This is Part 1 of a 3 part series.

The Old Mindset

A while back I managed to upset a room full of university educators with one throwaway off-the-cuff comment. Mid-presentation I committed a heinous sin; I ridiculed the assignment word count.

My reasoning: in thirty-plus years of working, no manager has ever told me how many words my report or presentation needed to be. And more importantly, rewarding length rewards verbosity, and if the point of writing is to communicate, the student who can say the same thing in a third of the words should be rewarded for that, not penalized for falling short of an arbitrary number. Obvious, right? Well, the icy reception that my explanation received made it clear that while nobody in that room wanted to argue with my logic; this was a position they were neither used to hearing nor willing to contemplate.

That's not an AI story. It's a "how we've always done it" story, and I bring it up because much of the resistance I hear to AI in the classroom isn't actually about AI at all. Rather, it's the same reflex, just aimed at a new target.

I want to be really clear about where I stand: Schools still teaching, testing, and grading the way they did when I was in school are doing their students a disservice, and AI is forcing a reckoning that was overdue anyway. But I also want to be direct about the other failure mode, because it's just as real. Education has a habit of chasing the shiny new thing, declaring it the future, expending precious funds and resources, and then spending the next decade quietly walking it back. The ed-tech pendulum has swung too far too often. Pens and paper still have a place. So does slow thinking, and memorization, and doing something the hard way because the hard way teaches you something the easy way doesn't.

Interestingly, that correction is happening in a completely different corner of the curriculum right now. Schools spent decades stripping out home economics and shop class chasing test scores and college-prep metrics, and now there's real appetite to bring them back: how to change a tire, understand credit cards and debt, register to vote, build something with your hands. Nobody's calling that reckless. It's just the pendulum swinging back toward balance. AI in the classroom deserves the same instinct: don't defend the old way because it's old, and don't chase the new way because it's new. Figure out what's actually true.

Ok, So What's Actually Happening?

With that framing out of the way, here's what's actually happening in classrooms right now. Not a hot take, just the honest state of things.

Educators: On the educator side, adoption is real and growing fast. Something like six in ten K-12 educators now use AI in some capacity, and the ones who use it weekly are saving close to six hours a week, roughly six weeks over a school year. Higher ed is moving at least as fast: a 2026 global survey of more than 45,000 students and faculty found 77% of professors now use AI in their teaching, up from 61% just the year before. Mostly it's the unglamorous stuff: lesson planning, drafting rubric feedback, generating practice materials, writing the parent email that would otherwise eat twenty minutes. And the good news is that nearly all of them say it's saving real time on the administrative load that was never really the job in the first place.

Students: On the student side, it's not a future trend, it's already the water they swim in. Roughly 86% of kids nine to seventeen use AI, and 85% of those kids are using it for schoolwork. Some of that is exactly what you'd hope: brainstorming an essay structure, getting a concept explained a different way after the classroom explanation didn't land, checking work before turning it in. Some of it is exactly what you'd worry about: having it just spit out the thing. Both are true at the same time, in the same classrooms, sometimes with the same kid on different days.

And here's the part that should worry people more than the cheating headlines do: only about half of kids say they've actually been taught how to use AI safely, even though the overwhelming majority are already using it daily. We've handed kids a tool and skipped the instructions. I'll come back to this point.

The Old Mindset, Correcting Itself

If you want a clean example of the old-mindset reflex actually correcting itself, look at New York City's public schools. In January 2023, the district banned ChatGPT outright, citing safety and accuracy concerns. By May, they'd reversed it. Chancellor David Banks (someone I worked with when we rolled out Adobe Express throughout the district) put it about as plainly as an official ever does: the ban had been "knee-jerk fear," and it overlooked the fact that these kids are going to work in a world where understanding this stuff is essential. That's a big system admitting it got the first move wrong, in public, within four months. That's not nothing.

Case Study: Khanmigo

The most useful case study for "how is this actually going," though, is probably Khan Academy's Khanmigo, mostly because Khan Academy has deservedly earned the trust of educators, and has been unusually honest about how things have been going.

The growth is real: Khanmigo went from 40,000 K-12 students to 700,000 in a single school year, on track to cross a million this year. One pilot school, Enid High School in Oklahoma, reported zero students failing geometry after a semester of using it. Teachers in participating districts are reporting hours saved per week, same story as the broader stats above.

But here's what I actually respect about it. Sal Khan has said plainly that the first version of Khanmigo "did not change student learning as much as many hoped it would," and that the underlying model "confidently lies to students" when it hits a math problem it can't actually solve. Hallucination in math is a real, documented, unsolved problem, and Khanmigo doesn't currently have a way to flag to a nine-year-old when it's guessing instead of knowing. Math teacher Dan Meyer's line about it, that Khanmigo doesn't love students, is a fair shot. It doesn't. It can't. And the most common complaint from younger kids is that when they want an answer, they get a Socratic question back instead, which is by design, and still frustrating if you're eight.

None of that is spin. That's a company saying: this didn't work the way we hoped, here's specifically where, here's what we're doing about it. The first real independent randomized trial of whether Khanmigo improves outcomes isn't due until later this year. Nobody's hiding from that either.

Getting It Wrong, Faster

I don't see any of that as a mark against Khanmigo. On the contrary, I see it as the whole argument for how this is supposed to go.

Education has gotten a lot of things wrong before, and the track record on correcting those mistakes is not good. New Math showed up in the '60s, confused a generation of parents, and took years to unwind. The phonics-versus-whole-language fight over how to teach reading ran for decades before the research finally settled it, and settled it again, and is still occasionally re-litigated. No Child Left Behind turned a decade of American classrooms into test-prep factories before anyone seriously reckoned with the cost.

One-to-one laptop programs are the freshest entry on this list, and the correction is happening right now instead of decades later. Maine was one of the first states to put a laptop in every student's hands, back in 2002, and fifteen years later its test scores hadn't moved. This year, a middle school in Kansas quietly took its Chromebooks back after discovering kids were mostly using them to watch videos and torment classmates over school Gmail, not to learn anything. Sweden is spending a purported $100M to replace tablets with textbooks as part of a plan to wind back an all-digital policy that often prioritized technology over pedagogy. Districts in North Carolina, Michigan, Virginia, and Maryland are walking back their one-to-one policies for the same reason, and a neuroscientist who testified before the Senate pointed to international test data showing that frequent in-class computer use correlates with meaningfully lower math and science scores. The most painful line in his testimony:

Gen Z may be the first generation in modern history to score lower than their parents did.

Each of those took the better part of a generation to recognize and fix. Khan Academy shipped a tutor, watched part of it underperform, said so out loud, and started iterating, and that whole cycle took about three years. That's the actual shift AI makes possible in education, and it has nothing to do with test scores. We used to get things wrong for decades before we noticed. Now we can get something wrong, say so, and fix it in the time it takes a kid to go from third grade to fifth.

That's not a reason to move recklessly. It's a reason not to be afraid of moving at all.

That's not to say there are no risks, as I'll discuss in the next post.