In a ninth-grade biology class I observed years ago, the teacher stopped eleven minutes into a lab on osmosis, crossed the room, and crouched next to a boy by the window. He had not raised his hand and had not spoken all period. His mother had been taken to the hospital the night before and nobody had told the school. She had read it off his face. She went back, picked the sentence up where she had dropped it, and walked him to the office at the bell.
I think about that whenever someone asks me whether artificial intelligence will replace teachers. Assume that in ten to twenty years a system exists that explains any concept in the K-12 curriculum more clearly and more patiently than the average adult in the room. I think that is closer to likely than not. It still does not settle what people are actually asking.
What a teacher actually does
Most arguments about AI and teaching treat the job of teacher as the same thing as the task of instruction, but they're different. Because human teachers offer a lot more than what AI can do currently.
Take an ordinary elementary day. There is instruction. There is also supervision and physical custody of minors for six or seven hours. There is classroom management, which is mostly the work of holding attention in a room where attention is scarce. There is motivation, which is largely relational. There is noticing, which is what the biology teacher did at minute eleven. There is mandated reporting, a duty that attaches to a licensed adult and not to software. There is implementing accommodations under a 504 plan or an IEP, a compliance obligation with a paper trail and a due-process appeal. Underneath all of it sits legal accountability, because when something goes wrong in a school a named adult answers for it. Instruction is just one slice of that.
| Teacher function | Outlook by 2040 | Why |
|---|---|---|
| Lesson planning and materials | Likely automated | Text generation against a standard. Already happening. |
| First-pass grading of routine work | Likely automated | Pattern matching to a rubric, with teacher review on exceptions. |
| Differentiation and practice sets | Likely automated | Per-student adaptivity is what the technology does best. |
| Explaining content, first-pass tutoring | Likely automated | The clearest case for AI. Assuming it wins here. |
| Progress monitoring, IEP and 504 paperwork | Partly automated | Drafting yes. Sign-off and due process stay with a licensed adult. |
| Diagnosing why a student is stuck | Partly automated | Academic causes yes. Home, health and social causes rarely reach the data. |
| Motivation and relationship | Resistant | AI doesn't have feelings and can't relate to students. |
| Noticing the child who is off today | Resistant | Needs physical presence and a baseline built over months. |
| Supervision and physical safety | Resistant | Custody of minors is not an information problem. |
| Mandated reporting and child protection | Resistant | A statutory duty held by a person. |
| Legal accountability | Resistant | No jurisdiction has assigned it to a system. |
That table is my forecast for various teacher functions and if AI can replace it by 2040.
Earlier predictions
In an interview published in the New York Dramatic Mirror on July 9, 1913, Thomas Edison said of motion pictures: "Books will soon be obsolete in the public schools. Scholars will be instructed through the eye." The line is often misdated to the 1920s and often reproduced with "public" dropped. Instructional radio followed in the 1920s and 1930s with similar promises. B.F. Skinner published "Teaching Machines" in Science in 1958, then spent a decade arguing that programmed instruction could deliver individualized reinforcement at scale. Educational television came next. In 2012, the New York Times declared the year of the MOOC (Online Classes).
In the first year of open online courses from Harvard and MIT, 841,687 registrations produced 43,196 certificates, roughly five percent. Six years later Justin Reich and Jose Ruiperez-Valiente reviewed the same platform in Science and reported that completion rates had not improved, that most learners never returned after their first year, and that growth had concentrated in the world's most affluent countries. The courses themselves worked. Enrollment patterns and institutions barely moved.
"It never happened before" tells you very little about a technology that is different in kind. Film could not respond to you. Skinner's machines could not diagnose you. A 2040 tutoring system plausibly does both. Anyone reaching for the Edison quote as proof that teaching is safe is arguing from intuition about a machine they have never examined. What the history establishes is narrower than that. In every prior case the instructional bottleneck turned out not to be the thing holding the institution together.
The tutoring research
The research everyone cites is Benjamin Bloom's 1984 paper in Educational Researcher, "The 2 Sigma Problem." Bloom reported that students taught one-to-one with mastery techniques performed roughly two standard deviations better than conventionally taught students, and posed the challenge of finding group methods that could match it. That figure is the ceiling in nearly every AI-tutoring pitch written since The two-sigma result has been hard to replicate, it came from a small set of dissertation studies run under tightly controlled conditions, and most researchers treat it as an upper bound rather than a settled effect size, which makes it a reasonable target and a poor promise.
The better-grounded finding is about human tutoring. Andre Nickow, Philip Oreopoulos and Vincent Quan pooled 96 randomized trials of PreK-12 tutoring in the American Educational Research Journal in 2024 and found an overall effect of roughly 0.29 standard deviations. Results were strongest when tutors were teachers or paraprofessionals, when sessions ran at least three days a week, and when tutoring happened inside the school day.
That supports the premise of the AI case, which is that intensive personalization works. It also describes a version that was relational, delivered by a person, inside the school day, at high frequency.
Schools as childcare
A public school is an instructional institution and a custodial one. Tens of millions of American parents hold jobs because a building takes their children in the morning and keeps them, fed and supervised, until mid-afternoon. In household economics that may be the primary function. A perfect AI tutor does not solve custody. It does not serve lunch, stop a fight in a hallway, or sit with a child having a panic attack, and it cannot be left legally responsible for a seven-year-old for six hours.
So anyone who says AI can replace teachers is also saying kids don't need an adult watching them all day. Most of them have not thought about that part.
The same holds for liability. If an AI system gives a child instructions that lead to harm, misses a disclosure a mandated reporter would have caught, or allocates attention in a discriminatory way, who is responsible? The vendor, the district, or the uncertified adult in the room? No state legislature has answered this, and the bills moving now are largely disclosure and procurement rules. Maryland's districts were put on a deadline to adopt AI policies before anyone resolved what those policies make anyone liable for. The binding constraints here are legal, political, and economic rather than technical.
Where it is already happening
Some displacement is already underway. Districts that cannot staff a physics or Latin position fill it with synchronous virtual instruction. Credit recovery ran on software long before anyone called it AI. Rural districts with one applicant per posting take what they can get. In each case an adult monitor stays in the room, which is what my argument predicts: the instruction goes first to AI but the adult supervision stays.
A current example is Alpha School, the private chain opening campuses in Edmond and Tulsa, Oklahoma, in August 2026. Students spend two hours each morning on the school's adaptive platform, thirty minutes per core subject, then move to project workshops in the afternoon. The adults are called guides. Per Oklahoma Watch's reporting, they do not teach curriculum, they are not required to hold teaching credentials, a bachelor's degree in any field is acceptable, and they are paid between $100,000 and $150,000. Tuition is $40,000. State Representative Cyndi Munson told the outlet that such schools are "taking away the seriousness of what it takes to be a classroom teacher."
Here, the AI is taking over the lecture and not the custodian. That is the model at the top of the market, where families are already leaving traditional schools for AI-forward private options.
At the other end, RAND's American School District Panel found that by fall 2024, 67 percent of low-poverty districts had provided teachers with training on AI use, against 42 percent of middle-poverty districts and 39 percent of high-poverty districts. Capacity to use the technology well is distributed the way everything else in American schooling is distributed. My worry for the next decade is stratification more than replacement. Wealthy families keep small human classes with AI software as an accelerant, while under-resourced districts get software instead of staff they cannot hire. AI is usually sold as an equalizer.
My forecast for 2040
Pretty Confident: AI absorbs planning, first-pass grading, differentiation, and initial tutoring across most well-resourced districts by 2040. Also confident: the classroom adult's role shifts toward coaching, behavior, and judgment, and evaluation systems lag that shift by years, because the metrics we rate schools on were built for the old job description.
Moderately confident: total teacher headcount falls at the margins rather than in the aggregate. The pressure points are secondary electives, credit recovery, and rural staffing gaps, in that order
Speculation: I do not think the adult leaves the room in any jurisdiction operating under a legal framework resembling today's. Supervision, mandated reporting, and liability for minors have no automated substitute any legislature has been willing to define, and defining one would mean deciding that a corporation can hold a duty of care to a child.
The teacher in that biology class had known the boy since August and she was in the room to see him. Better AI would have changed what she spent the other fifty minutes doing. Someone still has to be standing there to supervise.
Figures here come from the sources below, each checked in July 2026: Bloom (1984), the HarvardX and MITx first-year report (2014), Reich and Ruiperez-Valiente in Science (2019), Nickow, Oreopoulos and Quan (2024), RAND's district panel (2025), and Oklahoma Watch's July 2026 reporting on Alpha School. Forecasts are the author's own and labeled where they appear.
Sources
Quote Investigator: Books Will Soon Be Obsolete in the Schools
ERIC: Bloom, The 2 Sigma Problem, Educational Researcher (1984)
ERIC: Nickow, Oreopoulos and Quan, The Promise of Tutoring for PreK-12 Learning (2024)
MIT Teaching Systems Lab: The MOOC Pivot (Reich and Ruiperez-Valiente, Science, 2019)
RAND: More Districts Are Training Teachers on Artificial Intelligence (2025)
Oklahoma Watch: A $40,000-per-Year AI School With No Teachers Is Opening in Oklahoma
