Sixty-eight English secondary schools in the EEF Teacher Choices trial
Verified evidenceStep before the decisionCreator to judgeTeachers can use a general model for preparation while retaining curriculum, quality, and classroom decisions.
Enrollment · England, United Kingdom
Collections: Human still decides
Executive brief
Bounded preparation tasks can save time without degrading reviewed artifacts, but student outcomes remain a separate question.
AI value · Weekly Year 7-8 science planning time
Verified56.2 minutes in the ChatGPT group, 25.3 minutes or 31% lower.
Teachers used ChatGPT for roughly one-third of relevant lessons, so this is not whole-workload reduction.
Before
Creates questions, quizzes, activities, and differentiated resources without generative AI. → Reviews and adapts prepared resources.
After
Uses an implementation guide and prompts the model for questions, quizzes, activity ideas, or tailored materials. → Reviews, corrects, and decides whether to use each resource.
Human boundary
Teachers decide prompts, edits, suitability, and classroom use.
Why it matters
Teachers can use a general model for preparation while retaining curriculum, quality, and classroom decisions.
Before
How the work ran before the change.
Science teacher
Creates questions, quizzes, activities, and differentiated resources without generative AI.
ControlTeacher judgment and school curriculum.
Teacher
Reviews and adapts prepared resources.
ControlTeacher quality assurance.
What changed
Teachers can use a general model for preparation while retaining curriculum, quality, and classroom decisions.
Decision rightHuman decision retained
After
How the same work runs now.
Science teacher with ChatGPT
Uses an implementation guide and prompts the model for questions, quizzes, activity ideas, or tailored materials.
ControlUse limited to lesson preparation.
Teacher
Reviews, corrects, and decides whether to use each resource.
ControlTeacher retains pedagogical authority.
Exception path
Teachers reject poor output and prepare the resource conventionally.
Work removed
Decision authority
Teachers decide prompts, edits, suitability, and classroom use.
Before
Science teacher
Creates questions, quizzes, activities, and differentiated resources without generative AI.
Control: Teacher judgment and school curriculum.
Teacher
Reviews and adapts prepared resources.
Control: Teacher quality assurance.
After
Science teacher with ChatGPT
Uses an implementation guide and prompts the model for questions, quizzes, activity ideas, or tailored materials.
Control: Use limited to lesson preparation.
Teacher
Reviews, corrects, and decides whether to use each resource.
Control: Teacher retains pedagogical authority.
Work that left the path
Human role before
Teachers generated preparation materials directly.
Human role after
Teachers direct and quality-assure generated resources and reinvest saved time.
AI roleDecision mode: draft generation for lesson preparation; no grading or student decision authority.
81.5 minutes in the non-GenAI comparison group.→56.2 minutes in the ChatGPT group, 25.3 minutes or 31% lower.
Weeks 6-10 of a ten-week 2024 trial. · 259 teachers in 68 schools; school-randomized.
Teachers used ChatGPT for roughly one-third of relevant lessons, so this is not whole-workload reduction.
Blind expert ratings of non-GenAI resources.→No apparent reduction in quality for ChatGPT-assisted resources.
2024 trial. · Expert panel blind to trial allocation.
Resource quality is not the same as student learning outcome.
Anti-pattern
Claiming reduced planning time proves improved learning.
Questions
Portability conditions
Reputation risk
low
Evidence and authority
Current · updated
1 independent; publication outcomes are verified.
Bundle 1.0.0 · reviewed 2026-08-23 · stable ID a7cc28afea359c61
TD Bank Group · Step before the decision
That preparing a mortgage or HELOC file for a credit decision requires a human to read the document package end to end - classifying documents, extracting figures, calculating income, checking consents and writing the summary - because borrower documents are too variable for a rules engine to handle reliably.
American Airlines · Step before the decision
Live flight, runway, and gate data can recommend the nearest gate; operations teams retain disruption control.
John Hancock / Manulife · Step before the decision
A licensed underwriter need not assemble every preliminary before an agent gets a non-binding indication.
Boston Public Schools · Coordination compressed
Routes can be optimized as one district system instead of built school by school and then combined.