Evidence · strength 1 of 5 · searched to 2 Nov 2026
Does AI save teachers time on lesson planning?
Not enough evidence yet to say whether AI saves teachers time on lesson planning without lowering quality. Across 2 studies, preparation time changed by −31% to −24%.
What randomised trials found when teachers prepared lessons and resources with generative AI: time saved, the quality of what they made, and effects on students.
VerdictNot enough evidence yet
No good evidence. Studies so far point to a gain, but there are too few or too weak to rate it.
ImpactModerate gain
Across 2 studies, preparation time changed by −31% to −24%.
Evidence strength1 of 5
We have very low confidence in this estimate.
Cost per user a year$
One trial used the free version of a general AI assistant; the other a free tool from a public body (Oak National Academy).
Pilot rating under our v0.3 rules. The rules are published on our method page but not yet independently reviewed, so ratings may change as the method is checked and new studies arrive. Our method.
What the studies found
Time (headline)
Moderate gain−31% to −24%- Roy et al. (2024): −31% Lesson and resource preparation time, weeks 6-10 (weekly teacher diary), ChatGPT group as a share of the non-GenAI group (69%) (69% (95% CI 53% to 90%); p = 0.006; Hedges' g -0.38 (-0.65, -0.11)). Self-reported.
The lesson and resource preparation time spent by the ChatGPT group was 69% of the time spent by the non-GenAI group in weeks six to ten of the trial.
- Roy et al. (2026): −24% Weekly hours of lesson and resource preparation over ten weeks (diary), Aila as a share of non-Aila (76%) (76% (95% CI 64% to 89%); p = 0.001; standardised effect -0.22 (-0.35, -0.09)). Self-reported.
The lesson and resource preparation time spent by the Aila group was 76% of the time spent by the non-Aila group during the ten weeks of the trial.
- Checking time not measured: none of these studies counts the time spent checking the AI’s output separately.
Learning
No clear effect- Sungu et al. (2026): −0.02 SD Student standardised exam performance (spring term), AI v control, controlling for term 1 score (95% CI -0.088 to +0.050; p = 0.58). Not statistically clear.
we find no statistically significant effect of teacher AI access on academic performance, measured by students' standardized exam performance (β = -0.019 SD, 95% CI [-0.088, +0.050], p = 0.58
Quality
No clear effect- Roy et al. (2024): +0 no detectable difference in blinded panel ranks Quality of lesson resources ranked by a blinded panel of five science teachers (57 resources: 30 ChatGPT, 27 non-GenAI) (U = 184, p = 0.84). Not statistically clear.
The results from this test suggest there is no significant difference in ranks between lessons and resources provided by ChatGPT teachers and non-GenAI teachers, U = 184, p = 0.84.
- Roy et al. (2026): +1.01 points on a 0-25 quality score Lesson resource quality score (0-25) by six independent teacher-reviewers, 42 lessons; Aila minus non-Aila, non-inferiority margin -2.5 (95% CI -0.34 to 2.37; standardised 0.25 (-0.08, 0.58); non-inferiority p < 0.001). Not statistically clear.
Lesson quality score 21 17.7 (16.9, 18.5) 21 16.7 (15.6, 17.8) 42 0.25 (-0.08, 0.58) 1.01 (-0.34, 2.37)
Wellbeing
Worse- Sungu et al. (2026): −0.11 SD Student intrinsic motivation (composite of interest, importance and enjoyment items), teachers given AI v business-as-usual control (95% CI -0.199 to -0.023; p = 0.015). Self-reported.
Providing teachers with the AI teaching tool significantly reduces their students' intrinsic motivation (β = -0.111, 95% CI [-0.199, -0.023], p = 0.015)
- Sungu et al. (2026): −0.09 SD Student confidence (academic confidence and perceived progress), AI v control (95% CI -0.196 to +0.016; p = 0.097). Self-reported. Not statistically clear.
and marginally decreases student confidence (β = -0.090, 95% CI [-0.196, +0.016], p = 0.097)
A minus sign on time means the task took less time. Bands are provisional; see how impact is banded. Effects are not pooled yet: each band is set from one result per study, and where studies use different units we show only the direction.
How sure are we?
- Starting level3 of 5, from 3 grade A or B studies.
- Down one: old models0 of 3 studies show they used a model released since May 2025 (the 18-month window); the rest used older models or do not say which.
- Down one: self-reportedThe headline outcome is mostly self-reported rather than measured.
- Publication biasnot testable with fewer than 10 studies.
- Result1 of 5. We have very low confidence in this estimate.
Behind the average
- Teachers who find/create/adapt their own resources v those using pre-prepared resources (Roy et al. (2024)): No differential effect by planning approach (interaction p = 0.833)
- Teachers with low pedagogical confidence at baseline (Roy et al. (2026)): Planning time 0.63 times that of similar non-Aila teachers (37% less); no clear saving for high-confidence teachers (0.95)
- Teachers with fewer years of experience (Roy et al. (2026)): Planning time 0.68 times that of similar non-Aila teachers (p = 0.003); experienced teachers 0.87, not significant
- Teachers with high technological confidence (Roy et al. (2026)): Planning time 0.71 times that of similar non-Aila teachers (p = 0.019); low tech confidence 0.80, not significant
- Teachers with low subject knowledge (Roy et al. (2026)): Planning time 0.68 times that of similar teachers without Aila (p = 0.004).
How, where and why it works
Why it works
1 of 4. Teachers draft lessons and resources faster; the trials do not show which steps get quicker.
For whom and where
2 of 4. One trial found larger savings for less experienced teachers and those with lower teaching confidence.
What it takes
3 of 4. Both trials report training, guidance and how much teachers used the tool over ten weeks.
The full cost
2 of 4. The tools were free; the trials report minutes saved a week, from teachers' own time diaries.
Four separate scores from 0 to 4, never added up: how much the studies tell us about why it works, for whom, what it takes and what it really costs (the EMMIE method).
What to do next
Measure your pilot →If you are trying it yourself
- Use AI for a first draft of resources, then adapt it to your class.
- Keep a simple diary of planning time for a few weeks.
If you are rolling it out
- Give staff short training and a shared guide.
- Compare planning time and resource quality with a group not using AI.
What it takes
- Teachers in the trials got a short training session and guidance first.
- Savings appeared within the first weeks and held for ten weeks.
What can go wrong
- In one trial where teachers used an AI assistant for their teaching, students' motivation fell slightly, with no change in exam results.
- Time saved is from teachers' own diaries, not measured independently.
- Check AI-made resources for accuracy before use; quality was no better, only no worse.
The studies
| Study | Grade | Design | People | AI model (released) | Company role |
|---|---|---|---|---|---|
| ChatGPT in Lesson Preparation: A Teacher Choices Trial (Evaluation Report) Palak Roy et al. (National Foundation for Educational Research), 2024. Education Endowment Foundation evaluation report, December 2024 (government report) | B | field RCT, deployed work | 259 Year 7-8 science teachers in 68 English secondary schools randomised by school (34 schools/129 teachers ChatGPT; 34/130 non-GenAI); 211 teachers in primary analysis | ChatGPT 3.5, free version (OpenAI updated the free version to GPT-4o / GPT-4o mini during the trial) | none |
| Lesson planning using AI lesson assistant, Aila - A Teacher Choices Trial (Evaluation report) Palak Roy, Helen Poet, Ruth Staunton, Katherine Aston, David Thomas, Ben Styles, 2026. Education Endowment Foundation evaluation report, October 2026 (government report) | B | field RCT, deployed work | 464 Key Stage 2 (Years 3-6) teachers in 108 English primary schools randomised by school; 416 teachers in primary analysis | Oak National Academy Aila, built on OpenAI GPT-4o with retrieval and prompt engineering | none |
| Generative AI Can Harm Teaching Alp Sungu, Benjamin Lira, Angela L. Duckworth, 2026. SSRN working paper 7007339 (working paper) | B | field RCT, deployed work | 193 teachers (analytic sample) teaching 2,816 students in grades 5-12 across 14 Turkish K-12 schools; 85 school-department clusters randomised; 14,198 student-course observations; 2,343 students answered the endline survey (538 teachers randomised overall) | GPT-4o in a custom chatbot for teachers built by the research team (May 2024) | none |
Grades: A, a well-run randomised trial in real work; B, a strong natural experiment in real work or a well-run trial on realistic tasks; C, a trial with clear limits or a matched comparison; D, before-and-after, surveys or case studies. How studies are graded.
Studies we found but do not count
- Generative AI in Instructional Design Education: Effects on Novice Microlesson Quality (Moore et al. (2025), grade C). Not work: graduate students building micro-lessons in a course. The paper is paywalled; the figure comes from the university's project page. Shown as context, not counted (section 4, rule 1).
These studies are real and often cited, but they fall outside our rules for this rating. What counts as a study.
Does the rating hold up?
- Run without company-run studiesNo change: no such studies.
- Run without working papers and preprints2 studies: strength 1, verdict no good evidence, headline band moderate gain.
How up to date is this evidence?
Method →Evidence searched to 2 Nov 2026. Studies cover models released May 2024 to May 2024. Rating computed 5 Nov 2026 under Module 15 v0.3 (pilot).
How this page was made: AI research agents searched for studies and read each paper; a second pass checked every number on this page against the paper itself, with the quote and where it appears. The rating is computed by fixed rules, not chosen. No person outside AI Adoption has reviewed it yet. Editorial owner: AI Adoption editors. Spot a mistake? Corrections.
Disclosure: AI Adoption uses AI tools, including Claude from Anthropic, in its research and editorial work. No company pays for or has any say in a rating.