The Cognitive Revolution · 2025-10-08 · 112 min

Education in the AI Age: a Teacher Rethinks Learning & Purpose, w/ Johan Falk of Graspable AI

Erik TorenbergNathan LabenzJohan Falk

YouTube ↗

TL;DR

AI makes education’s labor-market mandate newly contingent: if society moves toward UBI, Johan Falk says supplying employers with competence could become “basically irrelevant.” Schools would then need to emphasize citizenship, cooperation, well-being, arts, and personal growth—yet Falk’s children are seven and 10, with roughly another decade in a system that takes about five years merely to change a curriculum.
The binding constraint on education technology is institutional speed, not model capability. Falk points to South Korea, Singapore, China, and especially agile, data-ready Estonia as early movers, while warning that research arrives roughly two years late and can rely on tiny, poorly generalizable samples. His policy call is deliberately asymmetric: establish national direction, fund teacher competence, define data boundaries, and act because “non-action is a great risk.”
AI tutoring could democratize what was once available only to royalty, but motivation remains the scarce input. A conversational expert costing perhaps $20—or even $2—a month may accelerate learning dramatically, yet Khanmigo’s early experience showed self-directed students racing ahead while already-confused students remained stuck. Nathan Labenz’s formulation survives the caveat: “There’s never been a better time to be a motivated learner,” but never an easier time to fool yourself about learning.
The near-term opportunity is teacher augmentation, while mandatory classroom deployment is premature. AI can turn four hours of research into 20 minutes, produce differentiated exercises, organize notes, and help teachers reclaim time; Falk’s practical target is to “save five minutes every day.” The investable category is workflow leverage with human review—not an AI silver bullet imposed uniformly across classrooms.
Assessment is both a major automation market and a governance trap. Falk rejects AI detectors outright and opposes unstructured AI grading because teachers may “fall asleep at the wheel,” but he accepts tightly governed second opinions—and, under pressure, a system of explicit rubrics, subgroup bias tests, transparency, and appeals. Nathan adds inference-time redundancy: multiple models or prompts, consensus thresholds, and mandatory human review when grades diverge.
AI literacy is more urgent than AI-mediated instruction, especially around learning self-deception and synthetic relationships. Falk would prohibit AI companions below age 18, while hedging that “90%, 95%” of interactions may cause no concern; the danger is concentrated in dependency, emotional harm, and products optimized for retention, ads, or commerce. Tutor and companion will blur because relationships drive learning, making incentives and institutional ownership as consequential as model quality.
Falk thinks “the age of grades is coming to an end,” because continuous AI observation makes a single letter or standardized test informationally crude. Grades also create the incentive to cheat, whereas a student motivated by learning receives mostly upside from AI. The deeper transition is from schools designed for the first industrial revolution toward systems that cultivate agency, authentic expression, relationships, and worthwhile lives—even if AI eventually outperforms most people economically.

Digest · the substance, structured for research

1. AI has reopened the question of what education is for

Falk divides education into three purposes: supplying competence to the labor market, forming functional and responsible citizens, and helping people grow through arts, knowledge, and experiences that are “good for you” without being economically necessary. AI destabilizes the first purpose while making the contested values inside the other two impossible to avoid.
Sweden weights those purposes differently from the United States. Falk describes a social-democratic culture with high taxes, broadly available healthcare, and free higher education, producing less educational competition and a stronger conception of the common good; Labenz notes that employability has clearer external validation than either “good citizen” or “good person.”
Falk’s uncertainty is personal: his children are seven and 10 and will spend roughly 10 more years in school, yet he has “no idea” what society will look like when they finish. Reading, writing, self-understanding, and cooperation remain durable; quadratic equations and additional languages might be useful, but he no longer assumes they are necessary.

2. A five-year curriculum cycle cannot track fast-moving AI

Falk identifies agility as the system’s most important missing capability. Changing a curriculum in five years counts as relatively fast, but “five years in the AI world” contains massive change; educational policy can therefore become obsolete before implementation, just as Labenz has watched startup features become outdated during a development cycle of only months.
The institutional drag is both emotional and structural. Organizations cling to sunk costs and familiar paradigms, while genuinely new approaches have not been stress-tested; large public systems add procurement and organizational inertia that can make the mismatch “an order of magnitude harder, maybe two.”
Falk finds no jurisdiction that has solved the problem, but sees useful pieces: Singapore combines top-down adoption with education-data standards; South Korea couples deployment with limits on using learning data for unrelated purposes; Estonia pairs national agility with strong digital infrastructure. China has long developed AI curriculum and offers Squirrel AI, while U.S. deployment includes Khanmigo’s planned expansion toward one million students and teachers, reportedly with Microsoft support.
His national playbook is concise: declare AI strategically important, resource teacher training, state what is permitted, and draw firm data boundaries. Countries hesitate because the right move is unknowable, but Falk’s judgment is that “doing something and improving along the way is better than doing nothing”; otherwise, they risk widening the digital divide.

3. Falk’s incomplete cube separates four different education markets

What initially looked like an undifferentiated mass eventually became “four different sides of a cube” for Falk: using AI for learning, using it to reduce teachers’ non-classroom workload, teaching students about AI, and preparing for system-level changes to schools, curricula, and teachers’ roles.
The missing two faces are deliberate: “we don’t have the full picture yet.” The framework prevents leaders from treating every AI-and-education question as tutoring or cheating while acknowledging that new categories may emerge as products and institutions evolve.
Counterintuitively, Falk calls the two most visible categories—student learning tools and teacher productivity—the least important at the moment. Teaching AI competence is more urgent because children already use these systems independently, while systemic consequences may ultimately reshape the entire purpose and structure of schooling.

4. Thin evidence argues for experimentation, not universal mandates

The research points in opposing directions: some studies suggest students can learn twice as much or halve the required time, while others report learning losses. Falk’s objection is not that the findings are useless, but that one might rest on 25 adults in Nigeria and another on 18 people instructed to use a chatbot for an essay—weak foundations for a Swedish middle-school mandate.
Research also reaches decision-makers roughly two years after the intervention, by which time language-based AI may have changed substantially. Falk distinguishes these fast-moving “thinking machines” from slower, traditional applications of neural networks to educational data; pooling both under “AI” can make ostensibly evidence-based policy misleading.
His recommendation is classroom-level discretion: interested, competent teachers should try AI where it fits their students and goals, but systems should not yet require everyone to use it as a learning tool. Labenz accepts that hedge while warning that “not proven for everyone” can become an excuse for most educators to avoid learning anything at all.
Falk’s answer is to build teacher competence without declaring a silver bullet. Teaching a class is “bizarrely difficult,” involving perhaps 2,000 decisions daily; AI may help one child practice German, another confront mathematics, and another become curious, while the correct intervention for a restless student may simply be running outside for 10 minutes.

5. AI tutoring amplifies agency more reliably than it creates it

Falk sees enormous potential in conversing with an expert on almost anything—an experience previously available mainly to “some kind of royalty” with personal tutors. At perhaps $20 or even $2 per month, such access could matter even more in poorer countries; Labenz notes Khan Academy’s U.S. retail price around $4 and the generosity of ChatGPT’s free tier.
The constraint is whether the learner can articulate a goal and keep moving. Salman Khan reportedly observed some children immediately “run” with Khanmigo while others remained confused and stuck; teachers recognized the same students who, without AI, could not explain what they were doing or what help they needed.
Falk’s former math classes contained advanced students who were understimulated and 16-year-olds still struggling with negative numbers. Adapting work to actual knowledge could restore interest, but the broader teacher role becomes motivation, initiation, and persistence—the human work Alpha School assigns to mentors, guides, and coaches rather than content presenters.
Labenz’s own AI-assisted reading of technical and biology papers supports the booster case, yet Alpha School’s two-hour learning model reportedly uses no chatbots at all, instead combining internally built and licensed applications. The contrast matters: strong results attributed broadly to “AI education” do not establish conversational tutors as the operative mechanism.

6. The best classroom uses are still small, situated experiments

Falk imagines sending a student fascinated by black holes to converse with Gemini, then asking for a Thursday presentation plus three difficult questions classmates would genuinely want answered. The teacher chooses the tool, subject, deadline, and social output; the chatbot expands available depth without replacing educational judgment.
His strongest anecdote began when a visiting child grew bored and reached for a phone. Falk generated an age-adapted interactive story with choices; half an hour later, the child’s mother said she had rarely seen him read so closely or intently. The point is exploratory, not causal proof: experimentation reveals affordances that standardized policy cannot anticipate.
Falk also describes an “AI wizard” acquaintance rapidly building tools for dyslexia, adapting them for non-native Swedish speakers, and possibly creating a tool to interpret shaky handwriting associated with Parkinson’s disease. AI-written code lowered the cost of moving between adjacent needs, illustrating how individual teachers and parents may discover niches before formal procurement processes can name them.
Yet children are already conducting uncontrolled experiments at home. Some use AI to deepen learning; others “fool themselves into believing that they’re learning stuff while they’re actually not,” fall behind, and accumulate academic and social consequences. That is why instruction about using—and deliberately not using—AI cannot wait for consensus about classroom tutoring.

7. Teacher workflow gains are immediate but require tolerance for imperfection

Outside the classroom, Falk expects AI to help teachers convert rough notes into parent communications, digest large information sets, plan lessons, and investigate unfamiliar student needs. Research and brainstorming that once took four hours might take 20 minutes, freeing attention for students rather than merely increasing administrative output.
A teacher could feed in an old mathematics test and request three variants, including one themed around soccer, then edit the results. Falk treats generated material as inspiration rather than automatically classroom-ready content; the more important personalization is usually matching difficulty to knowledge, not wrapping every exercise in basketball references.
His operating threshold is pragmatic: material that is 90% or 95% good may contain exercises that fail, but the net outcome can still improve if the saved time enables conversation and follow-up. Sometimes an experiment will even be net negative while yielding learning—a tolerance schools need if they expect teachers to discover responsible practices.
Learning-style doctrine is the wrong justification for multimodality: Falk says the science is clear that fixed visual or auditory learning styles “don’t exist in that way,” though the subjective feeling does. Audio on a bus, images or text where useful, and movement between media can still help; he cautiously recalls evidence favoring blended modalities, adding, “don’t quote me on that.”

8. Interactivity is a new primitive, not proof of better learning

Falk expects interactivity to support engagement through anticipation and physiological mechanisms involving dopamine, and he recalls language-learning research attributing gains to that feature. He remains appropriately uncertain about a general causal claim, but AI dramatically expands the number of contexts in which students can ask, respond, branch, and revise.
Two voices are often more engaging than a monologue, which helps explain NotebookLM-style conversational audio. Labenz already gives such synthetic discussions a minority—but growing—share of his listening because they can cover a paper or small collection for which no human podcast exists; the relevant comparison is not perfection but reading and falling asleep.
NotebookLM’s narrated presentations suggest another convergence of text, slides, speech, and interaction. Erik says that, in AI-and-education topics he knows deeply, he can produce better content, but has seen teachers and other presenters who do not do a better job; Labenz jokes to wait another six months.

9. AI assessment needs governance before it needs scale

Falk’s verdict on AI detectors is categorical: “They don’t work. Dead end.” On grading, his initial answer is also no, because a teacher who accepts the model’s first assessment may “fall asleep at the wheel,” reproducing hidden disadvantages for non-native speakers or other atypical groups.
A second opinion is more defensible: grade the work first, ask AI independently, and revisit large discrepancies. Falk’s experience overseeing Swedish national mathematics tests complicates human exceptionalism—evaluations of oral-test scorers found reliability comparable to “tossing dice and sometimes worse.”
AI can be more consistent, but one LLM applying one bias to everyone may be worse than a thousand teachers whose biases partly dissolve into noise. Detected model bias can theoretically be corrected more easily than human bias; Falk nevertheless warns that individual teachers should not casually deploy generic chatbots for consequential grades.
Pressed by Labenz’s hypothetical teacher who will automate regardless, Falk proposes decomposed rubrics with numeric subscales, documented aggregation, subgroup comparisons, transparency, and a route for students or parents to appeal. Labenz adds multiple models or prompts: accept strong consensus, but require human review for a 3–2 split or evaluations separated by two points on a seven-point scale.

10. AI literacy now includes relationships, incentives, and power

The first urgent competency is recognizing when AI supports learning and when it substitutes for it. Children have inhabited a world of chatbots for almost three years, Falk says, yet most schools still do little to explain the technology, the learner’s responsibilities, or the distinction between productive assistance and self-deception.
The second is AI companionship. Falk would prohibit companions for anyone under 18; he has heard of cases involving severe harm and warns of risks including suicide as well as emotional and social harms. He simultaneously preserves the denominator, estimating that “90%, 95%” of interactions may create no concern. Education should teach warning signs rather than imply that every attachment is catastrophic.
Labenz’s pushback is structural: relationships are among the strongest contributors to learning, so the best tutor may necessarily resemble a friend. A long-lived relationship may also become an application moat because users do not abandon a friend merely upon meeting someone smarter; educational rapport, emotional dependency, and commercial lock-in can therefore emerge from the same product feature.
Falk’s partial safeguard is incentive design. A nonprofit optimizing education has better chances of avoiding dark patterns than a subscription business maximizing retention, while Labenz worries even more about free, ad-funded companions steering children toward engagement or purchases. “We still lack the map for navigating this terrain,” Falk admits.

11. Outdated mental models cause institutions to underestimate AI

Falk sees many users treating ChatGPT as “Google but from OpenAI,” entering “vegan pancake recipe” and missing its roles as collaborator, critic, assistant, and problem solver. Six- or 12-month-old experience is also dangerous evidence in a field whose capabilities may change before a school completes a policy review.
Labenz argues that the hallucination narrative was formed around weaker systems: GPT-5 or Claude 4 may not be less accurate than Wikipedia for elementary-school material, although neither speaker claims infallibility at the fringes. The corrective has swung too far when occasional error becomes reassurance that educators can ignore the technology.
“It just predicts the next token” is similarly incomplete. Labenz points to language-independent internal representations of shared concepts and to reinforcement learning that rewards correct final outcomes rather than any particular token sequence; research into hidden reasoning traces has even found compressed, nonstandard shorthand unlike ordinary training text.
Falk adds the institutional implication: AI is not traditional software whose decisions can be recovered by reading source code. Researchers can illuminate fragments only with major effort; the systems are “essentially black boxes grown rather than built,” demanding empirical oversight even when outputs look reliable.

12. Continuous evidence makes grades and standardized tests look primitive

If an AI continuously observes what a student attempts, learns, struggles with, and excels at, Falk says reducing that record to A–F “makes no sense.” Standardized tests remain legible to incumbent institutions—one reason Alpha School still uses them—but they are a thin snapshot beside longitudinal, task-level evidence.
Falk’s most radical claim is that “the age of grades is coming to an end.” AI-assisted cheating is chiefly a problem when the grade or exam supplies the motivation; for someone driven by learning itself, AI is overwhelmingly a tool. He wants education to move from extrinsic ranking toward the desire to learn, while admitting, “I don’t know how to make that transition.”
Abolishing grades would hurt and create new allocation problems: universities would need admission tests, lotteries, or another mechanism. Falk does not pretend every consequence improves, only that grade-centered motivation may already do more harm than good and will deteriorate “for every six months that passes.”
The historical framing is intentionally large: today’s educational system was built for the first industrial revolution nearly 200 years ago. AI tutors, persistent assessment, disappearing entry-level work, and changing concepts of knowledge make it doubtful that the same standardized architecture fits the present revolution.

13. Authentic expression may survive after economic superiority does not

Labenz proposes replacing minimum-length essays with stringent caps—even tweet-length assignments asking students to state something they fully endorse and could read aloud in 15 seconds. Falk offers a companion exercise: write about a beloved hockey team or band, where an AI’s first draft will omit details the student considers essential and therefore provoke genuine revision.
Falk resists defining human purpose as beating AI. A person may enjoy an instrument without becoming the best professional; the reward is development, expression, and expanded capability. That may offer a path forward even when machines can perform economically useful tasks better.
Labenz asks whether schools should instead find the top 3% capable of Einstein-level paradigm shifts. Falk sees experimentation value but doubts society should run an entire system around that screening problem when AI may gain the capability five years later; meanwhile, it still must create meaningful lives for the other 97%.
Falk is less worried that people will complain simply because they lack jobs than that a small minority will capture all resources and leave everyone else without UBI. Personally, he could happily play board games, read, and spend time with family; the challenge would be helping people feel they are living well rather than merely being unemployed and useless.

14. Teachers should model learning by making more mistakes

Labenz’s preferred culture is “teach and learn every day”: adults and children are on the same AI timeline, regardless of age, so teachers should experiment, share failures, and model adaptation. Falk sharpens that into “make more mistakes”—after identifying the costly ones, schools should celebrate ordinary failures as evidence that learning is occurring.
Falk’s practical mantra is for every teacher to use AI to “save five minutes every day.” Repeated utility builds intuition about what AI can do, what students need to understand, where it belongs in teaching, and where it fails, all while returning rather than consuming scarce time.
Science fiction can widen institutional imagination because its useful form asks what society becomes when one or two premises change; dismissing scenarios as “just science fiction” ignores how much former fiction is already present. Labenz extends that into a student assignment: imagine a different future and practice the agency required to steer toward it.
Falk’s closing division of responsibility is calming but firm. Teachers should focus on students and learn enough AI to decide what their subject requires; principals and school owners must fund professional development; districts and national leaders must plan across three-, five-, and 10-year scenarios. The individual starting point remains deliberately small: “save five minutes a day using AI.”

Erik Torenberg

Today, just back from The Curve in Berkeley, where I had so many amazing conversations that I ended up losing my voice, I'm pleased to share an exploratory conversation about the impact of AI on education with Johan Falk, author, speaker, and AI analyst based in Stockholm, Sweden. Johan spent several years as a classroom teacher and at Sweden's National Agency for Education before pivoting to focus on AI and education full-time in the wake of ChatGPT. Today, he's making videos to help teachers use AI in their work and classrooms, which you can find on his new Substack, Graspable AI, at graspable.substack.com.

The context for this conversation is that I was recently invited to give a keynote to an audience of 500 public school administrators in my home state of Michigan, where my kids are now in first grade and preschool in the Montessori program of our neighborhood elementary school. In another timeline, I could easily imagine myself being a sort of tiger dad when it comes to my kids' education. But in this timeline, with the long-term outlook for the job market and, frankly, the structure of society as a whole subject to such radical uncertainty, I had, for the most part, been letting my kids be kids and watching how things develop before committing to a specific strategy for their education.

As I've prepared for that talk, including by requesting ChatGPT Deep Research reports, trying various learning apps for kids, and talking to teachers and principals in my personal network, I've been struck by just how perplexing the challenge that AI presents to educators really is. AI, of course, offers unprecedented access to information, unlimited feedback, and one-on-one tutoring, as we learned in our recent episode with Mackenzie Price of Alpha School. And yet, at the same time, it enables cheating like never before and also raises fundamental questions about the very purpose of education itself.

To help understand how educators in different contexts around the world are approaching these challenges, I invited Johan on to help me explore the many facets of the relationship between AI and education. We ended up covering a lot of ground, including what existing research shows about AI's effects on student learning and how the desire to be evidence-based is fundamentally challenged by the pace of change; how different countries, from South Korea to Estonia, are approaching AI adoption in education; what students need to learn about AI itself, including the risks associated with AI friends; how teachers can model a positive approach to AI, including a willingness to explore and make mistakes; whether the age of grades and standardized tests might soon be coming to an end; how education systems might need to fundamentally reimagine their role in students' lives as economic productivity becomes less relevant; and lots more along the way.

While we ended as we started, with more questions than answers, I found Johan's frameworks extremely helpful for organizing my thinking. I hope you enjoy this wide-ranging exploration of AI's impact on education with Johan Falk, author of Graspable AI.

Johan Falk

Thank you. I'm thrilled to be here. I'm excited for this conversation as well.

Erik Torenberg

You're coming to us from Stockholm, Sweden, and you, like me, have managed to make a full-time job out of trying to keep up with AI, which is something that not too many people have been able to do. We're in a very privileged, small minority who have the luxury of spending so much time really thinking about what matters in this space and trying to translate that to people who do have full-time jobs that keep them from keeping up with the increasingly dizzying pace of events.

For starters, we're obviously primarily going to focus today on the impact of AI on education, although I'm sure there'll be some digressions. But for starters, do you want to introduce yourself, tell us who you are, what you do, and who you do it for?

Johan Falk

Yeah, sure. I have a mixed background. I'm a physicist originally by training. I also spent 3 years as a science journalist and web developer, then became a teacher, and finally decided to go into education, which I'd been longing for for a decade or something.

I spent a few years as a teacher and then went on to the National Agency for Education here in Sweden, fully engaged in improving mathematics education in this country, which I managed pretty well. Then ChatGPT was launched, and everything was turned upside down. I initiated a team at the agency working with AI and education, and after 18 months with that, I decided to leave the agency to focus on AI more broadly because I think it is such a huge question for the whole society and the entire world.

Right now, I'm looking for a way to have a positive impact on AI risks or AI safety. I'm running my own business while looking for a way to do that.

Erik Torenberg

Yeah, cool. We're definitely very kindred spirits in that respect. Obviously, Sweden is quite a different country, context, and philosophy from what I'm used to here in the United States.

Super big picture, a couple of things that keep ringing in my head right now are, first of all, Sam Altman saying that his kid, who was just born, is going to grow up in a world where they're never smarter than AI. I think that's a really striking reality. For us, it's sort of happening in our lifetimes, but already people are born who are never going to be smarter than AI. That's pretty wild.

There's also, of course, Dario Amodei's forecast that we might see very significant labor-market disruption in the not-too-distant future. I would say in the U.S., it's often a tacit assumption, but the working assumption behind education, for the most part, is that it's about teaching you to be economically productive, so that when you're done with your education, you can enter the workforce, make a good living for yourself, and make a contribution to society—to the GDP.

How imperfect and incomplete as that measure is, and all the more so all the time, how do people in Sweden think about it? Is it the same? Is it different? Is it starting to change, in your mind and potentially in society more broadly, in light of AI?

Johan Falk

Yeah. AI forces us to discuss these questions. It puts most things in perspective, not least education. I would say in Sweden, and basically all over the world, there are more or less 3 different perspectives on education.

We have supplying competence for the labor market, which is what you talk about. We also have the idea of fostering citizens, providing them with the knowledge required to be functional and responsible citizens, but also with the values we have in society.

Then we have the third part, which would be just growing as a person, learning arts and things that aren't productive or really necessary in any functional way, but are good for you and good for the world because when you feel good, it's good for society. Things like that.

I would say those 3 parts are more or less the ingredients in the purpose of education in every country.

Erik Torenberg

Yeah, it is striking that the latter 2, certainly here in the U.S., are just radically more contested than the first. It seems like we have a pretty clear sense of what it means to be employable. You can either get a job or you can't, so there's some sort of ground truth that emerges there.

When we think about making you a good citizen, that obviously raises the question of, well, what is a good citizen? And when we think about growing as a person, it's, well, in what ways and with what values, toward whose measure of what a good person should be?

I imagine that those ideas are probably a little less contested in Sweden, but maybe not.

Johan Falk

Yeah. I haven't lived in the United States. I've been there once or twice, but my impression is that in the U.S., you have much more of a culture of competition than we have in Sweden.

We have—not communism, but—we have a strong social-democratic culture in Sweden. We have high taxes, free healthcare, well, to a certain extent, and free education. Even higher education is free in Sweden. So we have a much stronger sense of community and the common good in our society, which I guess lowers the competition aspect of education. There's just a bit less competition in society overall, I guess.

Erik Torenberg

So do you see any change, if only in yourself, in terms of how you think about the purpose of education or what you're trying to do as you advise school systems? Are you reframing in your own mind what the purpose of the whole enterprise is?

Johan Falk

Yeah, I am. I think this first part—providing the labor market with competence—should be questioned. It depends so much. We don't know where AI is taking us. You might think there's a 2% chance that we have a radically different future, or we have a 99% chance of a radically different future.

If we're moving toward a future with UBI, that means that competence for the labor market is basically irrelevant, and schools should be focusing much more on questions like: How do I, as a person, as a human, feel healthy? How do I have well-being? Things like that. We have science for doing that as well, but those parts aren't taught very much in schools nowadays.

You could also question how much we should try to reorganize schools today when we don't know the future. I have 2 kids. One is 7 and one is 10 years old. They're going to spend roughly 10 more years in school, and I don't have any idea how the world and society will look when they're done with school.

I know that everyone needs to learn how to read and write, how to understand themselves, how to cooperate with others, and things like that. But solving quadratic equations...

I don’t know. Learning new languages, I think that might be good and useful, but is it something that is necessary? Probably not. Yeah, so much is unknown right now.

I think the most important thing that we can learn is to increase the agility in the educational system, because that’s one of the things I keep coming back to when I think about and analyze what AI is doing to education in society: the pace of change is a huge challenge. It takes years to change a curriculum. For example, 5 years is rather quick. And if you look at 5 years in the AI world, it’s massive changes.

That means if we want to tackle the changes that AI brings about, we need to be much more agile than we are today. And that’s a real challenge. I don’t know how to do that.

Erik Torenberg

Yeah.

Nathan Labenz

Yeah. That’s a long answer to a short question. Someone once said, “Don’t worry about that at all.” The pace-of-change challenge is real for everyone. It’s even real at technology startups.

I’ve experienced multiple instances where we’ve been working on some new feature, some new capability, whatever, for the product, and the way we are building it becomes essentially obsolete before we’re even able to launch it. Then you’re faced with this weird situation—and that’s usually a couple of months at most from a feature starting development to actually getting online for users.

It creates probably the same challenges that are familiar everywhere, where it’s like, well, we’re kind of far along on this. There’s some sunk-cost-fallacy attachment. Should we rip it all out? What do we do now? Then there’s always the question of, well, this new way seems better, but we haven’t really stress-tested it, whereas we’ve kind of gotten comfortable with the paradigm that we were building on over here. So what do we do?

All too often, I think even in technology startups, there’s a reluctance to switch paradigms because of all these different reasons, which largely boil down to emotional reasons. You could also say that risk management, in a non-emotional way, is part of that as well.

Erik Torenberg

And structural slowness—just the inertia of big organizations or small organizations.

Nathan Labenz

Yeah, it’s an order of magnitude harder, maybe 2 orders of magnitude harder, at different kinds of organizations. So I guess, well, maybe let’s go right to that: is anybody doing a good job of this today?

I’ve got an invitation coming up to speak to a bunch of public school administrators in my home state of Michigan. One of the things that I’ve been thinking about with them is that they might need to rethink their procurement methods.

By the way, the same thing is happening at the Pentagon and DARPA. I know there’s major anxiety there, where usually we go through this whole super-long process, and now we don’t necessarily know. There could even be a new entity that doesn’t exist yet that we want to be buying from in a year’s time. How do we even conceptualize that?

Is there anybody that you’ve seen in the education sector, or maybe a little more broadly in the public sector, who has figured out a workable model for keeping up with the pace in any respect?

Johan Falk

Well, no, basically. But I’ve done some research looking at where in the world we have seen some good examples of what to do, and South Korea and Singapore are doing a good job of implementing AI in education. They have strong top-down incentives and approaches to just making AI happen.

I think it’s Singapore that also has good standards for data in education, which makes it easy, or easier, to apply AI to a lot of things happening in schools and education. There are some other examples as well. I think the US is partly a good example.

You have Khanmigo from Khan Academy being rolled out. They were saying that it was going to scale up to 1 million students and teachers, which is pretty good. I think it was Microsoft giving money to make that happen, or, well, making the economics of it work.

China is doing a pretty good job. They have actually been working with AI in the curriculum for quite some time. I don’t know if it was 10, 15, or 20 years, but now that has started to become reality. They also have something kind of similar to Khanmigo. It’s called Squirrel AI in China. Obviously, it’s a bit different when it comes to values.

Estonia is a good example. It’s a small country, and they’re agile, which I think is crucial. They also have good data infrastructure in Estonia, so they have a good platform to build on, and they rolled out AI support for learning.

Everyone is wrestling with this, trying to understand what to do and how to do it. The technology is moving so fast that we can’t really know what’s working or not, because all the studies we get are 2 years old when they’re published.

It’s also difficult because AI is so many different things. I’m often focused on language-based AI. I think that’s the most transformative thing happening right now. Those are the thinking machines.

In neural-network machine learning, you can apply it to many different things, and that is important and changes a lot of things. But language-based AI is advancing so rapidly, and it’s also being adapted rapidly, compared to, I don’t know, applying neural networks to educational data, for example.

Some of the studies concern using AI in more traditional ways, and that is not as interesting, I think, because it’s not moving at the same pace. If I were to pick out some things that I think are important at some kind of national level, it would be that you have a strong strategic initiative from the top down, saying, “This is something we should do. Everyone should use AI, or every teacher should learn AI in schools.”

Provide resources for actually doing that. Give some kind of clear guidance saying that this is okay, that is okay, and these 2 things we shouldn’t do. It’s prohibited to use learning data for purposes other than education, for example. That’s something I think is done in South Korea.

Then also just be bold. Non-action is a great risk right now, because you miss out on so many opportunities and you risk ending up on the wrong side of the digital divide if you’re not acting now. Many countries are not acting because they don’t really know what to do. That’s quite understandable, but doing something and improving along the way is better than doing nothing, I would say.

Erik Torenberg

So, for one thing, we can always take inspiration from Estonia. I’ve been amazed on multiple different levels with the quality of governance in Estonia and, of course, Singapore and, in many cases, South Korea as well. But Estonia, I think, is a little bit of a hidden gem when it comes to the good governance that it has. It has, in many cases, pioneered really interesting paradigms.

You mentioned, of course, the fundamental challenge of pace and then the fact that research, which obviously takes time, is often outdated when we get it. I do want to come back, though, and take a little survey, because I know you’ve done the hard work of going through the literature that exists and trying to get a sense for what research we do have. What does it say?

Then I definitely want to unpack the guidelines that you would recommend in a little more detail as well. We have all the time you need for that.

Maybe before we do those 2 things, how do you organize this? What I think is another fundamental challenge that people have with something like AI is that it touches everything. I experienced this in my own way, where I'm trying to keep up with AI. Well, what does that mean? Increasingly, it means a whole layer across the entire—not just economy but society, right?

And now I'm thinking, well, geez, I really can't keep up with everything going on in society. It sounds pretty ridiculous if you say, “I'm going to try to keep up with everything going on in society.”

Nathan Labenz

But that's kind of what trying to keep up with AI is converging to.

Erik Torenberg

So there's kind of no way to do that except to have some shortcuts, heuristics, and some sort of taxonomy of exactly what bucket any given thing falls into. I suspect there's probably even a gap there for a lot of people. When we think about AI in education, that can mean a lot of different things, right?

It could mean that kids are using ChatGPT to cheat on their homework. It could mean personalized learning. How do you organize that in your own mind? How do you recommend people set up a mental framework for breaking that down so they can then zoom in on the different categories, or perhaps allocate them to different people in their organization if that's what they're trying to do?

Johan Falk

Yeah, really good question. For a long time, it was just a big mix of things for me, but then different themes emerged. I'm calling it 4 different sides of a cube, which means we still lack 2 of the sides. I think that's important to remember: we don't have the full picture yet.

There are 2 aspects of AI in education that are often discussed and take a lot of time for teachers, students, and principals, but that I don't think are that important. The first one is using AI for learning—that is, the Khanmigo stuff, for example, but also just having an AI study buddy and things like that. That's 1 part: using AI for learning. We'll get back to that, I guess.

The second one is teachers using AI just to save time with things they do outside the classroom, outside teaching. For example, using Excel, but having more powerful tools so you can get work done more quickly with AI.

The third thing, which is strangely silent, is teaching about AI: having AI as part of a curriculum and helping students develop their AI competence. I think that's really important and really urgent, because kids have been living in a world with chatbots for almost 3 years now. In Sweden, at least, and in most countries, I would say, we're still not helping them understand what this is, how they should relate to AI, or how they should and shouldn't use it.

The fourth aspect is the system-level effects on education: how schools might change in themselves, what education is, or how the role of teachers might change. How we use books and a lot of other things might change, and we don't really know how, but that's 1 important bucket as well.

So, 4 different buckets, and 2 buckets we don't know what they are.

Erik Torenberg

Which 2 of those 4 were the ones that don't matter?

Johan Falk

The first 2? Well, today, looking at the state of AI in education, I would say it's not that important to use AI for education as a learning tool, which is a strange thing to say. But I think I can back it up.

The second one is using AI as a tool for teachers outside the classroom. Those are the 2 most immediate things you come to think of when you think about AI in education.

When it comes to using AI as a learning tool, the tricky thing is that we don't really know how it works or if it's a good thing for learning. There are more and more studies, and you've probably heard some of them, saying that you can halve the time you need, or learn twice as much in the same time, when you use AI. Some of them are saying that if you use AI, you learn less.

Those studies are useful, but when you dig deeper into them, you realize that one study was based on 25 adults in Nigeria, and it might not translate to my middle-school class in Sweden, for example. Or the study showing that you learn less when you use AI was based on 18 people who were told to use a chatbot to write their essay. It's really difficult to generalize from that.

And that means that we can't really use what we know. I think that's strange, because I'm into AI, but I don't think we should push for introducing AI as a learning tool for everyone. I think it's fine to use it as a learning tool for teachers who are interested in doing that and feel that it could work for their students in a particular situation.

But saying that teachers should use AI as a learning tool is, I think, still a mistake. That being said, we need to keep a close eye on the research, because there are quite a few promising results. It's promising enough that we might conclude this is really good and that we should start using it for everyone who is 16 years old or 13 years old when learning languages, or for kids with special needs, or something like that.

But we should only do that for everyone, or categorically, when we have good research showing that we should.

Erik Torenberg

So this doesn't recommend a sort of classroom-by-classroom approach, which effectively means a teacher-by-teacher approach. It would be a matter of their style, their enthusiasm, and I guess the upshot would hopefully be that, in the diversity of approaches that naturally emerge, collectively people will learn and students will have a variety of experiences.

Is it essentially a hedging-our-bets strategy because we don't really have a clear answer?

Johan Falk

Yeah, I think so. But it's also about considering AI as 1 of many tools that we should or could use. People are different. As you say, teachers have different styles and different AI competence. There are different situations and different things you want to teach the students.

I think it would be wrong to treat AI as some kind of silver bullet that works for everything and everyone. Teaching is really complex. Teaching 1 person is difficult; teaching a whole class is bizarrely difficult. Teachers make around 2,000 decisions every day when it comes to how they teach.

If we raise teachers' AI competence, they will have a much easier time seeing when AI could be used to inspire students, help them practice mathematics, or just speak German, or inspire a student to actually read for half an hour every day. But it might also be a great idea just to go outside: “Let's have a lesson outside today.” Or this kid should really just run for 10 minutes, and then he might be able to sit still. AI isn't going to help with that, I guess.

I think AI can bring a lot of good things and a lot of good tools when it comes to teaching possibilities. The biggest possibility is personalized learning: adapting learning to every single student's level of knowledge, interests, pace of learning, and things like that.

It's possible that this might actually work. There are some signs that it is working, but it's also quite possible that it doesn't work. We've seen this promise before, starting with cassette tapes: “Now we can individualize learning for everyone.” We have YouTube, and we have books.

It's possible today to have individualized learning, but only for students who are active learners themselves. I think that's something that's becoming more and more important in the world of AI.

Nathan Labenz

So if I was going to take the booster angle for a second, which is a pretty natural position for me to take, I suppose, one of the things I often say about AI is that there's never been a better time to be a motivated learner. The flip side of that is that it's also never been easier to cheat on your homework.

We can come back to the cheating part, but I'm glad you touched on individual tutoring because, at least from my perspective—and I'm far from an expert—it seems to me that, from the conversations I have, one of the most widely cited and generally believed ideas in education is that individual tutoring, one-on-one tutoring, is the gold standard and delivers the best outcomes.

We know about the 2-sigma effect, and the idea is that maybe we can have the 2-sigma effect for everyone. If we can get these AI tutors to really work, it does feel, when I want to learn something, that AI is just an unbelievably useful way to go about it.

For me, that's often taking a paper. It could be a machine-learning paper, even more so if it's an AI-for-biology paper. There's just so much stuff that I don't know, and there are important missing pieces in my knowledge that I need to make sense of what I'm currently trying to make sense of.

The AIs are just so good at answering those questions. I really do feel like, man, this is an indispensable advantage in terms of my ability and my confidence that I'll be able to make sense of things like this. In the past, I might get 2 paragraphs into the abstract and think, “I don't know. Maybe I have to come back to this another time.”

Now I can throw it into whatever chatbot and start asking my questions, and I usually do get pretty far. We also, interestingly, I did an episode, as you know, with MacKenzie Price from Alpha School not too long ago. They're the 2-hour-learning folks, and some of the stats that are most widely cited originate with them in terms of compressing the academic day down into 2 hours and still making great progress over the course of a calendar year.

Erik Torenberg

Notably, though, one thing she told me was that they don't use any chatbots in their mix of AI tools. They have an in-house development team that builds some of them, and they have a procurement team that goes out and licenses apps and whatever. They kind of hodgepodge this whole thing together.

So, it's striking to me that what I'm doing, which I find really valuable, isn't even part of their mix. All of that said, I do think a certain amount of humility—and not forcing this on everyone, or making it a one-size-fits-all thing just yet—probably does make sense. I think there's some prudence there.

At the same time, I'm worried when I think about my own kids going through school that if we tell educators, "This isn't really proven yet. It's not for everyone," then you might end up with a majority of people saying, "Well, if it's not for everyone, it's not for me. I don't have to worry about it." Then that's great—I can just keep doing what I'm doing. That doesn't seem to be the right answer.

So, to finally land on a question here: We don't have research to prove everything, but what do you believe about learning? What do you do when you want to learn something new? What do you want your kids to have, given that you've got to make some choices? They're only going to be 7 and 10 for a while, right? What do you want for them right now, while they are 7 and 10, even as the research's final verdict remains out on a lot of these questions?

Johan Falk

Yeah. Well, there are several layers to this. I think for 7- and 10-year-olds—at least 10-year-olds—it's quite possible to use chatbots to learn more, and I agree with your experiences. I, too, have the experience of being able to learn or do more with AI, but when I want to learn something, I often watch a YouTube video.

What do I believe? I think there is tremendous potential in AI when it comes to learning. The ability to converse with an expert on just about anything could accelerate learning in a way that we haven't seen before—unless you're some kind of royalty and can have your own personal expert tutor. This is something we haven't seen before.

But I also think that this would make it really important that kids want to learn stuff, because this is an enabler for someone who wants to learn. I listened to an interview with Salman Khan from Khan Academy, and when he talked about the early effects of Khanmigo, he said that some kids get it right away and just run and learn a lot of things, while some kids are stuck. They're confused, they don't really know what they're doing, and they can't get anywhere.

When they saw this, they talked to the teacher and said, "We're having problems with these kids. We don't really know why they're not getting anywhere." The teacher said, "It's the same thing in the classroom without Khanmigo. You can ask a question: How's it going? Do you need help with anything? And they can't articulate what they're doing, what they need, or what they want."

I think that will become much more important: activating students and making them more active, self-going, self-driven learners. Maybe AI can help with that as well. I'm drifting toward a new question. Sorry.

I think teachers will have a much more important role in motivating and inspiring kids, helping them get started learning, and then letting them learn anything they want using AI for the subject knowledge. Perhaps teachers can also keep things going, because I used to be a math teacher, and I know there is a wide span of math knowledge in every class.

Some kids are understimulated, and some kids are just lost, struggling with negative numbers when they're 16 years old. If you can adapt teaching and learning to their actual level, then I am certain that it will be beneficial for their interest in learning.

My basic answer is that I think AI has tremendous potential for accelerating learning for everyone. That is true in rich countries like Sweden and the United States, but also in poor countries. If you can have an expert tutor for $20 or maybe $2 a month, that's great. That could change so many things.

Erik Torenberg

Yeah, I think the retail price of Khan Academy in the US has been $4 a month, and with a little subsidy or whatever here or there, that's quite accessible. The free version of ChatGPT now is quite generous with its limits. I do think it's well worth keeping in mind that I often quote Biden on this: "Don't compare me to the almighty; compare me to the alternative."

For folks who are in disadvantaged positions and really don't have great alternatives, the AI option becomes a no-brainer in a lot of situations—not just for education, but even for medical questions, which is something I often use it for. Your comment about motivation also definitely resonates.

That was a huge part of the chat with Mackenzie from Alpha School. They're even renaming teachers as mentors, guides, and coaches, and the job description has totally changed. They are not responsible for presenting the content anymore at all, and I don't think they're responsible for grading homework or doing any sort of evaluation. All that stuff is done on the tablet via that sort of hodgepodge of AI-powered systems. The adults in the room are entirely focused on everything else—everything that's not the content.

Are there any other models like that that you're aware of? Like, Vinnie[?]—I mean, she's had a lot of good press recently. Are there any other pioneers that you would point to at the school level who are doing interesting things?

Johan Falk

No, I don't think I've seen any school models like that. I've seen individuals doing a lot of interesting stuff, including parents creating AI tools for their kids to practice or learn things, and just going on learning adventures. But I haven't seen anything organized at the school level.

I listened to your interview about Alpha School and had heard about it earlier. I think it's really inspiring to hear schools using different concepts for learning and reimagining what schools can be, with or without AI. Maybe the closest thing I can find is Khanmigo, and also perhaps Khan Academy without the AI stuff, where you have learning maps and a gamified environment where you collect stars and have spaced repetition and apps, whatever. But it's not organized as a school.

Erik Torenberg

Are there any other highlights from individual parents or families that you think are remarkable enough, or differentiated enough, to merit a mention?

Johan Falk

Well, I have a friend or acquaintance who is an AI wizard. Every second night, he's awake and inventing something new: AI musicals one night, and another time doing AI stuff that could help people with dyslexia. Then he realized this might be good for people who aren't native Swedish speakers, so he started tweaking it. All the code is built by AI, of course.

At some point, I think he also had a tool that could interpret shaky handwritten text caused by Parkinson's disease. I think it was. We need more exploring, and that's also why I think teachers should be free to use AI in the way they think is suitable for their classroom, their students, the whole class, or just individuals.

Then you can see that a student needs to be challenged. He's interested in black holes, or something. Here's a chatbot: Go talk to Gemini or whatever and learn more about black holes. On Thursday, I want you to tell me what you know about black holes, and also write up 3 questions that you think are really difficult but interesting—questions that your classmates would be interested in learning about.

Or it could be that I have a friend—well, a friend’s kid—who was visiting and got tired of us adults talking. He wanted to just play on his phone, and instead of doing that, I got my laptop and started an interactive story for him, adapted to, I think, a 10- or 12-year-old, whatever he was at the time. It started reading, and he was kind of confused: “What’s this?” Well, okay: read it and picture the options. A scenario, a scene, was being described, and then he picked, I don’t know, option 2 and went on.

Half an hour later, his mother said that she hadn’t seen her kid read this closely or intently for so long before. If you experiment, you will find new ways of using AI. But still, I don’t think it’s urgent to start using AI as a learning tool, and this could be a way to switch over to something else because kids are using AI for learning by themselves.

What you do in the classroom is one thing. You can have policies and guidelines, or just say, “No, no AI in this classroom.” But of course, kids will use AI anyway at home, and they will use it for learning or for cheating. Some of them will use AI in a good way that actually enhances their learning, but some of them will fool themselves into believing that they’re learning stuff while they’re actually not. I think that’s a great risk—or one of the greater risks—when it comes to AI in education.

If you use AI in a way that harms your learning too much, then you will get behind in school, and that will lead to accumulating problems. Then you might get your degree and not get a job, or just feel outside society. That is happening right now and has been happening for almost 3 years now. So even though we don’t know ourselves how to do it, we need to teach our kids, our students, how to use AI in a good way for learning, and also how they shouldn’t be using AI for learning, because you might cheat yourself and cause more harm than good.

Erik Torenberg

Yeah. So this is broadly in the bucket of things kids need to be taught about AI, right? Of course, there’s a challenge there, which is that the adults don’t necessarily know it either. This is part of why the taxonomies get tricky, because they all sort of bleed together.

Teachers have a lot of opportunity, as we all do in our white-collar work, to use AI to become more productive. You said that’s not necessarily super urgent; they could do it if they want to, whatever. But presumably, getting hands-on in those daily utility sorts of ways would translate into a much better understanding from which to teach kids what they need to know about AI.

So let’s take a beat on that second bucket of what teachers can be doing, what they should be doing, to get their own time back. We did an episode on this with respect to doctors as well. I was amazed by—I mean, I sort of knew this—but I think it’s similar across medicine and teaching, where you have your day at work and then you have your whole extra night shift.

For doctors, a lot of times it is translating notes, actually doing the paperwork that follows the actual interactions with patients. For teachers, it’s a lot of grading homework and processing all the stuff that got produced by the students during the day. It could have been homework for students as well.

What do you think teachers should be doing? I personally, again, go to AI all the time with my writing and ask for critique, and I usually find it’s at least somewhat good. I don’t necessarily take every suggestion that I get, but I seldom don’t ask. I can say that at this point for sure. What should teachers be doing? And then that obviously informs what they should be teaching the kids about AI itself.

Johan Falk

Yeah. My view on it is that what I tell teachers is: you should use AI in your work outside the classroom if it’s useful, if it actually helps you. But of course, it will help you, because you could use AI, for example, for writing, as you said.

I agree completely with the doctors-and-teachers analogy. Teachers have so much to do outside the actual teaching: taking your notes and turning them into something actually readable, something that you can send to students or parents or your boss, the headmaster, or someone else; going through a lot of information; and understanding, “How should I work with these 3 students who have these challenges? I’ve never met those challenges before.”

You can research and summarize, then spend some time talking to a chatbot about how you could approach this in your classroom and getting some ideas you could try out the next day. That could take 20 minutes instead of 4 hours just reading through research papers. Just finding them previously would have taken a lot of time.

What else? Planning stuff, getting ideas for creating material—that’s another thing. Here’s an old math test. I want to have 3 different versions of this. One of them should be on the theme of soccer. Then I get ideas from the AI that I can’t use straight off sometimes, but this is inspiration for me, and I can adjust it, tweak it, and use it after that.

How about personalizing content for kids? Khanmigo can sort of do this. The Learn mode—or I think it’s Study and Learn mode—from ChatGPT can do this.

Erik Torenberg

Really, you can just prompt pretty much any chatbot, I imagine, to do it. It doesn’t even have to be education-specific in terms of its design.

But I have no idea. It sounds nice to be like, “Oh, you’re interested in basketball, therefore it’ll make all the problems about basketball.” Does that really matter to kids? My sense is that sounds nice. It seems like it would wear off pretty quickly if you like basketball but you’re not that into math. How many problems are you really going to go through because they’re framed in terms of basketball? But maybe there is research there that—

Johan Falk

Yeah. I agree. I think it’s an easy hook to use, but in the long run, it doesn’t work except for very particular types of students. The biggest reward, I think, is the feeling of learning something. When you can get to that, you have something really good going.

When it comes to individualized learning, you could use AI—and I think that’s a great opportunity—to individualize exercises or something. Sometimes you can use basketball or soccer, depending on which student it is, but mostly I would say adjusting to their level instead.

There is a problem: either you need to trust the AI and say, “This is probably good. I’ll just send it off to the students right away without reviewing everything,” or you need to spend a lot of time on it, which teachers don’t have. But we should also be open to the idea that this is 90 or 95% good. Some of these exercises will be bad or won’t really work, but that’s okay. I’ve saved so much time on this that I can instead talk more to my students or follow up in different ways, and the net result is positive.

I think that’s important to keep in mind. It has to be okay to make mistakes, and sometimes it’s okay to have a net negative because we learned something along the way. But a lot of times, what we have becomes not as good as if I had made it myself, while I get so much time to do other things that we get a real boost in learning.

Erik Torenberg

Another dimension of personalization that comes to mind is modalities. This is another thing where I’m not really sure what the research would indicate, but there are at least these concepts of, “I’m a visual learner. I’m an audio learner.” I do feel that myself. Part of the reason I’m in the podcast game as opposed to writing a blog is that I absorb audio content a lot better.

Even in bed at night, I can stay engaged with audio for a long time, but the minute I try to switch over and read something, I very quickly end up going to sleep. It feels real to me. How real is that according to the research? And do we have—I’m thinking NotebookLM-type products here, perhaps—that could even take you from a textbook to a conversational format, even interactive? I mean, NotebookLM is even interactive these days. What do you see there?

Johan Falk

The science—the science when it comes to learning styles—is pretty clear that learning styles don’t exist in that way, sadly. But the feeling of them existing is real, so it’s a really difficult misconception to get rid of.

That being said, I think there are good ways of using different modalities for learning—not because someone is necessarily a visual learner, but because the different modalities can be used in different ways. If you’re on the bus going somewhere, maybe listening to something is great. I can’t come up with an idea of when an image is good, but there are probably places where images are useful as well, or text-based stuff, of course.

Moving between modalities is useful as a tool, and I think the science when it comes to learning styles says that we should blend between them. Don’t quote me on that, but if that’s true, then AI is useful as well, of course.

Erik Torenberg

How about the interactive mode in NotebookLM? It also suggests interactivity as a fundamentally new primitive in education. The chatbots bring that as well, so you can have interactivity across modalities. Is there any established truth about the value of—

Johan Falk

I’m not sure, but I would definitely think that interactivity is good.

I've seen some research on AI improving learning, and they explain it through interactivity. I think that was based in language learning, but the learning process physiologically is connected to dopamine and stuff, and interactivity increases anticipation and stuff, which means that you probably are learning more when stuff is interactive.

Yeah. So, in that way, I think it's fair to say that interactivity is good, and AI increases the possibilities for interactivity tremendously. I think also that that is why—at least I think—listening to 2 people talking about something is more interesting than a monologue. When you listen to the NotebookLM podcast stuff, 2 people talking to each other is a good format for getting engaged.

Erik Torenberg

Yeah, it's coming. It's coming for all of our jobs before too long. Podcasters included.

Nathan Labenz

And truly, I'm not joking when I say that. It's still a minority of my audio-listening time that goes to NotebookLM, but the fact that it can take in whatever paper, a couple papers, whatever it is that I'm immediately focused on, where there's literally no other content out there—again, compared to the alternative for me, the alternative is reading and probably falling asleep reading—so, as imperfect as it is, it is definitely starting to—

Erik Torenberg

Steal some share of my time and attention, and I assume that that's only going to go up and up and up from here.

Nathan Labenz

Have you tried the presentation video stuff at NotebookLM?

Johan Falk

No, I've done audio only, primarily, and a bit of the interactive. So, it makes slides now, too?

Erik Torenberg

Yeah, yeah, it does. It talks to the slides, and I feel—I still, when it comes to AI in education, where I know a lot of things, can do better presentations and better stuff. Maybe I don't do better at presentations, but I have better content in them. I know more about what's important to communicate, but I've seen people on stage, as a teacher and in other cases, who don't do a better job than not.

Nathan Labenz

So, yeah. Wait another 6 months and—yep.

Erik Torenberg

Interesting. So, just finishing up on what teachers should be doing: one thing I think—unless you're going to very much surprise me—is well established as something they shouldn't be doing, which is using AI for grading.

Nathan Labenz

AI detectors, basically.

Erik Torenberg

Oh, well, yeah, grading is another interesting one.

Johan Falk

Yeah. Oh, well, yeah, just first: AI detectors. No, they don't work. Dead end. Using AI for grading is really tricky, and I would say no. If someone asks me again, I would say no again.

Then I would say, well, maybe you could use it as a second opinion. When you've graded your essays or tests, you can run them through an AI and see what the AI says. If there are any big differences, you could have a second look at those. But if you use it as your first assessment, there's a great risk that you will fall asleep at the wheel and just use the AI assessment, which probably will mean that you give some disadvantage to, I don't know, kids who don't have English as their first language or some kind of atypical groups. We don't want that to happen.

Erik Torenberg

Are you aware—I mean, this sort of reliability of AI grading is a huge question in a lot of respects right now, because notably, when you read the technical reports on the new models that are coming out, a lot of the data that they are reporting about the new models is itself AI-generated, right? It's like... I was even involved in a little research that worked this way, where we were trying to assess the coherence, which was one dimension, and alignment of a particular model. The way we were doing that was just doing all these generations, taking all the generations, and feeding them into another LLM for alignment assessment.

Johan Falk

Yeah.

Erik Torenberg

And I'd say it works at the level of major differences in aggregate scores being meaningful, right? So if we have model A and model B, and model A gets a much higher alignment score than B, I would believe that reflects something real going on most of the time, at least. There's also the question of just how consistent or reliable the human raters are, right? Sometimes what you hear, what you see in these technical reports, is, “We validated this strategy by sitting down with some experts and looking at their assessments and comparing our assessments to their assessments,” and kind of talking ourselves into the idea that our AI assessment was similarly good.

Johan Falk

Yeah.

Nathan Labenz

So, we're going to be judged by AI before too long. Is there any way that that's not going to happen?

Johan Falk

No. No, it's not. I think we'll need the human evaluations to validate the AI assessments to make sure that they're on the right track, and 1 or 2 generations down from that, we won't have the competence to do the human evaluations anymore.

When I worked with the National Agency for Education, I was for a few years responsible for national tests in mathematics. At that time, we had oral tests; part of the national tests were oral in mathematics. We did some evaluations, and the results, when it comes to how reliable human evaluators were, was that they were basically like tossing dice, and sometimes worse.

In those cases, it would make sense to use AI assessment as well instead. Maybe the problem with AI assessment, even if we can have more consistency with AI assessment, is that we have a single LLM doing all the assessments. That means that whatever biases you have in that LLM will affect all the assessments.

When you have 1,000 teachers assessing instead, you at least have some noise, making it less probable that some groups of students are disadvantaged. On the other hand, you can, of course, find these biases for LLMs—at least when you find them—and you can adjust for them, which is much more difficult when it comes to humans. But, yeah, it's tricky.

I think if you want, as a teacher, you shouldn't, on your own, just start using a chatbot for grading students' stuff. It requires quite a bit of a framework and tests and stuff to have it reliable enough.

Erik Torenberg

Well, let's say I'm a teacher and I come to you and I say, “All right, I heard your warning, but I'm too busy. I'm going to do it anyway. So, I'm going to do it now. You tell me how to do it as well as I possibly can.”

It sounds like a couple ideas that have come to my mind just while listening to you talk about the challenges are, 1, maybe use multiple different LLMs; maybe use different prompts. I mean, this is sort of in keeping with the general trend toward scaling inference compute, right? Instead of grading the thing 1 way, grade it 5 different ways, and then you could maybe have some sort of resolution idea where, if they all agree, you go with it.

If there's a 4-to-1 vote, maybe you go with it. If it's 3-to-2, maybe you have to do it yourself. If you see any grades that are more than 1 grade apart—if you're on a 7-point scale and you see any evaluations that are 2 points apart—you have to go in and read that one yourself as a human.

Johan Falk

Yeah.

Nathan Labenz

I'm not going to listen if you say no. So, what else would you tell me to try to get it?

Johan Falk

What I would start by saying is, I definitely understand your needs. You need more time, and so let's try to do this in a good way. And, by the way, if you do this properly, you can start a company and make a lot of money off this.

I think I would say break down your assessment into different scales. So, if you're assessing essays on—I don't know—English, just writing skills, make sure that you have 4 different scales that you're using: typos, using rich vocabulary, I don't know, different ways. I'm not a language teacher, so I don't know these things, but there are different aspects that you use when you grade or assess essays.

Erik Torenberg

A rubric, in short.

Johan Falk

Yeah, yeah, and have these also as kind of numeric scales—A to F, 1 to 10, something—and then have some kind of method of composing the results into a final grade. Because that means that you can check afterward, or the student can check, or their parents can check: this LLM, or the assessment, says that this was bad in the essay. We don't think this is bad. It's a 3; it should be a 7. Then you can look at that and see, “Yeah, you're right. We should change this.”

That's much easier to do than the AI says it's a 5 in general. That's 1.

Second, you should try to identify groups that could be disadvantaged. So, you have non-native English speakers, English as a second language. You probably want to separate boys and girls to see if any of these groups are advantaged. You might have some other categories as well.

Then you just look at what your assessment says for these different groups and see that the results are okay. I mean, it could be that some groups are actually performing worse than others, but it should be proportionally so when the AI assesses them.

And then you should also have some way of complaining or saying that this is wrong. So, a student or their parents can say, “I want you to have a second look at this.” You should also be transparent about this being assessed by an AI.

If you do all of these things, I actually also think that you have complied with the EU AI Act, and then you can start selling this in the European Union.

Speaker 1

Okay.

Speaker 2

On the other hand, it might take more time than you had as a teacher. Well, once you get to there—

Speaker 1

Yeah, you go to Lovable and get it going. Yeah. Okay, cool. Coming back, then, to part 3, or things that schools and teachers need to be thinking about teaching their students about AI itself.

Erik Torenberg

What do they need to be teaching students about AI itself?

Johan Falk

Yeah. All right. So, I've already mentioned learning to use AI for learning. That is an AI competence in itself. Even if you don't use it in the classroom, you need to teach kids how to use or not use AI for learning. And that is one of the really urgent things, I think.

Another really urgent thing is helping students and kids understand AI friends or AI companions. We've probably both heard of really sad cases where AI companions have caused severe harm, and there are risks—not just suicides and things like that, but also emotional and social harms—that we should avoid.

Personally, I think that AI companions should be prohibited for anyone below 18. You don't have to go that far, but you need to help students—well, just discuss with kids what AI friendships are. How should we relate to them? What are some warning signs? When should you be worried? When shouldn't you be worried?

Because even though I'm against AI friends for kids, I'm pretty sure that 90% or 95% involve no worries at all. But you should be aware of what to look out for when it comes to AI friends. So, those are urgent.

We need more skills. Deepfakes and critical thinking are really important, though not as urgent as the other 2 I mentioned. And then we have general skills. Some of them are the same things that teachers should learn, like using AI for writing, managing information, and automating stuff. Probably more that I don't think of right now.

But then we have so much more that isn't the practical use of AI tools. It's understanding how AI is affecting democracy, how it's affecting the balance of power in the world or in society, the concentration of power, the effects on the labor market, and how rapidly AI is evolving.

It's learning the basics about AI as a technology, because that helps you understand what to expect from AI and what not to expect, as well as more legal and ethical aspects. I feel I could go on for quite some time. I'm writing a book about this.

Nathan Labenz

Yeah, this could easily be like an hour out of every day to cover all that.

Johan Falk

Yeah, it could be. But also, I think AI competence should be taught from preschool to adult education, and it affects most subjects. In many cases, it's a natural part of the subject, I think.

In some other cases, it's new content that you need to add. Writing, for example: if you have English writing or creative writing, then you probably teach feedback processes, stuff like that. You can incorporate AI into that as well, and you can use AI to get feedback on your text.

But you can also go a step further and create a panel of readers—typical readers—and have an AI mimic those and give you feedback based on what, I don't know, middle-aged white men think about this text. Things like that. So, sometimes it's just a small tweak.

In other cases, if you're teaching programming, then you need to bring in AI tools and help students learn to code with AI. And that is probably more of a shift than you have in writing.

Erik Torenberg

Going back to the AI friends thing for a second, what are the challenges? I always—and I do worry about this for my kids. They're a little young for it now, but I'm expecting probably this year, if not this year, it'll have to be next year, that we're going to start to get toys floating around that are kind of AI friends.

There are some already on the market, of course, but I haven't been asked for one, which I think means my friends' kids don't have any, which means they're just not that popular yet. And I'm sort of like, man, I agree intuitively with the precautionary principle here. I don't necessarily want to run some crazy AI girlfriend experiment on my hypothetical teenage son or whatever.

My kid's only 6, but if I project forward a few years, when he's not so basic, I don't think I want his first girlfriend to be an AI girlfriend. That just seems like too much, maybe.

Nathan Labenz

But then again, I'm also like, well, this tutoring thing has so much upside. What makes a good tutor, right? It's obviously a lot of things. It's having the knowledge and the skills to impart, but it's also building the relationship and the rapport, having fun together, and making it a time and a way to spend time that you look forward to.

So, I think it seems like the line between an AI tutor and an AI friend itself is going to get really blurry. I always think back to Eugenia Kuyda, who started Replika. She was actually one of the very first guests on this podcast, and she said a number of really interesting things.

One of which was that she thinks the moat in AI applications is ultimately going to be relationship. Meaning, for most people, you don't switch friends. You don't abandon your friends because you meet a smarter person than your current friend. It's the history that you have together. It's the experiences you've shared.

It's all this intertwinement that you have with a friend that makes them your friend, and not the person who might be smarter and might be better at some things. Maybe you want to be friends with them too, but you don't abandon your friend.

And I thought about that for a long time. It's been over 2 years now since she originally said that. So, I'm like, boy, this is definitely playing with fire, but it's kind of hard to imagine the best available AI tutor wouldn't have a lot of these elements.

Erik Torenberg

So, I don't know—how do we make sense of that, right? It's tough.

Johan Falk

Yeah. It's really tough. When it comes to the science of teaching and learning, it says that the relationship between the teacher and the students is one of the most important factors when it comes to learning. So, having a relationship between the student and the AI tutor will probably be important too.

I don't know how to find a good way around that. And that means that we should tread carefully, I guess. On the other hand, we also have the competitive environment, meaning that those who move more quickly will get more benefits.

Maybe one way of at least reducing the risk significantly is to look at the incentives. If the incentive is to get as much money as possible from people who are buying your services and paying a monthly fee, then you will have more dark patterns that make people stay stuck with your stuff, more or less against their will.

If the incentives are to get as good an education as possible and this is run by a nonprofit entity, then you have better chances of not getting those dark patterns and dependencies. But, I mean, we still lack the map for navigating this terrain.

Nathan Labenz

Yeah. And there's—I mean, not only do you have to worry about potential dark patterns from people who are trying to get you to subscribe, there's also the ad-driven model and the commerce-driven model, which is starting to take shape right now as well.

So, I feel like, as much as I don't want to get hooked in, I don't want my kid to have some sort of weird parasocial relationship with an AI that's designed to make me not churn off the subscription. I'm maybe even more reluctant to think about the free version, where it's trying to get them to be excited about buying whatever, or just maximizing engagement, paid by ads.

Johan Falk

Paid by ads. Yeah.

Nathan Labenz

There's a lot of weird stuff, and I'm not a hater. We just did an episode on advertising in AI apps, and I do think it has its place. It is easy to demonize that whole idea prematurely, too. But, yeah, I wouldn't—I'm definitely thinking, "Oh, God, I don't want to have my young kids be the subjects in that experiment."

We got a lot of questions. Okay. How about systemic changes? Well, maybe—is there more? I mean, you covered a lot there on things that we should be teaching kids about AI. This is sort of what we do. Can we try to get the kids to understand the most important parts of it?

Maybe misconceptions are something that I think is really interesting at the moment, especially because I'm thinking about this: What do I want to communicate to educators here in my home state coming up? One misconception that I hear a lot these days is the idea that hallucinations are still as frequent and severe as they were, with a big emphasis on hallucinations, which obviously do still exist, although I would say both their frequency and severity have come way down relative to when the hallucination narrative was formed.

So, in today's world, I'm not really even sure that the output you get from GPT-5 or Claude 4 or whatever is less accurate than what you would get off of Wikipedia.

Johan Falk

Yeah. Certainly not when it comes to stuff that you teach in elementary school.

Nathan Labenz

Yeah.

Erik Torenberg

It's going to stray—

Nathan Labenz

Pretty far out on the fringes. Yeah.

Johan Falk

Yeah. That's one misconception. I think one common—I don't know if it's a misconception, more like a missed opportunity—is just using chatbots as a search engine.

It took me too long to realize that a lot of people use ChatGPT, for example, and type in "vegan pancake recipe" and get a recipe for vegan pancakes back. They're thinking they're using Google, but from OpenAI, and instead they're missing out on using AI or chatbots as an assistant or someone to discuss stuff with, or using them as a problem solver, things like that.

That is something I think is worth showing people: you can do so much more with AI tools. Well, we have the cheat detectors, AI detectors—worth pointing out.

Nathan Labenz

What else? Misconceptions? I think, well, it's impossible, I guess, even for you, to stay up to date with what AI actually can do. Basing your assessment of what AI can do on experiences that are 6 or 12 months old is a huge risk of being wrong because things are changing so quickly.

Another one I had in mind was the idea that LLMs are just predicting the next token, which I think has kind of become wrong in 2 ways. When people say that, they sort of have—and they probably haven't even heard this term—essentially the stochastic parrot model of a language model in mind. There are 2 things I want people to know. First, even when the model was, in a very literal sense, just trained to predict the next token, that doesn't mean—and in fact, we have very good evidence to the contrary—that they don't have any higher-order conceptual understanding, right?

If you're just predicting the next token, you wouldn't expect things like a language-independent representation of certain concepts, right? But we do know now that at least common concepts seem to be represented in a way that's detached from the English word for that concept, or the Swedish word for that concept, or whatever, right? These concepts that exist across humanity have representations that are higher-order than the specific tokens or words that are used to represent them. It looks in that way a lot more like thinking, right?

Johan Falk

Yeah.

Nathan Labenz

I don't think primarily in tokens. I have to kind of cash my thoughts out to tokens, but there's something going on inside that is higher-order processing. It's not—you know, it is a lossy process to reduce that to a single token. So I think people really fail to understand that.

And then another one that's just even more literal, but really important too, is they're no longer being trained exclusively to predict the next token. Now they are being trained to get the right answer. The whole reinforcement learning paradigm doesn't really care what tokens you strung together to get to the right answer. In many cases now, you are just judged—as the language model—on whether you got the right answer or not. The reward is based on that final outcome, not on the token-by-token chain of thought that you took to get there.

This is also, I think, really quite well demonstrated by some of the recent chain-of-thought visibility we've gained through research from folks like Apollo Research. That's another recent episode where they got access to the chain of thought the public doesn't see from the OpenAI models. What they found inside these chains of thought is that this is becoming its own dialect: the AI is now kind of speaking in weird shorthand, using terms in nonstandard ways. It's like just very—

Johan Falk

Flam, disflam, watchers. Yeah.

Nathan Labenz

That is not predicting the next token in any text corpus I've ever seen. So whatever's going on there, it's definitely a different kind of thing. Anyway, I don't know. Maybe I'm just rehearsing part of my upcoming talk here, but those really stand out to me.

Why does that matter? The biggest reason is that I think it leads people to underestimate where the technology is really at. From my perspective, not just for education but society-wide, one of the biggest mistakes we could make would be to underestimate how good this stuff is—how powerful. We can leave good and bad aside for a second, but just how powerful it's become.

Johan Falk

Powerful, yeah.

Nathan Labenz

We're not doing anybody any favors if we allow them to comfort themselves or figure that they don't need to worry about this because it's not that good. Again, for so many questions, it comes back to whether people are comforting themselves or convincing themselves that they don't need to worry about this because it's just predicting the next token and it makes so many mistakes anyway.

I think those things have kind of become dangerous memes that have outlived their usefulness. It is still important to know that you're not going to get 100% infallible accuracy from AIs either. It does feel to me like that has swung in the other direction, where people are underestimating their capability rather than—originally, they were perhaps overestimating it and needed to be corrected, but now it seems like it has swung the other way.

Johan Falk

Yeah, yeah. I agree that is an important message. Perhaps in the same vein, I think it's also important to point out that AI models are not traditional software. You can't look at the code and see what kind of decisions are being made. We can, with a lot of effort, understand small parts of what's going on inside, but they're essentially black boxes grown rather than built.

Nathan Labenz

Yeah, yeah. I'm taking a note on that as well. Okay, how about on the systemic changes front? That could go a lot of different directions. I have a couple of candidate ideas that I want to throw at you, but what's top of mind for you in terms of what people need to be preparing themselves for?

Johan Falk

I have such a mix of stuff that I don't know where to start. AI tutors are probably one of the most salient things that you could look at. If we start having stuff like Khanmigo actually working at full scale, it means that the role of the teachers will change. They will no longer be teaching subject knowledge, but will be more like mentors, as at Alpha School. That paradigm will become more common, which is a huge shock to the educational system.

We have the labor market. What is demanded from schools and education will change when the labor market changes in different ways. It could be that the demand for, I don't know, accountants will go down heavily. Okay, then we need to shift the balance between different kinds of education.

We also have entry-level jobs disappearing. That means maybe we should have 2 more years in school before you start working. Maybe we should have some kind of mentoring program—more mentoring or trainee programs at work—that is sponsored by the state or something. We have potential mass unemployment in whole sectors, and that would call for reskilling, which requires mobilizing education in different ways.

What else? Well, taking it all the way to UBI and stuff, that will change how education works. We have stuff like, what is knowledge? What is being human? We're kind of used to being the only entities on this planet that are able to think in some kind of higher order, and that has changed already, I would argue, and seems to be changing more in the next few years. That will change how we view learning and understanding, stuff like that, which will affect education, of course.

Nathan Labenz

Yeah, and there are so many things. You had a recent episode with Emad Mostaque, where he talked about the economy going to really strange places, which I think could be a low-probability outcome for that particular situation or scenario coming true. But we have to plan even if we gave it merely a 5% chance that some radical things will change in the next 5 or 10 years.

We need to plan ahead because the educational system is so slow. It's a huge ship that we need to turn around in some way, and we need to start planning now. It's just looking at different scenarios, seeing what might happen.

So that's kind of a motivation or an argument for one of the ideas that I wanted to get your take on, which is basically: will we see the end of standardization? You sort of alluded to that a little bit with some teachers who may or may not want to use AI in different ways. Don't make it mandatory.

I'm kind of wondering, even just across the board, should we be thinking that standardized education is a thing of the past, even to the point of standardized tests maybe going away? When I spoke to Mackenzie from Alpha School, I was struck that she was kind of one foot in 2 paradigms, right? She's pushing the AI-based content delivery and all that stuff as far as she can, but she's still holding herself and her schools accountable for results on standardized tests.

I was kind of like, well, that's interesting, because I sort of get it in the sense that that's what the rest of the world understands right now. If for no other reason, she's kind of got to prove as a pioneer that this can work on terms that everybody else accepts as valid. But it also struck me that the whole thing is kind of superfluous relative to the depth of understanding. That's one snapshot, one morning, one set of problems, whatever, compared to the depth of understanding that her AI system has about the students. It seems like it's a pretty limited signal, right?

So this also connects to being judged by AIs, but it seems like in the future, we will probably not need to sit for one test one time and get a score. Rather, whether we're in school or potentially just doing my thing, I can imagine turning on a recorder of some sort.

I've previously hypothesized that people should be paying me to watch me use my computer for reinforcement-learning data purposes. I'm still waiting for those offers, people. I hear they're starting to materialize in some places. But another version of that would be just assessment, right? Instead of me doing a test, maybe I install some software, even as a professional, and it just kind of watches what I do. Then it can, a week later or whatever, be like, “Okay, based on 40 hours of this dude's computer use, here's what we can tell you about him.”

Erik Torenberg

And I would expect that would be, in many cases, really valuable signal with a lot of texture to it that you wouldn't get from people sitting and doing multiple-choice tests. So, yeah, how far do you think this sort of—

Erik Torenberg

Educational system goes?

Johan Falk

I think that what you say makes total sense. If we have an AI that monitors what I do, what I learn, what problems I have, and what I'm excelling at, it makes no sense to reduce that to a letter grade from A to F. It's good because it fits with current metrics, but otherwise it's just insane.

And I think this is, weirdly, perhaps one of the most radical ideas I have, but I think that the age of grades is coming to an end. Grades are something that—well, cheating is a problem in particular when you have AI in the picture, because it's so easy to produce stuff that looks good but doesn't reflect your knowledge. This is only a problem when you're actually going for the grade, when the grade or the exam is the thing motivating you.

If you're motivated by learning, then AI is basically nothing but a great tool, not a problem. And I think we need to make that transition from having an educational system largely driven by grades to an educational system largely driven by the will, the desire, to learn. I don't know how to make that transition, but I think getting rid of grades would be a big kick in the right direction.

That would hurt a lot, but I think it would also, in the end, be very beneficial. And, well, not everything will be better. There will be new problems that you have to solve in some way. You want to go to university, which means you have to have some kind of admission test, or you'll just have random access. I don't know. There are other things, but I think it becomes more and more problematic to have grades as the motivator for kids to learn stuff.

It might already be more harmful than good, but it will get worse for every 6 months that passes, I think. From a wider perspective, this current educational system was built for the First Industrial Revolution, and that was almost 200 years ago. I'm not sure that the same system is fit for the fourth—or whatever it is—Industrial Revolution, with AI going on now.

Fifth, I don't know. I'm into math, not social sciences or history. But that's a really big perspective.

Erik Torenberg

Do we have any alternatives? I'll throw one at you, and you can tell me what else you've seen in the wild. I guess this isn't exactly an alternative to grades, but in terms of what would motivate people, maybe I should start with what I think demotivates people.

What seems highly demotivating is: I'm being made to learn these things, I'm being made to do this work, and I'm then getting a grade on it. And the whole time I know that, no matter what I do, I'll probably never be as good as AI at doing this thing. So that just plain sucks from the beginning, right?

What can we do that is inherently not something AI is going to be better than you at? I think when it comes to things framed in terms of economic contribution, that's potentially a vanishing set. But one thing that could still be inherently not something AI can do for you, and also might be motivating, is figuring out and expressing clearly what you really think.

I've been toying with this idea: What if we reframed writing assignments not as a minimum word length or a minimum—it's got to be a 5-page essay, or a 5-paragraph essay, or a 500-word essay, whatever? Usually those are framed as the minimum in schools, at least in my experience.

What if we instead put caps, and potentially low caps—tweet-length caps? Sometimes I find my own thinking is most pressed when I have a feeling or a thought or an impulse on something and I want to tweet about it, but now I have to express myself in that limited space. Maybe I'm just precious about this, but I feel like I want to put something out that I fully stand behind.

Could a writing assignment be reframed from showing that you can write in an expository form to asking: Can you put something forward that you could read to the entire class in 15 seconds, perhaps, but which you fully stand behind?

Johan Falk

Yeah. I think exercises like that are really valuable—trying out new forms and new ways of doing things. Another thing I've been thinking about is writing about something that you really care about: your favorite hockey team, your favorite band, or something.

You could have an AI write that for you, but you won't be satisfied with the first output that you get, because there are so many aspects that you think are important about this hockey team that you want to get into the text. And that is kind of the same thing as you're saying: expressing yourself becomes the important part. I guess that is one path to having something that you are better at than the AI.

On the other hand, I'm not sure that is the important part—that you should be better at something—because you might want to play a musical instrument. You will never be as good as the professionals, but you'll still enjoy playing that instrument eventually. You're not doing it for excelling at something. You're doing it because you want to develop, because you like the feeling of being able to do more, to express yourself better, and things like that—growing as a person. Maybe that is a way forward. I don't know.

Erik Torenberg

Yeah, I think these things—I like the idea of just something you really care about. And I do agree that there's—I have, I'm thinking of my one friend, Chris, who's a real AI whisperer. It reminds me of what you said about your friend earlier: just constantly creating new things. He's of the mind, and at times coaches me, like, “Dude, nobody really cares whether you did it or the AI did it.” If the AI is as good as you—and it probably is—it doesn't really matter.

And I think that in many ways that's true, from the perspective of the person consuming the output. That probably is true more often than not. But if it's something I care about—if I'm asking, “Is this what I want to put my name on? Is this what I really want to project into the world? Is this what I think other people should seriously consider?”—then it really does matter to me.

I do end up toiling over my intro essays for the podcast, probably well in excess of what I need to, because I feel like it's me.

Nathan Labenz

The AI might be able to do a better job of analysis, but if I have anything to offer, it's got to be something that's at least sincerely felt. That's one thing I can—and I still definitely can, I mean, for no surprise—but at the end of the day, I can still articulate something that I feel I want to—I'm prepared to stand behind—better than I can get an AI to do that for me.

If that stops being the case—if the AI gets to the point where it can read all my writing and write something that's, “I literally couldn't have said that better myself,” or that captures my perspective on the matter as well as I could ever have hoped to articulate it—then we're in a really weird world. We're already getting into a weird world, but that's going to be extremely weird.

Johan Falk

Yeah.

Nathan Labenz

There were some really weird results in a blind, randomized blind test for the Turing test, where people—I don't know if it was 60% or 67% of the time—thought that the AI was the human. The maximum level of that test should be 50/50, but in some way the AI is more human than human. I mean, what's that? That means we should be able to tell which is the AI by who is more human.

Johan Falk

Yeah.

Erik Torenberg

Yeah. I don't have any good answers for that one.

Nathan Labenz

Yeah. Here's another idea in terms of just the purpose of education, or rethinking some fundamental assumptions. I suspect this one might be more controversial in Sweden than it would be in the US, although it wouldn't be without controversy here by any means.

But again, working from the premise that maybe—I think already, right?—we've seen the GDPval from OpenAI in the last few days that showed, notably, Claude Opus 4.1 at the highest level was almost at parity in terms of how effective it was at basically doing domain-work tasks created by experts. Then Opus and an actual human expert do the task, and another set of experts evaluate the task. Opus is almost at 50% in terms of how often the evaluating expert prefers its output to the human expert.

So, you don't have to go too much farther there. Then we're in this world that we've been envisioning, where most people are probably not going to be able to make an irreplaceable contribution to the economy. And so maybe, at least for a time—and I feel pretty confident that we're going to get there—what I don't know, although I certainly don't rule it out either, is whether the AIs will begin to deliver paradigm shifts for us at some point. Will there be a sort of Einstein-level contribution from AI that reframes things that we thought we knew, but in fact there's a whole new paradigm that both explains the previous paradigm and unlocks a new depth of understanding?

Honestly, I think probably the AIs will get there, but that's much less certain. So, anyway, for the time that we're in this zone where the AIs are competitive with, if not better than, our sort of mass expert culture, but maybe not yet at the truly revolutionary, Einstein-level stage, one thing you might think of education as serving to do is identify and cultivate uniquely special genius. In other words, there's an Einstein in your country right now somewhere for probably 10 or 20 different things. Can we figure out who those people are?

Erik Torenberg

Identifying the top 3% that really could excel at something?

Nathan Labenz

Yeah, people who actually could change things.

Erik Torenberg

Well, maybe we still need them if the AI can't deliver that kind of paradigm shift.

Johan Falk

Yeah. All right. But we're doing pretty okay right now without focusing on the top 3% exclusively. So I don't think we as a society need them unless we feel like, oh, the aliens are coming and we need to mobilize or something. I think it's an interesting idea, and I think it's worth exploring because we should explore many different ideas, but I don't see that society will invest in that.

I think from a societal level, if you want the GDP or productivity from these, say, 3% of people, then it's a huge investment to run the full educational system to find them. On the other hand, you could of course use some screening, but the alternative is to wait 5 more years and then you have AIs doing that stuff as well. And I think we're still, from a societal viewpoint, left with a conundrum of what to do with the 97% of people, and that's something we need to solve. Should they have their UBI and be happy?

How can we do that in a way that they don't feel that they're just unemployed and useless, but instead feel like they're having a nice retirement and spending their lives in a good way?

Erik Torenberg

How worried are you about that? My perhaps naive sense is—the way I put it to Jake Sullivan was—if the political class can figure out the international relations, I believe in the working class's ability to spend the peace dividend. In other words, I think people will probably figure it out.

Johan Falk

Yeah. Me too. If I'm worried, it is that a small minority of people will gather all the resources and the others will be left without the UBI. If I were unemployed today and provided for by the government or something, I would be happy. And I have a job that I find really interesting, but I would love to just play board games with my friends, hang out with my family, do stuff, and read books. Ninety percent of people don't have that, I think. So, yeah, I'm not worried about people complaining because they don't have jobs.

Nathan Labenz

Yeah. How about any other habits of mind you think people in the education world should be adopting or cultivating? One, I'll give a shout-out to my longtime teammate Matt Kall from Waymark, who is the son of 2 teachers. He's the technology lead at our company, and he's made one of his mantras “teach and learn every day.”

What I think is interesting about that, and maybe generalizable, is the idea that we all have a lot to learn. And when it comes to AI, we're all on the same timeline. We may be different ages, but we're all experiencing these advances and these changes to our reality at the same time, regardless of what age we are and regardless of what we have or haven't otherwise experienced in life.

So, I feel like if I wanted to give one bit of advice to teachers, it might be that they should just embrace the idea that they need to be learning right alongside the students and being hands-on—doing their experiments, making their mistakes, sharing all that stuff with the students, and showing that this is an active learning process for me and, therefore, of course, it's going to be the same for you.

Johan Falk

Yeah, I agree. And one way that I phrase this for teachers and principals is: make more mistakes. We need to raise the number of mistakes that we make because we need to learn and use AI at the same time. This technology is important, and it's evolving so fast that we need to play with it.

That means that, just as we allow our students to make mistakes while learning, we should apply that to ourselves also. We need to identify the most costly mistakes, but as long as we keep those in mind, we should celebrate our own mistakes because it's fine to make mistakes. It means that we're learning.

Erik Torenberg

Any other habits of mind, best practices, all-star standout examples, or approaches that we should highlight?

Johan Falk

Yeah. Well, another mantra of mine is that everyone, not least teachers, should find a way to use AI to save 5 minutes every day. The best way of learning AI—learning about AI—is to use it. Everyone can find a way to save 5 minutes a day using AI.

If you do that, you will learn more about the technology, how you can use it, what it means for the world, what it means for students, how you could use it in your teaching, what your students should learn about AI, and stuff like that, while still saving time. So, use AI to save 5 minutes a day.

Erik Torenberg

How about, let's say, reading assignments you think should be more emphasized? These could be things the teacher should read or the students should read, or maybe we should all read them together—like science fiction, visions for a positive future, or just anything that you think is really useful cultural capital to bring into the AI era?

Johan Falk

Wow. I don't know. Science fiction is a good thing. Well, science fiction is not lightsabers and stuff. Science fiction is picking an idea, or several ideas, about the future: what would happen to the world or society if this was different or that was different?

That kind of science fiction is a great way of expanding your views of how things might be. Sometimes people say, “Well, this or that happening, that's just science fiction,” which to me is a weird thing to say. A lot of science fiction stuff is present right now in our world.

Erik Torenberg

Yeah.

Johan Falk

So, reading more science fiction or watching science fiction is a good way to expand your views when it comes to AI. Otherwise, I don't know. But being curious is great also when it comes to literature.

Erik Torenberg

Another great assignment, I think, is just trying to get the kids to do some of that future-vision work. I don't necessarily expect the hit rate on great literature coming out of the typical classroom to be very high, but if only as a reminder to people that, at least for now, we still have agency to think about what the future could look like, think about what we want it to look like, and actually try to steer it in that direction.

With so much happening all around us, and so much feeling like it's happening to us or happening by some exponential process that's got a logic or momentum or a life of its own, it does seem like a useful mindset to cultivate. Just imagine a future. Imagine something different than what exists right now.

Johan Falk

We have a say.

Erik Torenberg

Yeah.

Nathan Labenz

Yeah. Yeah. Yeah. Maybe you should be a teacher. That's a good idea.

Erik Torenberg

That's kind of what I'm trying to do. To a degree, that's what I'm trying to do with this podcast, very much with a teach-and-learn emphasis on learning, no doubt.

Nathan Labenz

Yeah.

Erik Torenberg

I think we've covered all the bullet points that I had outlined, and I appreciate all your time. Is there anything that we haven't talked about that you would want to make sure we cover? Any other thoughts to leave people with?

Johan Falk

No, I think we covered things pretty well. It's been a fun conversation. It's not often I have this opportunity to geek out on AI in education at this length.

Yeah. Well, I think teachers might feel overwhelmed: “What am I supposed to do? Everything is going so fast. I don't understand what AI is.” What teachers should do is focus on their students. That's the role of the teacher.

When it comes to AI, that means that you should learn enough about AI to understand: Is this something that your students should learn about? Should you teach your students in your subject, and if so, how should you do it? That's the main responsibility of teachers when it comes to AI.

Apart from that, principals and school owners need to invest in AI competence for their teachers' professional development. And from a strategic level—national level, perhaps school districts—we need to look ahead and see what's going to happen in 3, 5, or 10 years. How can we prepare for that? That is a big challenge.

But as a teacher, try to relax a bit, focus on your students, and find a way to save 5 minutes a day using AI.

Erik Torenberg

I love it.

Johan Falk

All right. So, I’m starting something called Graspable AI, with short videos aimed at teachers and others working in schools, trying to explain what AI is, what it means for education and society, and what it means for our students.

That’s a place where you can find more about this.

Erik Torenberg

So, Graspable AI.

Johan Falk

Yeah.

Erik Torenberg

Tell me the URL again one more time.

Johan Falk

Graspable.ai.

Erik Torenberg

Graspable.ai. All right, cool.

Johan Falk

Yep.

Erik Torenberg

I got a chance to preview some of these things, but I hadn’t actually been to the URL, so that’s great.

Cool. Well, Johan Falk, Graspable AI, thank you for being part of The Cognitive Revolution.

Johan Falk

Thank you.