AI adoption in schools hits 77% of professors; honest reckoning replaces knee-jerk bans

The article opens with the author's anecdote about upsetting university educators by criticizing word-count requirements in assignments, arguing that no workplace manager ever specifies a word count for professional writing and that rewarding length instead of clarity sets students up poorly. The author frames this as emblematic of education's tendency to defend old practices simply because they are old, while simultaneously warning against chasing new technologies simply because they are new. The real task, the author argues, is figuring out what is actually true.

On the educator side, adoption is widespread and accelerating. A 2026 global survey of more than 45,000 students and faculty found that 77% of professors now use AI in teaching, up from 61% the year before. Among K-12 educators, roughly six in ten use AI in some capacity. The work they report is unsexy but consequential: lesson planning, generating rubric feedback, creating practice materials, and drafting parent emails—administrative tasks that educators say consume time that could be better spent on teaching. Educators using AI weekly report saving close to six hours per week, or roughly six weeks worth of time over a school year.

Student adoption is even more universal. Roughly 86% of children ages 9 to 17 use AI, and 85% of those use it for schoolwork. Some usage aligns with educational aims—brainstorming essay structure, requesting alternative explanations of concepts, or checking work before submission. Other usage does not—students simply ask the tool to produce the answer. Critically, only about half of students say they have been taught how to use AI safely, despite the overwhelming majority already using it daily. The author frames this as handing kids a tool and skipping the instructions.

New York City's public schools offer a concrete example of institutional learning. In January 2023, the district banned ChatGPT outright, citing safety and accuracy concerns. Within four months, in May, the district reversed the ban. Chancellor David Banks, whom the author notes he worked with on an earlier initiative (rolling out Adobe Express), explained that the original ban was "knee-jerk fear" and overlooked the fact that students will need to understand AI technology in the world they will enter. The author highlights this as a rare case of a large institution admitting error publicly and in a timely way.

Khan Academy's Khanmigo provides the article's main case study. The tool grew from 40,000 K-12 students to 700,000 in a single school year and is on track to cross a million students this year. At one pilot school, Enid High School in Oklahoma, zero students failed geometry after a semester of using Khanmigo. Teachers in participating districts report saving hours per week. However, Sal Khan has said plainly that the first version "did not change student learning as much as many hoped it would." More problematically, the underlying model "confidently lies to students" when encountering math problems it cannot solve—a documented hallucination problem that remains unsolved. The tool has no mechanism to signal to a nine-year-old that it is guessing rather than knowing. When students seek answers, they often receive Socratic questions instead, which is intentional but still frustrating for young learners. The author cites math teacher Dan Meyer's observation that Khanmigo doesn't love students—it cannot. Khan Academy is not hiding from these shortcomings; the first independent randomized trial of whether Khanmigo improves outcomes is due later this year.

The author uses this transparency to illustrate a larger historical pattern. Education has gotten major initiatives wrong repeatedly: New Math in the 1960s confused parents and took years to unwind; the phonics-versus-whole-language debate over reading instruction spanned decades before research settled the question (and the question still occasionally resurfaces); No Child Left Behind turned American classrooms into test-prep factories for a decade before the cost was seriously reckoned. One-to-one laptop programs are the freshest example. Maine, one of the first states to put a laptop in every student's hand in 2002, saw no improvement in test scores fifteen years later. A middle school in Kansas recently removed its Chromebooks after finding that students used them mainly to watch videos and harass classmates, not learn. Districts in North Carolina, Michigan, Virginia, and Maryland are walking back their one-to-one policies for similar reasons. A neuroscientist who testified before the Senate cited international test data showing that frequent in-class computer use correlates with meaningfully lower math and science scores, concluding that Gen Z may be the first generation in modern history to score lower than their parents.

The author's central argument is that AI creates a structural opportunity for faster correction. Khan Academy shipped Khanmigo, identified underperformance, said so publicly, and began iterating—completing a full cycle in about three years. In contrast, past education mistakes took "the better part of a generation" to recognize and fix. The shift is not in technology but in the timeline for learning from error. The author frames this not as a reason to move recklessly but as a reason not to be afraid of moving at all, though acknowledging that risks exist and will be discussed in the next installment of the series.

The article frames AI in education not as a revolutionary tool but as a test of whether schools can learn from past mistakes faster than they have in the past. The author recounts how previous innovations—New Math, laptop programs, and No Child Left Behind—were either oversold, misimplemented, or took a decade or more to correct. One-to-one laptop initiatives, for example, showed no improvement in Maine test scores fifteen years after launch, and districts in North Carolina, Michigan, Virginia, and Maryland are now walking back those programs after discovering students used the devices primarily for entertainment and harassment rather than learning.

Khan Academy's Khanmigo serves as a counterexample: the organization acknowledged that its initial version underperformed expectations and that the underlying model hallucinates on math problems it cannot solve—problems that particularly affect younger students who lack the experience to detect false confidence. Rather than hiding these shortcomings, Khan Academy is iterating openly and plans an independent randomized trial later in the year. This represents a structural shift in how education technology can be validated: where past missteps took a generation to recognize and reverse, AI tools can now be tested, found wanting, and improved within a three-year cycle (roughly the span from third to fifth grade).

The article also highlights a cultural blind spot: educators resisted the author's suggestion that word-count requirements reward verbosity over clarity—a reflex rooted in "how we've always done it." That same conservatism has historically delayed adoption of necessary corrections. Now, with 77% of professors already using AI and 86% of school-age children using it daily, the critical gap is not whether AI belongs in classrooms but whether students receive adequate instruction in using it safely, which currently only about half report receiving.

No comments yet. Be the first to share your thoughts!