The Teacher in the Age of AI

The fear that AI will replace teachers is, in the short run, wrong. The complacency that AI will therefore change nothing about teaching is also wrong. These two errors are mirror images of each other, and the education policy conversation is currently paralyzed between them.

What AI actually does to teaching is more specific than replacement and more disruptive than nothing. It eliminates some of what teachers currently spend time on, raises the floor on what a mediocre teacher can offer students, and raises the ceiling on what an excellent teacher can accomplish. The net effect on the profession depends almost entirely on how schools respond to those three things: which functions they stop asking teachers to perform, which ones they invest in, and whether the workers who survive the transition are doing the high-value or the low-value work.

Let’s start with what AI can actually replace in teaching, because it’s more than most teachers want to admit and less than most technologists claim.

Content delivery is the obvious one. A lecture explaining how photosynthesis works, delivered by a mediocre teacher to 30 students who learn at different speeds and have different prior knowledge, is an inefficient use of everyone’s time. An AI-augmented explanation, self-paced, with embedded comprehension checks that identify when a student is lost, can do the content delivery part of that lecture better than most teachers can. This is not a slight against teachers. It’s a recognition that lectures are intrinsically bad at adapting to individual learners, and adaptation is something AI is specifically good at. The teachers who resist this point are mostly resistant because “content delivery” is a large fraction of what they do all day, and acknowledging that AI can do it better is uncomfortable.

Routine feedback on written work is another category. Grading 150 student essays on whether the thesis is clear, whether the evidence supports the claims, and whether the transitions are coherent is genuinely exhausting work that takes teachers many hours and often results in comments students don’t read carefully. AI can do this work at scale, with reasonable quality, freeing teachers for the feedback that requires human judgment: the response to the student who is clearly struggling with something beyond the assignment, the recognition of a genuinely original idea that deserves encouragement, the identification of a student whose work quality has dropped sharply and who might need support outside the academic.

Answering routine student questions is a third category: “what does the assignment ask for,” “can you explain this concept again,” “is this example from the textbook on the exam.” It currently takes substantial teacher time and is often handled outside official hours, creating a hidden labor burden that disadvantages teachers at schools with higher student-to-teacher ratios.

None of these categories are small. Combined, they represent a significant fraction of a teacher’s working day. If AI can handle them adequately, the question becomes: what do you do with the freed-up time? And this is where school systems’ responses diverge sharply.

The optimistic case is that AI handles the routine work, and teachers spend more time on what only humans can do: noticing the student who is disengaged, building the kind of relationship that makes a student believe they’re capable of more than they thought, designing the challenging discussion that develops critical thinking, providing the experience of intellectual excitement that makes students want to keep learning. These are not things AI can do, and they are, arguably, the most important things teachers do.

The pessimistic case, and this is the direction several large school systems were moving in 2025, when fiscal pressure and AI hype combined, is that AI handles the routine work and that freed-up time is used to justify a reduction in teaching staff. The math is appealing to administrators: if one teacher with AI tools can manage what previously required 1.3 teachers, and if educational outcomes are roughly maintained on the measured dimensions, then you can cut 23% of your teaching force.

This is the wrong answer. It’s wrong because the unmeasured dimensions, student engagement, sense of belonging, relationship with a caring adult, motivation, will degrade, and the degradation will show up in student outcomes years later when the causal chain has become too diffuse to hold anyone accountable for.

The historical parallel that isn’t made often enough is the wave of administrative technology in healthcare in the 1980s and 1990s. Electronic health records, billing automation, and diagnostic decision support were all supposed to free physicians from paperwork, allowing more time with patients. What actually happened, in most health systems, was that the administrative savings were captured as organizational efficiency gains, fewer support staff, higher patient loads, and physician time with patients did not increase. The technology reduced administrative burden and simultaneously reduced the resources devoted to the human-intensive parts of care. Patient satisfaction data from this period tells the story clearly.

Education risks the same pattern. Technology that could free teachers from low-value tasks will, if administrators respond primarily to the efficiency signal, be used to justify reducing investment in the high-value tasks instead.

The teachers who will thrive in an AI-augmented classroom are not the ones who know the most about their subject. They’re the ones who are best at the specifically human parts of education: designing experiences that provoke genuine thinking, building environments where students feel safe to be wrong, recognizing what each student needs that isn’t in the curriculum. These are skills that are very difficult to train and very difficult to evaluate on the standard teacher effectiveness metrics. They’re also the skills that distinguish the teachers students remember 20 years later from the ones they’ve forgotten.

The fact that these skills are hard to evaluate has always been a problem for teacher training and compensation. AI makes the problem more visible by removing the easier-to-evaluate tasks, content delivery, routine feedback, from the picture. What’s left is the irreducibly human part, and school systems are going to need to decide whether to invest in finding, developing, and retaining people who are good at it.

There’s a subset of teachers for whom AI presents not a threat but a genuine superpower. A teacher with strong subject knowledge and genuine pedagogical creativity, previously limited by the time consumed by administrative work and routine instruction, can now operate at a level of sophistication that wasn’t practical before. They can spend class time on discussion, design, demonstration, and individual attention. They can assign more ambitious projects because they’re not paralyzed by the grading burden. They can be more responsive to individual student needs because the AI has already surfaced which students are struggling with what.

The teacher who should be worried is not the brilliant one. It’s the one whose primary value-add is delivering content reliably to a captive audience, who is, in effect, a competent but uninspiring lecturer. AI replaces that teacher’s marginal contribution without necessarily replacing their institutional role, which creates a different kind of problem: a teaching workforce where the distribution of human-specific value-add has shifted dramatically upward, but the institutional structures (tenure, standardized contracts, undifferentiated pay) haven’t shifted to reflect it.

The school systems that handle this transition well will invest heavily in training teachers on what AI is genuinely bad at, redesign classroom time around high-value human interaction rather than content delivery, and use AI tools to let teachers spend more time with the students who most need human attention. The school systems that handle it badly will use AI as a cost-cutting tool while maintaining the fiction that educational quality is being preserved because the test scores are still in the acceptable range. Both kinds of school systems exist right now. The difference between them will be visible in longitudinal outcome data in about a decade, which is too slow for the administrators making the decisions today to face any accountability for the choices they’re making.

The teacher training question is more important than the technology question, and it’s receiving a fraction of the investment. Teacher preparation programs in the United States were already struggling before AI. Teacher shortages in mathematics, science, and special education have been documented for years, and the pipeline of people choosing teaching as a career has been declining in most states since 2010. The average starting salary for a public school teacher in the United States in 2025 was $42,000. An entry-level software engineer makes roughly three times that. The pipeline problem is fundamentally a compensation problem, and no amount of AI tool integration changes the underlying economics.

What AI could do for teacher training, if the resources were directed this way, is compress the feedback loop between novice teacher practice and experienced guidance. A new teacher who uses an AI system that observes their classroom practice, identifies patterns in how students engage and disengage, and provides specific actionable feedback could develop pedagogical skill faster than the standard induction program, which typically involves monthly observations from a mentor who is themselves stretched across 10-15 new teachers. The technology exists. The investment in deploying it in teacher training programs is minimal compared to the investment in student-facing AI products.

This asymmetry, heavy investment in student-facing AI, minimal investment in teacher development AI, reflects the market structure of EdTech. Students (or their parents) are paying customers. Teacher professional development is a government expenditure line that gets cut when budgets tighten. The companies building AI education products are following the money, not the pedagogy. And the money is in the direction that helps students who already have advantages, deployed in schools that have the resources to buy subscriptions, in a way that compounds the existing inequalities in educational quality.

Getting AI in education right requires recognizing this dynamic explicitly and investing in the parts of the system that markets will underserve: teacher development, foundational instruction, and the human-intensive interventions that make the difference for the students who are furthest behind. The technology is not the bottleneck. The policy and investment priorities are.

The single most counterproductive thing a school system can do right now is use AI as a justification for cutting teaching staff without first documenting which teacher functions AI actually replaces, which it supplements, and which it leaves entirely untouched. The cutting comes fast; the rebuilding, when the inevitable gaps become visible in student outcomes years later, comes slow. The teachers who get cut are the first to go, but the students who pay the price are the ones left behind by an institution that mistook cost reduction for modernization.