I spent part of my time in high school working with my school district on a five year plan for a cybersecurity curriculum. Pre-GPT, that work meant months of research, interviews with teachers and administrators, and arguing with the Director of CTE for a programme I would never directly benefit from, as I would graduate before ever taking it. But it did teach me how slowly the curriculum development cycle moves, how many stakeholders have to sign off on a single course, and how much of the resistance to change comes from people who aren’t wrong to be cautious, nay, skeptical of a student telling them how to run their courses. Hence when I give opinions about the American education system in the age of AI, now in 2026, I come from having the inside experience, not just from watching it on the news.
I generally believe generative AI can be a net positive for society. I also believe the concerns that many experts raise are indeed real. Reduced cognitive engagement, weaker critical thinking, and the eroding ability to sit with a hard problem before reaching for an answer is somewhat of an existential threat to the core human experience. If you become reliant enough on these tools and you begin outsourcing parts of the core human experience, the struggle that learning is supposed to be, goes away. Both things, here, are true, and the mistake schools keep making is treating them as a dichotomy.
An outright ban is the wrong response. We should want students to be problem solvers who use every tool at their disposal, not just the ones an administrator approved, because that is exactly what will be expected of them in the workplace and in later schooling. That being said, I understand my professors' and former teachers' skepticism. When a student can produce a polished essay without engaging in a single concept, the tool is no longer augmenting the learning, it’s fully circumvented it. Right now, if I want Claude to write an essay for me, nothing in the process probes whether I understand anything. I have to consciously prompt Claude to work with me to make sure that I craft an essay that allows me to fully understand the learning process. And so, the teachers who worry about that are describing the fragile system accurately.
So the answer is not access or no access. The answer is to adapt and revamp curriculums that are outdated, in some cases ill-advised, make use of the tools these AI labs are producing because of what they enable rather than in spite of what they threaten, revisit which concepts we treat as important, and take advantage of the self-directed learning children exhibit naturally before school trains it out of them. The tool should be a support, never a crutch, and the difference between them is a design problem that schools, in terms of administration, have so far refused to take head on.
It starts with teachers. Without that foundation, nothing else stands. A teacher's week is consumed by work that, realistically, has nothing to do with teaching such as: formatting lesson plans, writing rubrics, generating practice problems, drafting the fourth version of a parent / teacher conference email. This is precisely the boilerplate that current models handle really well, and offloading it brings back the scarcest resource in most classroom settings, which is: teacher attention. A teacher with five reclaimed hours a week can spend them on the students who are falling behind via longer office hours, and no education technology over the last thirty years has managed to deliver that. The curriculum work still needs to be researched and rigorous (my district experience makes me allergic to shortcuts here), but the assembly of them should not eat a teacher's evenings.
Second, students need safe access, and safe access cannot mean handing a fifteen year old the same unrestricted chat interface I use everyday. Broad access to tools like ChatGPT and Claude is the very net negative teachers and administrators worry about, because there is currently no reasonable way to ensure the learning process is not being skipped entirely. What I would propose instead is a gradual unlock tied to demonstrated ability. A student who is consistently outperforming their peers, who has shown they can understand the material without the tool, has earned the right to use it the way another peer would at a similar level, and I see no reason to hinder them. A student still building the fundamentals gets a constrained version: one that asks questions back, that refuses to hand over finished work, that behaves like a tutor rather than a ghostwriter. Anthropic's Learning Mode and OpenAI's Study Mode both ship versions of exactly this behaviour today, but it barely scratches the surface of what is needed for students. These are modes that offer prompt toolkits and skills, but are not behaviours that are actually embedded in the classifiers nor the models. The issue, to start, is that schools have no framework for deciding which student gets which mode, and how to allow for students who do demonstrate the need for less guardrails to access those tools.
Third, we need to support parents, because education does not stop at the classroom door. Post 2025 it is clear these models are capable of real agentic work, and that capability should be in parents' hands, as much as it should be given to teachers. Parents who homeschool and need research-based curriculum alternatives they cannot build alone, and parents who simply want to know where their child stands and where they can help is a key demographic of people who could truly benefit from AI. The age of the AI tutor is already here, and economics matter more than ever today. In-person tutoring runs 35 to 40 dollars an hour where I grew up, which prices out most of the families I knew who are considered in “Extreme Poverty”. A capable AI tutor costs a fraction of that, and for a family that could never afford weekly sessions, a tool is now available to them, that in person tutoring simply cannot compete with.
And we need to support innovation itself, because some of the curricula in circulation date back to the 1990s, and I am not exaggerating. Teaching from a framework older than the students, and in some cases the parents, is a disservice to a generation that will graduate into a labour market that now requires vastly different knowledge and skills.
So what does the classroom look like once you accept all of that? It looks like in-person work. AI can and should carry the load outside the classroom, but face to face is where a teacher can gauge what a student has retained, which cannot be delegated to a model. If we give teachers the tools to offload the boilerplate, the classroom hours that remain become higher density with more interaction among all students and more focused time with the groups that need it.
It also looks like teamwork, deliberately fostered. We are raising students inside individually tuned TikTok feeds and personal echo chambers, and the classroom is one of the last places where they are forced to build something with people they did not choose. I say this as someone whose instinct is to work alone, because then at least the output is mine and so is the responsibility. But the group projects where I carried the brunt of the work and took the poorer grade anyway, taught me something solo work never could: how to say, out loud, that a situation needs fixing, instead of quietly resenting the concept of teamwork for years afterward. That is a skill that applies to all of life and should be fostered within the school setting so students are prepared for when they graduate.
Which brings me to assessments, and here I will die on this hill and say that the traditional essay and the traditional exam should go. They do not accurately, nor adequately, measure whether a student can retain and apply knowledge outside of the setting in which the information is relayed; they measure compliance and pattern-matching. I got an A in AP Physics, and then a 3 on the AP exam. The difference between those two grades/scores is the gap between my teacher's method, which was exam drills and whiteboard repetition, and actual understanding of the core concepts that would be on the exam. There were units where I knew that if the class had run labs and projects, and I had been paired with a more driven student, I would have pulled a 4 simply because their drive and motivation to learn would have rubbed off on me. Grinding the same fifty problems before a due date shows how well you have memorised previous exam styles. It does not show learning at all, and a multiple choice sheet cannot tell the difference.
What replaces it is presentation and inquiry. Project-based work where the teacher probes the student's process forces a student to reflect: yes, I used AI as a tool, and here is what I got out of it, here is the decision I made and why. Therefore, you cannot fake your way through a follow-up question about your own reasoning. That structure is also, conveniently, the only assessment format that makes student learning outcomes enforceable rather than aspirational, because the outcome is demonstrated live instead of inferred from a document of unknown provenance.
Cut homework down and move the work into the room, which is this whole thesis restated. I love being called to the whiteboard, working on a problem, having a peer review my work, and then discussing it with the whole class. Discussion is a measurement instrument. You can read for hours and write for hours, but if you never verbalise what you learned, neither you nor your teacher can locate where you are in terms of understanding the concept. Socratic seminars, structured discussions, conversations about the concepts show growth and offer more opportunities to learn.
This applies fully in STEM, where the culture often pretends otherwise. Working a maths problem as a class, toward a solution as a team, exposes the collective learning process, and when someone gets it wrong the response should be "how did you come to that conclusion?" rather than a direct correction, because in my experience eight times out of ten the student catches it themselves: ah, I missed that step. So a student who reviews and verbalises their own work is demonstrating the exact metacognition every curriculum document says they value. And when peers support each other inside the class, the whole class succeeds, and the burden of forming external study groups evaporates because the support structure already exists inside the room. Those relationships carry forward too; that is where "what section are you taking next, we should register together" comes from, and it is how cohorts form without anyone mandating them. Not only does this foster friendships but it further supports the idea of collective success. “When we all succeed, I succeed.”
STEM is notoriously isolating, and it is frustrating to watch some engineering and physics professors try so hard to foster collaborative work while their colleagues down the hall run silent lecture halls with no room for discussion and no way to sit down and focus on why students aren’t understanding that particular concept. Teachers, professors, and education faculty need to be on the same page, because when one professor participates in that culture and the next does not, the student is left renegotiating the rules of the classroom every semester, and that dissonance is its own tax on learning.
Students perform better when the whole class is invested in the success of its peers. Every proposal and opinion above: teacher tooling, gated student access, parent support, project-based assessment, discussion as measurement, points at that one fact. AI does not change it, but AI is the first technology in a long time that could let us build a school around it.