A student’s face goes flat in the third row. Something has shifted, you can’t yet say what, and the lesson you planned no longer fits the room you’re standing in. So you set the plan down. You kneel beside the desk, you change the question, you make the joke nobody saw coming, and the class comes back to life. None of that was written in your plan, because none of it could have been written. You read the room and you invented the next moment in real time.
That capacity, the one a good teacher uses a hundred times a day without ever naming it, is the one capacity artificial intelligence does not have. As AI grows faster and more fluent, the anxiety underneath every staff-room conversation about it narrows to a single question: what can’t AI do in education? The answer turns out to be the most ordinary and most human thing a teacher does, and it has a name.
What Teachers Do That AI Can’t
The move you just made, reading the room and inventing the next step, is a form of what Dr. Michael Jacob calls play. Jacob, a neuroscientist, psychiatrist, and Breathe for Change faculty member, defines play in a single word: possibility. “Play for me is possibility,” he says, and he means it precisely. Possibility is the capacity to sense that something different could happen, and to take the embodied risk of trying it before anyone knows whether it will work.
That capacity is not a frill on top of teaching. “Life requires possibility for us to evolve, for us to adapt,” Jacob says, “to take risks on new things, to make mistakes, and to find new ways of being.” Possibility is how a living thing meets a situation its past did not prepare it for. And it is the one thing a machine built entirely from the past cannot reach.
Why AI Can’t Improvise the Way Teachers Can
Strip away the marketing and a large language model is a statistical model. It predicts what is probable given everything that came before: the next word, the next sentence, the next paragraph. The context windows can grow and the weightings can be tuned, and the operation underneath stays the same. It recombines the past. What it cannot do, by design, is leap to something the past does not already contain.
That limit isn’t a hunch. A January 2026 position paper from Google DeepMind makes the case in technical terms, arguing that language models handle statistical induction well and formal deduction increasingly well, but lack abduction, the intuitive jump from lived experience to a genuinely new explanation (Zahavy, Google DeepMind, 2026). The paper is a position piece archived in the University of Pittsburgh’s philosophy-of-science collection rather than a peer-reviewed result, so it stands as an argument rather than a settled finding. But the argument is a strong one, and it lines up with something far older.
In 1980, the philosopher John Searle imagined a person sealed in a room with a vast table of Chinese characters and a rulebook for matching them. He passes responses back through the door that a fluent speaker would accept, and he never understands a single word of Chinese (Searle, Behavioral and Brain Sciences, 1980). The syntax checks out, but the meaning never arrives. Meaning, Searle argued, takes more than symbol-matching, and Jacob names what it takes: a body with something at stake. “AI doesn’t have any skin in the game,” he says. “It doesn’t have anything to lose.” A model can describe fear, grief, or wonder fluently, but it has never felt the floor drop out from under it. That is why emotional intelligence will define the next era of education, and it is why the brain is not a computer in the first place.
Possibility, then, is not a trick of language. It is a property of being alive. Jacob points to the work of his mentor Terrence Deacon, the UC Berkeley anthropologist and author of The Symbolic Species, to make the point at its largest scale. In work he is now developing on the role of play in evolution, Jacob says, Deacon traces the capacity to step off the predictable path all the way down to the cell, where new biology emerges precisely because life does not only repeat what already worked. Possibility is the engine of evolution, not a human luxury layered on top of it.
What This Means in Your Classroom
Teachers have never been in the prediction business. They don’t stand in front of children and recombine the past into the most statistically likely next sentence. They improvise. They notice the child whose nervous system has just gone offline, and kneel down. They take the planned lesson and turn it inside out because something in the room shifted. They make the leap that no record of the past could have produced, and make it while twenty-eight other things are happening at once.
This is also where not-knowing earns its keep. The poet John Keats called it negative capability, the capacity to stay in “uncertainties, mysteries, doubts, without any irritable reaching after fact and reason” (Keats, 1817). Every teacher knows the version that arrives in the shower, or at 4 a.m., or two weeks after they finally stopped forcing the lesson plan. The new idea needs the room that not-knowing makes, and a machine built to resolve the next token as fast as possible has no such room.
So the work that makes educators irreplaceable, as AI gets better and faster at the probable, is the work of the possible: imagining what could happen, creating the conditions for play, and refusing to clamp down on the unexpected when it walks into the room. That is the heart of Human Intelligence in the age of AI, and it is the part of an educator no model can stand in for.
If AI Can Do the Recitation, Teach What It Can’t
Take this seriously and the curriculum question shifts. If recitation and recombination are what machines do best, then the parts of teaching that matter most are the parts that demand a human in the room: the play, the art, the embodied practice, the repair of a frayed relationship, the moment of being genuinely together. John Dewey was making a version of this argument in 1899, in The School and Society, against a schooling built on recitation. A century and a quarter later it is more urgent, not less, because now there is a machine that can recite better than any of us.
For educators sitting with this question right now, the work is to develop the capacities a statistical model structurally cannot. Breathe for Change has been building that work for a decade, across 20,000+ certified educators, 250,000 educators transformed, and 20M+ students impacted. The nearest doorway into it is TeacherCon, the free three-day online event for educators, whose opening day is built on exactly this argument: the Human Intelligence framework and its role in education alongside AI, led by Breathe for Change co-founders Dr. Ilana Nankin and Michael Fenchel.
What the Birthday Poems Were Missing
Dr. Ilana’s birthday began with a poem. A close friend sent it, and then another friend sent one, and then four more, six poems in the span of a few minutes, each one composed and specific and warm. She read the first and was moved to tears. Then she learned the friends had written them with ChatGPT. The poems were still well made, and the qualities they named were still true, and yet something had been quietly stripped out, what Nankin later called “the art and the intention and the love that people pour into creating something for another person.” The information survived. The meaning did not.
That is the whole of it. The poems were accurate, composed, and hollow, because no one had taken the time, and the taking of time was the point. Educators are the ones who take the time. They notice what is specific about one child. They imagine what is possible for the student in front of them, in the room they are actually in, under the constraints they actually carry. The future of education belongs to the people who can still picture what isn’t yet probable, and who teach children to do the same.
Want to go deeper? Listen to the full conversation between Dr. Ilana Nankin and Dr. Michael Jacob on A Work of Heart, the Breathe for Change podcast. They cover the neuroscience of play, the structural gap between human and machine intelligence, and what irreplaceable teaching actually looks like.
