This is a fascinating meeting so I have just asked ChatGPT to summarize it without much modification. It is a long post, but definitely worth reading.

At our most recent Courageous Minority meeting, we returned to a question that has become increasingly difficult to avoid:

If AI can do so much, what should schools teach—and what should schools become?

The question is becoming more urgent because AI is changing faster than most schools can respond. In the opening of the meeting, I shared a recent experience in Vietnam, where I was invited to discuss the design of a new school. My suggestion was simple: if you are building a new school today, do not build one that will look outdated in five years. Build a school born in the age of AI.

That became the framing question for our conversation. What would such a school look like? What would students learn? What should teachers do? What should AI do? And perhaps most importantly, what should remain distinctly human?

I also acknowledged that I do not have a simple answer. What seems increasingly clear to me is that much of what schools have traditionally taught no longer makes the same sense when AI can perform so many of the cognitive tasks around which schooling has been organized. The purpose of the Courageous Minority is not to claim that we already know the future. It is to create spaces where educators can experiment with possibilities rather than wait for entire systems to change.

“The curriculum is dead”

Patrick opened the discussion with a striking description of his elementary classroom. His students, he said, can sometimes look as though they are sitting in a classroom from the 1960s—pencils in hand, barriers around them to prevent cheating, completing prescribed assessments because the curriculum requires them.

He contrasted those moments with periods of free exploration. When students are allowed to build board games, work with materials, explore questions, and simply be children, “they light up.” His conclusion was provocative: the curriculum is dead.

His frustration was not directed simply at one curriculum or one jurisdiction. It reflected a broader tension many educators are experiencing. Schools continue to require students to master prescribed content even as the conditions that originally justified that curriculum are changing dramatically.

The discussion then turned briefly to international testing. I argued that systems such as PISA increasingly tell us little about what young people need in a rapidly changing world. Governments continue to use test results because the numbers provide convenient measures of success and failure, but the deeper question is whether improving performance on measures built around an older educational paradigm actually prepares students for the world they are entering.

The growing backlash against AI

Several Australian colleagues then described something happening in the opposite direction: rather than rethinking schooling because of AI, governments are responding by restricting technology.

Peter described new policies in New South Wales that limit screen time and reduce the use of take-home tasks. Others added similar developments in Victoria and concerns that other states may follow. Participants questioned the evidentiary basis for rules such as particular percentages of allowable screen time or fixed numbers of minutes for younger children.

The concern was not that all technology use is necessarily good. Rather, it was that AI, social media, smartphones, and screen time are often being bundled together as one problem. The resulting response can easily become: remove the technology and return to pen and paper.

One participant compared the situation to prohibition: publicly, institutions restrict or condemn the technology while privately nearly everyone continues to use it.

This produces an increasingly difficult environment for teachers who want to experiment. Even in progressive schools, participants described AI fatigue. Teachers hear conflicting messages from governments, parents, media, technology companies, and educational leaders. Some respond by retreating toward familiar practices because those practices feel safer.

The discussion raised an important implementation problem. A relatively small group of educators may recognize the need for transformation, but most teachers still work inside systems organized around compliance, prescribed curricula, standardized assessments, age-based progression, and accountability requirements. How far can educators move when the system itself continues to reward the old model?

That question remained with us throughout the meeting.

Does AI make people lazy?

AI is frequently criticized for making students—and adults—lazy. I think the criticism contains an important insight, but perhaps for the wrong reason.

If schools continue asking students to perform tasks that AI can easily perform for them, then of course students may use AI to avoid doing those tasks. If an assignment asks for something ChatGPT can produce in seconds, students will eventually recognize the absurdity of spending hours producing it manually.

But there is another way to think about “laziness.”

Much of technology has been invented precisely to allow human beings to do less. We adopted calculators so we would not have to perform every calculation manually. We adopted washing machines so we would not have to wash clothes by hand. We adopted search engines so we would not have to remember where every piece of information was located.

Technology frees human capacity.

So the educational question should not be:

How do we force students to keep doing what technology can now do?

The more interesting question is:

If AI allows students to do less of one kind of cognitive work, what can they now do more of?

I referred to this during the meeting as leftover cognitive capacity. If AI takes over some forms of routine cognitive labor, education should help young people redirect their attention and energy toward activities that remain meaningful: creating, questioning, judging, collaborating, finding problems, developing relationships, and contributing value to others.

From using AI to redesigning learning

The second half of the meeting moved from discussing the problem to exploring one possible response.

Ruojun and Wanyu introduced the work we have been doing with YEE AI, an educational system developed around the principles that have guided much of our work: personalizable learning, problem finding and solving, and human interdependence.

I have to admit that I was initially skeptical about building a separate AI system. My reaction was: why not simply let students and teachers use powerful general-purpose large language models?

Over time, however, I became convinced that there is a significant difference between a system designed to give people good answers and one designed to support human learning.

Ruojun organized her presentation around a fictional learner named Alice. Alice notices food being wasted in her school cafeteria and begins wondering why.

From that simple starting point, she used Alice’s journey to explain three things YEE AI is intended to do:

see the learner, support the learning, and learn from what happens.

She added another principle that became one of the strongest ideas of the meeting:

Do not define the learner too early, and do not close the learning process too early.

Evidence, not labels

Suppose Alice notices food waste once. That tells us very little about her.

But suppose she begins observing the cafeteria, interviewing staff, collecting information, questioning her assumptions, and returning to the issue repeatedly. Over time, a pattern begins to emerge.

Perhaps she keeps returning to sustainability questions in different settings. Perhaps the formal project ends, but she continues anyway. Perhaps she revises her survey, recruits friends, or keeps working even when no grade is attached.

At some point, other students may begin asking her for help. Something she can do becomes valuable to others.

These behaviors may provide evidence of what we call SIPC: strengths, interests, passions, and curiosity.

The key is that SIPC is not intended to become four fixed scores attached to the learner. We should not decide early that “Alice is a sustainability person” and then build her education around that conclusion.

Doing so creates a serious danger. If the AI concludes that Alice likes sustainability, continually gives her sustainability opportunities, and then later observes that she keeps working on sustainability, the system may simply be confirming a pattern it helped create.

Personalization can become confinement.

Ruojun summarized the alternative beautifully:

Evidence, not labels.
Trajectory, not profile.
Development, not prediction.

The goal is to remain open to who Alice may be becoming.

Faster AI, slow learning

The second major challenge is almost the opposite.

AI is extraordinarily fast.

If Alice asks a powerful AI system, “How can we reduce food waste at our school?” it can immediately produce causes, solutions, a survey, an action plan, talking points, and perhaps an entire presentation.

From the standpoint of efficiency, this is wonderful.

From the standpoint of education, however, we need to ask: What is left for Alice to do?

What does she need to notice herself?

How does she know food waste is actually a problem?

Whom should she talk to?

What assumptions might be wrong?

What judgment should she make?

What should she try?

What happens when her first idea fails?

Some difficulties should be removed by technology. But not every difficulty should disappear. Sometimes the difficulty is where the learning occurs.

This led to another phrase that stayed with me after the meeting:

Faster AI, slow learning.

The educational question is not simply how AI can make students learn faster. It is:

What should AI make easier, and what should the learner still experience?

Premature convergence

Ruojun called one of the central risks premature convergence.

Large language models are designed to reduce uncertainty. Give them an ambiguous problem and they quickly organize possibilities into a coherent answer.

Usually that is a strength.

But learning often requires the opposite. Learners need time to remain uncertain. They need opportunities to explore alternatives, ask poor questions, discover better questions, change their minds, make judgments, and experience consequences.

If AI reaches the answer before the learner has truly entered the problem, part of the learning process disappears.

For that reason, YEE AI often responds Socratically.

Instead of immediately answering Alice’s question, it may ask:

What have you actually noticed?

How do you know this is a problem?

Who might see the situation differently?

What information do you still need?

The purpose is not to make the conversation unnecessarily long. It is to keep the possibility space open long enough for the learner to participate in forming the judgment.

Semantic drift

A second problem is what the team calls semantic drift.

Alice begins with food waste. Twenty minutes later, the AI conversation may have moved into climate change, recycling, app development, fundraising, and entrepreneurship.

Everything being discussed may still be intelligent.

But the conversation may no longer have much to do with the problem Alice originally encountered.

So good educational AI needs more than memory. It needs to remember what matters.

Who is Alice?

What is she actually trying to understand?

What are the constraints?

What is happening in the real environment?

What is the educational intention?

Ruojun summarized the desired behavior using three words:

Hold. Scaffold. Return.

Hold the context.

Scaffold the learner without closing the problem too quickly.

Return the learner to judgment, action, relationships, and the real world.

This is also where human interdependence becomes important. The purpose of educational AI should not be to create an ever more powerful private relationship between one student and one machine. It should ultimately return learners to other human beings and meaningful action in the world.

Learning from learning

YEE AI is also being developed around another idea: learning from learning.

Once Alice actually does something, new evidence becomes available.

She talks to cafeteria workers.

She tests an assumption.

She changes her mind.

She abandons an idea.

She persists with another.

Someone else begins depending on her work.

These outcomes give us evidence that cannot be captured simply by asking the learner, “What are you interested in?” or giving her another standardized assessment.

Over time, the process becomes:

outcome → evidence → patterns → future support

But again, the system should not use these patterns to lock the learner into a permanent profile. It should update its understanding while remaining uncertain enough to allow development and surprise.

In that sense, assessment becomes less about determining what a student “is” at one moment and more about following what the learner is becoming.

Doing is learning

Ruojun ended her introduction with perhaps the simplest statement of the entire design:

Doing is learning.

Alice does not learn deeply about food waste because AI explains food waste perfectly.

She learns because she observes something, asks a question, makes choices, talks with people, acts, receives feedback, revises her thinking, and acts again.

The same idea applies to us as educators.

We cannot fully design the future of AI and education in a conference room. We have to try things in real schools, with real teachers, real students, real constraints, and real consequences.

That is why the Courageous Minority community matters to the development of YEE AI. The system is not being presented as a finished product for educators to adopt. The hope is that educators will experiment with it, notice what works and what feels wrong, and contribute to its continued development.

In other words, we are also learning by doing.

What the system currently does

Wanyu then demonstrated the current version of YEE AI.

The platform guides educators through curriculum design using questions rather than generating a finished curriculum instantly. Teachers describe their context, students, goals, duration, challenges, and tasks. The system then helps them think through the learning design step by step.

A “seed pool” allows educators to draw from activities, resources, and ideas contributed by the system and by other members of the community.

Teachers can build challenges consisting of different tasks, add guidance and resources, seek input from colleagues or external experts, and decide how much flexibility students have in accessing the tasks.

Students also interact with the AI through questions. Rather than simply being given answers, they are encouraged to think further, upload their own creations, contribute resources, help others, and document what they have done.

Those activities become footprints of their learning.

Over time, those footprints can contribute to a richer picture of the learner: what they create, what they return to, whom they work with, where they contribute, and how others respond to them.

The system is therefore moving toward something quite different from a conventional grade book or transcript. It is closer to a continuously developing record of a learner’s activity, contribution, interests, relationships, and growth.

Beyond AI tutors

Wanyu also made an important distinction between YEE AI and many current educational AI products.

Much of the AI education market is focused on tutoring: helping students learn mathematics, vocabulary, science, or other conventional subjects more efficiently.

Those tools may work very well at what they are designed to do.

But if AI simply becomes a better way to drill students on the same knowledge, then AI has not transformed education. It has simply made the traditional system more efficient.

That is not the future we are trying to create.

I connected the system back to the three principles that have guided this work:

Personalizable learning: education should help each learner develop unique strengths and interests rather than force everyone toward identical outcomes.

Problem finding and solving: students should learn through identifying and addressing meaningful problems rather than primarily reproducing prescribed knowledge.

Human interdependence: education should help students discover how they can contribute value to others rather than simply compete to rank above them.

These principles require us to rethink not only curriculum but also teaching and assessment.

I also repeated something I have become increasingly convinced of: I do not think schools can prepare students for the age of AI simply by adding an AI literacy course.

Students need more than lessons about AI. They need experience living, learning, creating, judging, and collaborating in an environment where AI is already part of human activity.

Can transformation happen inside existing schools?

The final part of the meeting returned to a difficult question.

Can we really create this kind of learning within existing educational systems?

One participant described the tendency of organizations to respond to disruption by returning immediately to familiar practices. When the future becomes uncomfortable, the status quo feels safe.

He wondered whether truly transformative change is possible while educators remain bound by conventional curriculum, schedules, assessments, policies, and expectations—or whether transformation requires a blank canvas.

Another participant pushed the idea even further, imagining a personal AI agent that belongs to the child rather than to the school: an agent that develops with the learner over many years and supports a much more self-directed learning journey.
These questions do not yet have clear answers.

But they remind us why we formed the Courageous Minority.

We should not expect entire governments, ministries, districts, or school systems to transform at once. Large systems are designed to preserve themselves. Transformation is more likely to begin with relatively small groups of educators and schools willing to try something different.

Some experiments will fail.

Some will work.

Some will reveal possibilities none of us have imagined.

Our task is to create enough protected spaces for these experiments to happen, connect the people doing them, learn from their experience, and make the alternatives visible.

That is the role I hope the Courageous Minority can play.

A courageous minority, not a cynical minority

At one point in the meeting, as several participants described restrictive policies and growing frustration, I reminded the group that we are the Courageous Minority, not the cynical minority.

There are plenty of reasons to become cynical.

Schools are slow to change.

Policies often arrive before evidence.

Curriculum requirements remain powerful.

New technologies are quickly forced into old educational structures.

And the easiest response to uncertainty is often prohibition.

But throughout the meeting, participants also described students undertaking real-world capstone projects, schools experimenting with learner agency, educators building healthier learning communities, and young people contacting government officials about problems they care about.

These examples matter.

Perhaps the future of education will not begin with a national reform or a new set of standards.

Perhaps it will begin with small groups of teachers, students, school leaders, designers, and communities who are willing to ask a different question:

Not How can AI help us do school better?

But:

If AI changes what human beings need to do, what should learning become?

That remains an open question.

And that is precisely why we need courageous minorities willing to explore it together.