A curriculum rewrite takes five years. AI does not wait.

Johan Falk joins Nathan Labenz from Stockholm with a résumé that already spans several lives: physicist, science journalist, web developer, classroom teacher, then the National Agency for Education, where he spent years trying to improve mathematics teaching across Sweden. ChatGPT arrived and, in his telling, turned the place upside down. He stood up an internal team on AI in education, ran it for eighteen months, and left the agency to look at AI as a societal problem rather than a school-tool problem. He is now hunting for a way to have a positive effect on AI risk while running Graspable AI, a set of short videos for teachers.

Labenz has his own reason for the conversation. He is about to keynote for 500 public-school administrators in Michigan. His children are in first grade and preschool in the neighborhood Montessori. In another timeline he can imagine becoming a tiger dad; in this one, with the job market and the structure of society under radical uncertainty, he has mostly been letting the kids be kids. Both men spend their working lives tracking AI. Both have more questions about school than answers. What Falk offers is less a five-year plan than a way to sort the questions.

The tempo mismatch is the thing he cannot get past. Changing a curriculum takes years; five years is rather quick. Five years in the AI world is massive change. His own children are seven and ten and will be in school for roughly a decade more. He has no idea what society will look like when they finish. Everyone still needs to read, write, understand themselves, and cooperate. Quadratic equations? New languages? Useful, maybe. Necessary? Probably not. The educational system's most important upgrade, he keeps saying, is agility, and he does not know how to buy it.

Labenz notes the tacit American assumption: school exists so you can enter the workforce, make a living, and contribute to GDP. Falk says every country, more or less, mixes three purposes. One is supplying competence for the labor market. One is fostering citizens — the knowledge and the values a functional, responsible adult is supposed to have. The third is growing as a person: arts and other things that are not productive in any functional sense, but good for you, and therefore good for the society around you.

In the United States, Labenz observes, the last two are a brawl in every district — what is a good citizen, toward whose measure of a good person — while employability at least has a ground truth: you can get a job or you cannot. Falk's impression, never having lived in the U.S., is that America runs on competition in a way Sweden does not. Sweden is not communist, he says, but it is social-democratic: high taxes, healthcare that is free to a point, even higher education free. A stronger sense of common good lowers the competitive temperature of school.

AI, he thinks, puts the first purpose in question. Assign whatever probability you like to a radically different future — 2 percent or 99 — but if the path is universal basic income, competence for the labor market becomes basically irrelevant. Schools should then spend more time on how a person stays healthy and well, a science that exists and is barely taught. Reorganizing today for a future nobody can see is its own risk. Labenz offers a smaller rhyme from startups: a feature can become obsolete in the months between kickoff and launch, and the reluctance to rip it out is emotional as much as technical. School inertia, they agree, is an order of magnitude worse, maybe two.

Asked who is keeping up, Falk says basically nobody, then names the relative bright spots. South Korea and Singapore push from the top down; Singapore also has education-data standards that make it easier to plug AI in. The United States has Khanmigo, which Khan Academy has talked about taking to a million students and teachers, with Microsoft money in the mix. China has been putting AI in the curriculum for something like ten, fifteen, or twenty years, and has a Khanmigo-like product called Squirrel. Estonia is small, agile, and data-rich, and has rolled out AI support for learning. Language-based models, not older neural nets on education data, are the thing moving fast enough to matter. The studies arrive two years late.

Nationally he wants four moves: a top-down strategic instruction that this is something to do; resources and training attached to it; clear rules, such as Korea's line that learning data is not for other purposes; and boldness. Non-action is itself a great risk. You miss the upside and you can land on the wrong side of the digital divide. Doing something and improving is better than waiting until you know.

To keep the sprawl from becoming "keeping up with all of society," he uses a cube. Four faces are visible; two are still unknown, and he wants that remembered. Face one: AI for learning — Khanmigo, study buddies. Face two: teachers using AI outside the classroom, the Excel-but-stronger job of getting work done. Face three, strangely quiet: teaching about AI, the competence kids have needed for almost three years of chatbots. Face four: system-level effects — what a school is, what a teacher is, whether books survive. The first two are what everyone reaches for. Today, he says, they are the least important. Teaching about AI is the neglected emergency.

The learning-tool research does not yet license a mandate. Some studies say you can learn twice as much in half the time; some say you learn less. Dig in and one is 25 adults in Nigeria, another is 18 people told to use a chatbot to write an essay. Useful, not portable to a Swedish middle school. So he would not tell every teacher to use AI as a learning tool. Interested teachers, in situations they judge, yes. Labenz names the reverse risk: if officials say it is unproven, most people hear permission to ignore it. Falk accepts the hedge. Teaching is bizarrely hard — on the order of 2,000 decisions a day — and AI is one tool among many, including sending a restless kid out to run for ten minutes. Personalized learning is the biggest possible win and also a promise schools have heard since cassette tapes. It works for students who are already active learners. That, he thinks, is becoming the scarce skill.

Labenz, who finds AI an "indispensable advantage" when a biology paper has holes in his knowledge, still wants humility about forcing it on children. Sal Khan has said some kids take to Khanmigo and sprint; others freeze, and their teachers report they freeze in class too — they cannot say what they need. Falk wants seven- and ten-year-olds to be able to use chatbots, believes conversation with an expert on anything could accelerate learning in a way only royalty used to get, and thinks that makes *wanting* to learn the bottleneck. Teachers shift toward motivation and ignition; subject knowledge can follow. An expert tutor at $20 or $2 a month could matter in poor countries as well as rich ones. Alpha School's two-hour-learning model, which Mackenzie Price told Labenz uses no chatbots, is inspiring and, as far as Falk knows, unmatched as an organized school. What he has seen are wizardly parents and a friend who ships AI musicals, dyslexia tools, and Parkinson's handwriting readers, all coded by models.

Classroom policy is almost beside the point. Ban AI in the room and kids still use it at home, some to learn more and some to fool themselves that they are learning. That self-deception is, he says, one of the greater risks in AI and education. Fall behind, accumulate problems, miss the degree, miss the job, feel outside society — and it has been happening for almost three years. Adults who do not themselves know the difference still have to teach when to use a model and when not to. The second face of the cube, teacher time, is how they get the fluency to teach the third. Labenz's analogy is the doctor's invisible night shift of notes. Falk's pitch is blunt: use AI outside the classroom if it helps, and of course it will. Turn scattershot notes into something a parent, a student, or a headmaster can read. Research three students with unfamiliar challenges, talk it through with a chatbot, walk in the next day with something to try — four hours down to twenty minutes. Feed it an old math test and ask for three versions, one themed around soccer.

The soccer wrapping, they agree, is a cheap hook. Labenz doubts a child who likes basketball and not math will grind extra problems because the word problems mention jump shots. "It doesn't work," Falk says, except for particular students. The real reward is the feeling of having learned something. The personalization that actually moves interest is difficulty: he taught math, and in every class some kids were starving and some sixteen-year-olds were still lost on negative numbers. Matching the work to the real level is the lever. That requires either trusting the model enough to send work unreviewed or spending time teachers do not have. His compromise is cultural: 90 to 95 percent good is acceptable. Some exercises will be bad. The time back — more talk with students, better follow-up — can still be a net gain. Schools have to allow mistakes, sometimes even a net negative, because that is how you learn the tool.

Learning styles, the visual-versus-auditory story Labenz feels in his bones as a podcaster, are a scientific dead letter. The feeling is real, which is why the misconception is so hard to kill. Modalities still have uses — audio on a bus, images where images help — and blending them is the advice, not marrying one. Interactivity looks like a genuine primitive: some language-learning studies credit it, and physiologically it rides dopamine and anticipation. That is also why a NotebookLM conversation between two voices holds attention better than a monologue. Falk has watched teachers on stage who do not outperform NotebookLM's slide-and-talk videos. "Wait another six months," he says.

Should teachers use AI detectors? Falk does not hedge. They do not work. Dead end. Grading is trickier. Asked again, he would still say no, then allow a second opinion: grade the essays yourself, run them through a model, and look again where the two diverge. Use the model as first assessor and you will fall asleep at the wheel. The likely cost is systematic disadvantage for kids without English as a first language, or other atypical groups.

Labenz notes that new-model technical reports already lean on AI-generated evals, and that human rater agreement is often the fig leaf. Falk says human evaluation will be needed to keep the machines honest for a generation, and then the competence to do those checks will fade. When he ran Sweden's national mathematics tests, including an oral portion, scorer reliability was "basically like tossing dice, and sometimes worse." In those conditions an AI judge can look attractive. The new problem is a single LLM applying one bias to every student, whereas a thousand teachers at least make noisy, uncorrelated mistakes. LLM bias, once found, can be calibrated; human bias is harder to rewrite. A teacher should not just open a chatbot and start scoring. It takes a framework.

Labenz asks him to ignore the no and give the busy teacher a recipe. Falk starts with sympathy, then a joke: do this properly and you could sell it. Break the judgment into numeric scales — typos, vocabulary, structure — and compose them by an explicit rule so a student or parent can point at the dimension that is wrong. Identify groups that could be hurt: non-native speakers, boys and girls, whatever else fits, and check that the score distributions are not quietly warped. Leave a way to complain. Be transparent that a model scored the work. Do all of that, he says, and you have basically complied with the EU AI Act, which means you could in theory sell the thing in the Union. It may still take more time than a teacher has. These days you could also vibe-code the pipeline in Lovable.

On the third face of the cube, two items are urgent. First, how to use AI for learning and when not to — even if the school bans it, because home use will not wait. Second, AI friends. There are already sad cases of severe harm, not only suicide but quieter emotional and social damage. Personally Falk would prohibit companions under 18; we do not know enough. You need not go that far, but you do have to talk with kids about what an AI friendship is, how to relate to it, and which signals should worry them. Ninety to 95 percent of the time, no problem. They still have to know what the remaining slice looks like. After that come deepfakes, critical thinking, writing feedback, information management, automation; then the civic layer — democracy, concentration of power, the labor market, how the technology actually works, law and ethics. He is writing a book. It could eat an hour of every school day, and it should run from preschool through adult education, sometimes as a tweak inside writing class, sometimes as a real shift, as in programming with AI.

Labenz is already bracing for AI-friend toys. The line between tutor and companion will blur, because the science of teaching says the relationship is one of the most important factors in learning, and a good tutor will grow rapport on purpose. Eugenia Kuyda of Replika told him years ago that the moat in AI apps is relationship: people do not drop a friend because they meet someone smarter. Subscription incentives then pull toward dark patterns that make you stay; ad-supported free versions maximize engagement or steer commerce. He is not an ads hater, but he does not want his young kids as subjects in that experiment. Falk's mitigation is to look at the incentive: a nonprofit whose goal is education is less likely to addict than a company whose goal is monthly revenue. They still lack a map.

Misconceptions, Labenz thinks, are now doing harm in the other direction. Hallucinations still exist, but on elementary-school content he would not bet that GPT-5 or Claude is less accurate than Wikipedia. People also treat chatbots as a search box — "pancakes vegan recipe" — and miss the assistant, the discussion partner, the problem-solver. Stale six-month-old impressions of capability are a huge risk of being wrong. The slogan that models "just predict the next token" is wrong twice: even next-token training produced language-independent concept representations that look like higher-order thought, and reinforcement learning now rewards the final answer, not the token-by-token chain. Apollo's look at hidden OpenAI chains of thought found a private dialect, not any corpus. Underestimating power, Labenz says, is the dangerous meme now. Falk adds that models are not traditional software. You cannot read the code and see the decision. They are black boxes, grown rather than built.

System change, for Falk, is a pile: working tutors turn teachers into mentors along Alpha School lines; disappearing entry-level jobs might mean two more years of school or state-backed traineeships; mass unemployment means reschooling at a scale the ship cannot currently turn. Even a 5 percent chance of something radical in five or ten years is enough to start planning, because education is slow. Labenz asks whether standardization itself is ending. Alpha School still proves itself on ordinary tests, which he understands as a translation for the outside world and also sees as a thin slice next to what an AI already knows about a student. He imagines assessment as a week of watching computer use rather than a morning of multiple choice.

Falk agrees 100 percent, then offers what he calls, oddly, one of his most radical ideas: the age of grades is coming to an end. Reducing a rich model of a learner to a letter from A to F is insane except that it fits current metrics. Cheating is a problem because people are chasing the grade; produce something that looks good and does not reflect knowledge, and the letter still pays. If the motive is learning, AI is basically a tool, not a threat. He does not know how to make the transition. Universities will still need admission tests or a lottery. New problems will appear. Directionally he thinks getting rid of grades would be a painful kick the right way, and that the harm of grade-as-motivator is already arguable and will get worse every six months. The present system was built for the first industrial revolution, almost two hundred years ago. He is not sure it fits this one.

Labenz wonders whether, in the window before models deliver Einstein-level paradigm shifts, school should hunt the few people who still can. Falk's counter is "why?" Society functions without an exclusive focus on the top 3 percent. Running the whole system to find them for GDP is a huge investment; waiting five years for cheaper AIs to do that work is the easier bet. The live problem is the other 97 percent: UBI without the feeling of being useless. He is not very worried that people will complain about joblessness. If he were unemployed and provided for, he would be happy playing board games, hanging out with family, reading — and he already has a job he finds interesting, which he thinks 90 percent of people do not. What does worry him is a small minority gathering the resources and leaving everyone else without the UBI. Labenz's version, which he once put to Jake Sullivan, is that if the political class can handle international relations, he trusts ordinary people to spend the peace dividend.

Advice for Monday is deliberately light. Save five minutes a day with AI — Falk's most repeated mantra. You do not have to teach a unit or read a paper first; the minutes teach you what the technology can do and what students will need. Make more mistakes. Identify the costly ones, then celebrate the rest, the way you already let students err. Focus on students. Learn enough AI to decide whether your subject requires teaching it, and how; principals and owners fund the rest, and districts look three, five, ten years out. Labenz adds Matt Cahill's line: teach and learn every day. On AI, age and seniority do not confer a head start. Labenz also wants classrooms doing future-vision work — not laser swords, but the science-fiction habit of asking what happens if this variable changes — as a reminder that agency is not gone. Falk is putting short videos for teachers at graspable.ai, with longer pieces on the Graspable AI Substack.