Lesson planning with AI: evidence and limits

Generative AI can reduce some of the time spent preparing materials, but the available evidence is still limited. It can produce a first draft of activities, examples, or questions for a lesson. Choosing objectives, sequencing the work, and providing support for a particular class require information the system does not have.

In Chile, full-time teachers reported spending 8.7 hours per week preparing lessons and 30.3 hours teaching, according to TALIS 2024. These figures describe teachers’ reported time. They do not measure planning quality or the effect of AI on students’ learning.

Planning for a particular class

Chile’s Marco para la Buena Enseñanza places planning in teachers’ knowledge of students, the curriculum, and the local context. Planning combines the writing of a worksheet with decisions about what students should understand, which prior knowledge they will use, which obstacles are likely, and how progress will be observed during the lesson.

A model can suggest an activity about equivalent fractions. It does not know which procedures the class used the previous week, which representations have proved difficult, or which student needs particular support. Those decisions need to be made before asking for a draft. Without them, generated text may seem coherent while offering little value to the actual group.

It is also useful to distinguish between a document and a plan. An orderly outline, a sequence of questions, or a list of materials are documents. Writing them is only one part of planning, which includes the decisions that give those materials a purpose. AI can assist with the first group of tasks; the teacher remains responsible for the second.

A trial of science lesson preparation

The EEF and NFER Teacher Choices trial examined ChatGPT for preparing science lessons and resources. It involved 259 teachers from 68 state-funded secondary schools in England. The trial lasted ten weeks and focused on Key Stage 3 science, with a guide for participants.

Teachers using ChatGPT reported 56.2 minutes per week on lesson and resource preparation, compared with 81.5 minutes in the comparison group. The average difference was 25.3 minutes per week. The project also reported that an expert panel reviewed materials without knowing which group had produced them and did not detect a difference in quality.

The result describes one specific intervention: science teachers in England, using ChatGPT and a guide for ten weeks. It does not show that the same time saving will occur in every subject or age group. It also does not measure student learning or show how teachers used the recovered time. In Chile, TALIS reports working conditions and reported time, but it does not provide a comparable trial of AI-supported planning.

What to check in a draft

A useful draft may contain an activity that fits the timetable and available resources. It still needs review. The first review concerns content: concepts, examples, instructions, and expected answers. The second concerns pedagogical fit: the relationship to the objective, level of difficulty, prior knowledge, and opportunities to participate. The third looks for omissions: support for students who need it, materials, time for closure, and a way to check what students understood.

In a 45-minute lesson on equivalent fractions, a request can specify the objective, class, prior knowledge, and one concrete constraint. For example, students might explain with drawings why one half and two quarters represent the same quantity; they may already have worked with partitions of shapes; the lesson may use paper, pencils, and the board; and some students may need visual examples before writing. The response can propose a sequence. The teacher decides whether that sequence suits the class and changes what is needed.

A short request often leaves out conditions that alter a lesson. A model does not know whether an activity overlaps with an upcoming assessment, whether the group needs more time to read instructions, or whether an example may cause confusion. Providing that context improves the draft, but it does not remove the need to review it.

A short procedure

A limited use can follow four steps. First, define the learning objective and constraints: class, time, materials, prior knowledge, and required support. Second, request a concrete proposal, such as a sequence of activities or questions. Third, check the content, fit for the class, and missing elements before using it. Fourth, remove any information that identifies students from the request. General descriptions, such as “a student needs visual support,” are usually enough to frame an initial adaptation.

This procedure treats the output as working material. Its usefulness depends on the information in the request and on subsequent review. The available evidence supports a claim about time saving in a limited trial. Studies still need to describe effects on planning and learning in Chilean schools.