically within months.”
The ministry appears to agree. Its budget documents describe a new junior high subject under consideration, provisionally called Information and Technology, and say work on teaching materials will begin without waiting for the next curriculum revision to take effect.
Murakami is not pessimistic about schools. Asked whether compulsory education is better placed than self-directed learning to help students understand the potential, risks, and use of AI, she said: “AI makes it easier to learn on your own, but it can still be hard to study things you are not interested in. Schools provide a common foundation by giving students a structured learning environment.”
What’s Left
Having granted AI that much, Munidasa draws a conclusion that runs against his own interest. “But if that’s all a boot camp does, then yes, I think AI should replace it.”
What survives is the part that was never about information. “You don’t learn teamwork by reading about teamwork. You learn it by having to build something with another human being when you don’t agree.”
Asked what she would fund in Japan if she could fund one thing, Druga did not name a tool for learners. She chose “AI tools that assume a teacher is present and make her better.” Every person I spoke to for this piece said a version of the same thing.
A country where every child has a device, a teacher, and a teacher confident with both would be the first of its kind. Japan already has two of those three.
The third is the hardest. Teachers need confidence with these tools, as well as AI built for learning rather than for answers. Current AI products will not do that out of the box. They are built to answer. So tools that enhance and support the learning process have to be built.
MEXT is reconsidering the curriculum now, and the obvious move is to add AI to the syllabus. But nothing in a syllabus reaches a student except through a teacher. What we do know is that accountability matters more than it did, and that what is worth teaching is the reasoning rather than the result. As Munidasa put it: “AI may become an extraordinary teacher, but education still has to develop the human being.”
