The Teacher's Last Monopoly

Over half of U.S. college students now use AI for their assignments. Eighty-six percent of the time, their instructors cannot tell.¹

The reaction has been predictable. Institutions crafted guidelines — 94% of top universities now have them² — while the technology they were regulating had already moved on, becoming undetectable and writing in the students’ voice. The conversation focused on whether students were cheating. Educators also need to know what students are learning through these tools, and which parts of teaching still need their care and judgment. The tutoring studies offer a place to start.

How AI changes the role of the educator

Educators have historically traded on three forms of value: possessing knowledge others lack, transmitting it effectively, and providing mentorship and development. The first two are now measurably compromised.

At Harvard, a randomized controlled trial pitted an AI tutor against the university’s gold-standard: active learning classrooms. The AI-tutored students achieved more than double the learning gains, with effect sizes ranging from 0.73 to 1.3.³ Not a marginal improvement, but a categorical difference in the transmission function itself.

Meanwhile, Duolingo’s AI engine processes over 500 million lessons daily, personalizing across tastes, attention spans, and ability levels. The platform created roughly 150 new language courses in a single year. The first 100 had taken twelve years.⁴ And as video and audio generation tools approach fluency, the delivery medium itself is being commoditized alongside the content. Knowledge possession at scale is now an engineering problem, not a staffing one.

This is not the first time technology promises to “educate”. However, every previous technology predicted to replace teachers — radio, television, the internet, MOOCs operated on a one-to-many, broadcast style relationship. AI tutoring is the first technology matching the one-to-one granularity of human tutoring while being infinitely scalable. The mechanism is categorically different this time. And the institutions are already conceding it: faculty are “no longer simply transmitters of knowledge” but “facilitators.”⁵

From Train the Trainer to Train the Machine

Gregory Kestin, a physics professor at Harvard, spent months building an AI tutor called PS2 Pal. He gave it brief responses, one step at a time, and rules that kept it from revealing full solutions. Socratic questioning was enforced. Those choices encoded decades of pedagogical judgment about how to help a student think, allowing his teaching philosophy to reach hundreds of students simultaneously.⁶

The result was the RCT that doubled learning gains. The design of the tutor helps explain the result.

Kestin did not outsource his expertise to the AI. He crystallized it. The content was already commoditized (any large language model knows introductory physics). What the AI could not do on its own was teach well. It needed a teacher to teach it how to teach.

This is meta-pedagogy: the skill of designing how AI should teach, not what it should teach. And the distinction between meta-pedagogy and simply “using AI tools” is sharp. Making AI-generated worksheets changes how material is produced. Encoding your pedagogical rules changes what happens when a student needs help: what the AI should leave for them to work out, which struggles are productive, and when to push or wait.

Expert estimates suggest AI can automate roughly 40% of lesson planning, while context-based pedagogy remains overwhelmingly dependent on human judgment.⁷ These figures come from practitioner analysis rather than controlled measurement, but the directional split is consistent with what the Harvard and CMU data show: the line is not human versus machine. It is transmission versus design.

A CMU study tracking 350 seventh-graders over a full school year found that students with human-AI tutoring finished 0.36 grade levels ahead of AI-only students, with gains compounded over time. Eighty-eight percent of those students were economically disadvantaged.⁸ The human tutors were not lecturing. They were providing motivational scaffolding, contextual judgment, accountability. Applied meta-pedagogy, whether they called it that or not.

The shift from delivery to design becomes more valuable and echoes a broader pattern across knowledge work, where the question is not whether expertise matters but how it survives commoditization.

The Pedagogy Paradox

AI models’ knowledge about education reflects the consensus in their training data. It optimizes within known frameworks with extraordinary precision, but it cannot ask “what if we taught this completely differently?” The boldest educational innovations — flipped classrooms, problem-based learning, ungrading — came from educators who broke with consensus. These ideas would be trained out of any AI system because they contradict the training data.

This is the Pedagogy Paradox: the technology that commoditizes teaching execution cannot generate the next teaching revolution.

There is a plausible concern of course. Research on AI-assisted creative work suggests that students who never had AI assistance outperformed on creativity tasks, while those who lost AI access after relying on it performed dramatically worse — pointing to a possible erosion of the creative self-concept needed for pedagogical invention.⁹ The argument for personalized education does not depend on this single finding, but if AI dampens the capacity for divergent thinking even modestly, the tool that delivers knowledge so efficiently needs to be applied in conjunction with other techniques.

Educators who protect time for uncomfortable experimentation — pedagogical ideas that would make their colleagues wince — are cultivating exactly the kind of value AI cannot replicate. AI needs human-generated food for thought to build personalized learning paths around. An innovation that survives committee review unchanged probably does not challenge the teaching assumptions enough to matter. That is why these experiments may unsettle colleagues’ expectations.

What Duolingo Got Wrong (and Harvard Got Right)

Duolingo’s AI-first strategy cut roughly 10% of its translators, writers, and curriculum experts alike.¹⁰ The productivity gains were real: 150 courses in one year. But users noticed. Content became repetitive, robotic, stripped of the playful tone that made the platform iconic.

The company made a category error. Translation — pure knowledge possession — was genuinely commoditized, and AI handled it well. Curriculum design, engagement sequencing, and the motivational voice that kept learners coming back required human creative judgment. Duolingo conflated two distinct layers of human contribution and cut both.

Harvard went the other direction. Kestin preserved the meta-pedagogical layer entirely. He designed how the AI taught. The content was commodity. The pedagogical architecture was the irreplaceable contribution. Quality multiplied instead of degrading.

The pattern extends beyond knowledge work. In vocational welding programs, VR simulators with AI-driven assessment now analyze bead shape, fusion depth, and defects in real time and more consistently than any instructor could. But welding instructors did not disappear. They migrated to orchestrating when students move from virtual to physical practice, interpreting AI diagnostics for individual learning plans, coaching the embodied intuition no simulation captures. Even in the trades, transmission commoditized while design elevated.

The lesson is structural. Strip the design layer and quality degrades. Preserve it and impact multiplies.

Engagement and the “old wood ruler”

AI tutoring also depends on whether students use it consistently.

Khan Academy’s own efficacy data shows that only 9% of 350,000 students reached the recommended usage threshold of 30 minutes per week.¹¹ The technology works but only when students use it. The capability to outperform human transmission exists in theory, but the engagement problem prevents it from functioning at scale without human facilitation. Making students care about algebra on a Tuesday afternoon remains stubbornly, irreducibly human work. The old wood ruler was a crude attempt to keep students at the task. We need better ways to sustain that engagement.

This is the engagement trap: AI delivers information with superhuman precision but cannot own the learning relationship. The findings tell us where the human monopoly actually lives — not in knowing or explaining, but in making someone want to learn.

The equity dimension sharpens the argument further. The U.S. Department of Education warns that when AI automates instructional decisions at scale, achievement gaps could widen through biased assumptions, incomplete data, and poor pedagogical theories.¹² The CMU study’s population — 88% economically disadvantaged, 65% Hispanic — demonstrated that human facilitation matters most precisely where the stakes are highest.⁸

The Harvard AI tutor succeeded because researchers deliberately designed equitable constraints into the system.⁶ Educators who identify bias in AI tutoring and redesign for equity are performing exactly the meta-pedagogical work the thesis predicts will be most valuable. The engagement crisis and the equity risk are not arguments against AI in education. They are arguments for the humans who design how AI educates.

So where does this leave the profession?

Start by auditing how you spend your time. Track one work week across three categories: Knowledge Possession (preparing content you could look up), Knowledge Transmission (explaining things an AI tutor could explain), and Pedagogical Design (judgment calls about sequencing, motivation, and adaptation that require understanding your specific students). Most educators discover 60-80% of their time sits in the first two categories. That ratio shows how much of your working time is exposed to the change described here.

Then run a replacement test. Give an AI your three most time-consuming teaching activities. If it produces a 70% adequate response, that activity is commoditized. The bar is not whether the AI matches your best teaching moment. It is whether it is good enough at scale — your average Tuesday at 2 PM across 150 students simultaneously.

Build from there with one unit of your course. A custom AI tutor needs the choices you make when helping your students: what to leave for them to discover, which misconceptions to watch for, and which struggles are worth protecting. Encoding those rules lets more students benefit from your teaching judgment.

And protect 10% of your time for pedagogical experiments that feel wrong by current standards. The ideas that would be trained out of any AI are your durable advantage.

Two tiers of education AI professionals are emerging. Foundational meta-pedagogues — a smaller group — design AI tutoring architectures and audit them for bias and equity. Applied meta-pedagogues — a larger group, though likely smaller than today’s teaching workforce — select, configure, and orchestrate AI tools for specific student populations. The CMU data proved these orchestrators produce gains that AI alone cannot.⁸ Neither tier is traditional teaching. Both are genuinely valuable.

The teacher’s last monopoly is not knowing or teaching. It is designing how AI teaches — and imagining how to teach what has never been taught. That shifts the role toward instructional design and responsibility for how the tools are used.

References