Will AI replace teachers? What the trials and the jobs record show (2026)

If you teach, you have heard the pitch. Every student gets a patient tutor on a screen, the lesson plans write themselves, and the classroom needs fewer adults.

There is a record to test that against, and it points somewhere less dramatic. The trials that measured AI in learning got real gains, and in both of them a person was still running the session.

Start with the jobs outlook. The US projects high school teacher employment to show “little or no change” from 2025 to 2035 [source]. The reasons it gives are enrollment and budgets, not AI.

What “will AI replace teachers” actually asks

Will AI replace teachers asks whether software can take over the work a school pays a teacher for: deciding what a class needs, running the room, and answering for whether students learned.

Producing explanations is the easy part, and models already do it. The harder part is the judgment and supervision around them. That is what this post tests.

It uses three TIN case files: a Danish register study that measured teachers’ pay and hours, a Stanford trial where AI coached tutors, and a World Bank trial where students used GPT-4 under supervision.

Will AI replace teachers in the future?

Not on the official US outlook, which does not name AI as a driver at all. It projects teacher employment roughly flat for demographic and budget reasons.

The Bureau of Labor Statistics counts 1,087,500 secondary school teachers in 2025 and projects 1,085,000 in 2035 [source]. For kindergarten and elementary teachers it projects 1,530,500 falling to 1,524,400 [source].

The agency’s stated reasons: “Projected declines over the decade in both the school-aged demographic and student enrollment should constrain demand,” and public hiring “may depend on state and local government budgets” [source].

It still projects about 63,500 high school openings a year [source] and about 99,400 elementary openings a year [source]. These are forecasts, not measurements, and TIN did not check them.

Has AI changed teachers’ pay or hours yet?

No, on the one register study that names teachers. NBER Working Paper 33777 surveyed workers in 11 exposed occupations, and “teachers” is on the list [source].

The latest round covers 25,000 workers across 7,000 workplaces, linked to Danish records of earnings and hours [source]. Adopters report saving about 3% of their work hours [source].

The authors find “precise null effects on earnings and recorded hours,” ruling out effects larger than 2% two years after ChatGPT [source]. It is a working paper covering the first two years, and these are pooled results, not a teachers-only estimate.

Will AI replace teachers in the classroom?

Not in the two trials that measured it. Both produced real learning gains, and in both the AI worked inside a session a person was running.

In Nigeria, a World Bank team gave first-year senior secondary students Microsoft Copilot, powered by GPT-4, for English over six weeks [source]. Students worked in teacher-supervised computer-lab sessions ([source]).

The result was 0.31 standard deviations on the overall assessment and 0.23 standard deviations on English [source]. That is a large effect for six weeks of schooling. It was not a test of a chatbot without a teacher.

What happens when AI coaches the teacher instead?

The weakest tutors improved the most. Stanford’s Tutor CoPilot gave live tutors real-time AI suggestions, and the tutor, not the model, worked with the student.

The trial involved 900 tutors and 1,800 K-12 students [source]. Students whose tutors had the tool were 4 percentage points more likely to master topics, and students of lower-rated tutors gained 9 percentage points [source].

The cost was “only $20 per-tutor annually” [source], a figure K-12 Dive reported independently ([source]). Tutors also flagged suggestions “not grade-level appropriate” [source]. The person in the loop was catching the model’s mistakes.

The proof

TIN’s case file on the Danish study quotes both versions of the paper and prints where they differ: chatbots, pay and hours across 11 occupations, teachers included.

TIN’s case file on Tutor CoPilot checks every figure against the paper and two independent outlets, and names the limit that it is one unreplicated working paper: AI coaching for live tutors.

TIN’s case file on the Nigeria trial quotes the results tables and shows the press rounding conflict rather than merging it: six weeks of supervised GPT-4 tutoring in Edo State.

The tutoring-tool question, AI against a human tutor, is argued separately in is AI tutoring better than human tutoring. None of these files counts teaching jobs lost to AI, and this post claims no such number.

What each record shows

The bottom line

The forecast that AI replaces teachers treats teaching as explanation delivery, and models are already good at explaining. The record says the measured gains arrive when a person decides what the session is for, supervises it, and catches the model when it is wrong. Neither trial removed that person, and the official outlook does not even list AI as a reason teacher jobs will change. The useful test for any education AI is not how well it explains.

It is whose judgment it strengthens, and the strongest result so far went to the weakest tutors.