AI in Teaching · is it happening · updated quarterly

Is AI replacing teachers?

Updated Aug 2026Next review Nov 2026Method evidence first, then the mechanismSponsor none · nothing sold
Is it happening now?No.

No school system in the record has reduced teaching posts and attributed it to AI. Adoption is running ahead of policy: teachers use it weekly, most have had no guidance on it.

Could it?Parts, yes.

Grading, worksheet generation and online-only tutoring pass the replacement test below. Classroom teaching fails it. The one trial where an AI tutor moved learning had a teacher supervising every session.

What’s the real risk?Not replacement. Supervision.

The risk is the “one teacher, more screens” model: freed hours becoming larger classes or more sections, with the teacher supervising AI rather than teaching. Measurable, and rarely measured.

Is it happening? What the evidence would look like, and what it shows

If AI were replacing teachers, certain things would be visible by now. Each one, against what’s in the record.

If replacement were happening
What the record shows
Verdict
School systems cutting teaching posts, citing AI
None recorded. Teacher shortages are the dominant workforce story: UNESCO estimates 44 million additional teachers are needed worldwide by 2030.
No
Teachers using AI at scale
Yes: 30% of US public-school teachers use AI weekly and estimate 5.9 hours saved a week, about six weeks a year. ED-02 That is the “hours saved” this page is about.
Watch
AI tutors matching teachers on learning
The strongest positive result, a World Bank trial in Nigeria (~0.31 SD gain), ran under teacher supervision. ED-04 Unguarded AI access cut later exam scores by about 17%; a guardrailed tutor removed the harm. ED-05 The teacher, or the guardrail, is the active ingredient.
No
Evidence that AI improves outcomes without a teacher
Ofsted found little independent evidence that AI improves educational outcomes at all. TV-18 A Singapore vocational trial: better teaching process, no improvement in assessment; weaker students used it to skip thinking. TV-17
No
Systems planning for fewer teachers
Policy is behind, not ahead: only 18% of US teachers have received any formal guidance on AI, fewer than 1 in 10 for a specific task. ED-06 You cannot plan headcount around a tool you haven’t issued guidance for.
No
Firms and employers in general cutting for AI
AI-related employment decreases at 2% of US firms. BO-04 Vocational and public-facing roles are less exposed than average. TV-16
No

Reviewed quarterly. If a row changes to “Yes”, it will be the first line of the changelog, with the source.

Why AI replaces some jobs and not others

Jobs get replaced when three things hold at once. Classroom teaching fails all three; some of the work around it doesn’t.

1
Most of the job’s hours are in automatable tasks

A teacher’s day is managing thirty people, noticing who’s lost, adjusting mid-lesson, and dealing with parents. Planning, marking and materials are real hours, and they are the hours AI takes. Where the whole role is materials and marking, the test passes: content authoring, exam-marking pools, online-only tutoring.

Fails
2
The output can be checked cheaply

A wrong worksheet is cheap to fix. A wrong explanation absorbed by a class of thirteen-year-olds is not, and the Turkish trial is the evidence: unguarded AI help produced worse learning that only showed up on the next exam. ED-05 The check is the teacher in the room, and it can’t be done afterwards.

Fails
3
The employer keeps the saving as fewer people

A school keeps a saving as fewer teachers only by raising class sizes or cutting sections, both of which are contested, visible and often regulated. In shortage systems the saving goes to an unfilled post first. The exposed cases are private tutoring and online providers, where ratios aren’t regulated and the saving can be kept.

Fails

Run the tests on a marking contract, a worksheet-writing role or an online tutoring service and they pass. That is why the honest answer is “parts, yes”, and why the edges of the profession are moving while the classroom isn’t.

How it actually shows up: where the hours go

AI gives a teacher back about six weeks a year, on their own estimate. Where those weeks go is a decision the school makes, usually without saying so.

Hours savedLesson plans, marking, differentiated materials, parent emails. ~5.9 hours a week for weekly users. ED-02
→
More attention per studentSame class size, freed time on feedback and the students who are lost. What the tutor trials say actually moves learning.
or
More students per teacherLarger classes, more sections, or a teacher “supervising” AI-led sessions across rooms. No post is cut; the ratio moves.
Why this matters more than replacement

Replacement would be visible and fought. Supervision creep is quiet: a school adds an AI-tutor period, a teacher covers two rooms, and the tool is credited with the saving. The trials say the teacher is the active ingredient, so thinning the teacher out is the one thing the evidence warns against.

How to tell which is happening where you work

Two numbers before and after a rollout: students per teacher-contact-hour, and minutes of individual feedback per student per week. If the first rises and the second doesn’t, the saving went to the ratio.

What replacement would look like if it started

The leading indicators, so you don’t have to rely on us. Each is checked quarterly.

Would move first

  • Online tutoring services replacing tutors with a model plus a smaller escalation team.
  • Marking and moderation contracts moving to AI-first with human sampling.
  • “AI tutor periods” timetabled with one teacher supervising multiple groups.
  • A system reporting “teacher FTE efficiency” alongside an AI rollout rather than “hours saved”.

Would move last

  • Classroom posts, where the three tests fail.
  • Any model that has to show learning gains without a supervising teacher. None has. ED-04
  • Posts in shortage subjects and regions, where a vacancy absorbs the saving first.

What to do with this

If you’re teaching

  • Take the six weeks. Use them on feedback, which is where the evidence says learning moves.
  • Learn what your students’ AI tools do when unguarded. The 17% result is what you’re protecting against.
  • Keep a record of AI-generated material you corrected before it reached a class. It’s your case for the AI-lead and curriculum roles opening up.
  • Ask for the two ratio numbers above when an “AI period” is proposed.

If you’re training

  • Don’t pick a subject to escape AI. Every subject now has a “what do we do about AI” problem and few people to solve it.
  • Your programme should include designing a lesson that uses AI with guardrails. If it doesn’t, that’s the gap to fill yourself.
  • Teaching is in global shortage through 2030. That is the strongest protection a profession has.

For faculty of education, unions and ministries

Link or cite

  • The page keeps its URL. The evidence board and the changelog tell you what moved since you linked.
  • Cite a finding by record ID, which links to the primary source.
  • Nothing is sold and no programme or tool is recommended.

Correct us

  • A post cut and attributed to AI, a class-size change after a rollout, or a trial we’ve missed: research@straitsai.institute. Accepted corrections go in the next changelog with credit.

What changed this quarter

Aug 2026current

  • Evidence board: six rows reviewed, no verdict changed.
  • Gallup 2026 guidance survey added: 18% of teachers with any formal AI guidance. ED-06
  • Singapore vocational trial and Ofsted review added. TV-17 TV-18

May 2026retained

  • World Bank Nigeria tutor trial added as the supervised-AI benchmark. ED-04
  • PNAS Turkey trial added to test 2. ED-05

Questions teachers ask

Is anyone losing a teaching job to AI?

Not in the record. The moving edges are private and online: tutoring platforms and marking contracts. Those are jobs, but they are not classroom posts.

Why do the “AI won’t replace teachers” articles all say the same thing?

They say relationships can’t be automated, which is true and not the mechanism. The mechanism is that the teacher is the check on the tool, and the trials show the tool without the teacher makes learning worse.

What would change your answer?

A trial showing an unsupervised AI tutor matching a teacher on learning gains; a system reporting FTE reduction alongside a rollout; or ratios changing in a jurisdiction after an “AI period” policy. Any of those goes on the board as a “Yes”.

Can I cite this?

Yes. Cite the page with its updated date, or a finding by record ID, which links to the source.

Straits Institute for Applied AI · Industrial Research Unit · Singapore · Corrections: research@straitsai.institute

Record IDs point to entries in the State of Applied AI evidence base, each with its original source, date and link. Teacher shortage figure: UNESCO Global Report on Teachers (2024).