A problem for people who use AI
Connecting time savings to a workflow people keep using
Primary school teachers in the UK preparing lessons, school staff in Japan guiding pupils' AI use, and government staff in the US
Lesson preparation time saved; long-term retention unconfirmed
The lesson-preparation trial published in October ran September–December 2025, and the staff survey is also a 2025 observation. The Japanese school-staff survey is April–May 2026.
Editorial check: 2026-10-23
What people are paying attention to
The lesson-preparation AI trial published in October tests time savings and materials. The US staff survey also examines perceptions before and after. The main observations are all from 2025.
What is hard
Teachers say there is little need to build lesson plans from scratch and that fixing takes effort. The Japanese survey of school staff also has the checking and assignment-design burdens that come with pupils' AI use.
Workarounds being tried
A 10-week trial of lesson-preparation AI and an eight-week trial by staff with training have been run. The latter's role-specific training and consultation desk are requests raised by users.
What the sources confirm
In teachers' self-records, lesson preparation time was about 24% shorter than the comparison group. Meanwhile, use averaged about 7% of a week's lessons and declined over the period. Time savings and continued use are separate results.
What is still unclear
In which lessons and roles it keeps being used, and how the burden of the whole job including fixes and new checks changes, are unconfirmed. The results do not show pupils' attainment or long-term retention.
Workarounds and results in more detail
Introduce lesson-preparation AI and compare
School-level randomized trial
Preparation time is self-recorded, averaging about 49 minutes less per week. Material ratings cover the final versions teachers edited.
Scope and limits: Primary schools in England. 464 teachers and 108 schools assigned; main analysis 416 teachers and 106 schools. Both groups could use other generative AI.
Try Copilot at work together with training
Eight-week trial and before-and-after survey
Perceived usefulness on a five-point scale went from an average of 3.85 to 3.62. The authors interpret this as a revision of expectations.
Scope and limits: A single US transportation department. 124 analyzed after matching and quality checks. This is not a measurement of productivity or accuracy.
Read with care
- Time saved in lesson preparation and declining use coexist in the same trial. The results do not measure AI's raw output or pupils' attainment.
- The staff survey is a self-report at a single organization with no comparison group, and a preprint not confirmed as peer-reviewed. A change in perception is not a measurement of lower productivity.
- The Japanese burden survey covers different work and subjects, and cannot be compared as an amount offsetting the UK trial's time savings. The number of respondents to the burden questions is also unconfirmed.
What to check next
Records following which work continued to use it and which stopped, the reasons, the time spent fixing after generation, and long-term use and the burden of the whole job.
Generate a lesson plan, then finish it into materials that suit your class
In the comparative trial in which teachers of Years 3–6 in England tried Aila for 10 weeks, preparation time fell by about 49 minutes a week on average. Some users valued it when thinking up lessons on unfamiliar topics. However, they usually prepared by combining existing materials and other AI, and chances to build a whole lesson from scratch were limited.
Some said the output was too generic and lacked flexibility, and teachers edited it inside the tool or after downloading. Use in an average week's lessons stayed at about 7% and declined during the period. The editorial hypothesis to test is to look at each step—finding materials, structuring, adapting to the class, and checking—and decide where AI fits. The time-saving result concerns the preparation work in this trial and does not show pupils' attainment or long-term retention.
Questions to think with
- In your recent lesson preparation, how much was building anew and how much was fixing existing materials, and how many minutes did fixing and checking the AI draft take?
- In cases where the AI draft was not adopted, which decided it: accuracy of content, fit to the class, or editing effort?
Editing that checked the primary sources. The questions to check in your own work are proposals for verification.