Planning with AI is not cheating: it's exercising professional judgment

Lesson planning takes time that the system doesn't give you. According to TALIS 2024, Chilean teachers spend 8.7 hours per week on planning — 1.3 hours more than the OECD average — while also carrying the highest teaching load of any country in the study: 30.3 hours of instruction versus an OECD mean of 22.7. Given that margin, the question isn't whether you should look for better ways to plan. The question is how honest you are about how much of your planning time goes into the real pedagogical work versus formatting, searching for resources, and writing down things you already know. Here is the thesis: using AI to prepare lessons is not a shortcut or a delegation of your profession — it is a legitimate pedagogical decision that frees up cognitive time for the kind of judgment that Chile's Marco para la Buena Enseñanza (MBE) demands, as long as the teacher defines the objectives, course context, and differentiations, and the AI only drafts what the teacher has already thought through.

That distinction matters. The MBE 2021, in its Domain A, defines planning as a contextualized professional judgment: deciding what to teach, why, how, and for whom, with specific knowledge of that group of students in that context. AI cannot do that work because it requires you to know your class — their assessments, their rhythms, what happened last week. What AI can do is draft the document for what you have already thought. The problem is not the draft. The problem is not thinking before you ask for it.

What the available evidence actually measures

The most rigorous trial published to date is the Education Endowment Foundation and NFER study in England (2024): 259 science teachers across 68 secondary schools, randomly assigned, followed for 10 weeks. The group using ChatGPT to prepare Key Stage 3 units spent 56.2 minutes per week on planning, compared to 81.5 minutes for the control group — a 31% reduction. A blind panel of five expert teachers reviewed samples from both groups and found no statistically significant quality difference. The EEF itself flags that the quality finding rests on a limited sample, which is worth keeping in mind. But the time-saving result is robust.

A separate study in government schools in Karnataka, India (Shiksha Copilot, Microsoft Research, preprint CSCW 2025) tracked a planning assistant with 1,043 teachers and found savings of 2.02 hours per week, lower reported stress, and higher adoption of activity-based teaching methods. The low-resource public school context is closer to the reality many teachers face than the English trial.

In Chile specifically, 55% of teachers already use AI in their work — nearly 20 percentage points above the OECD average of 36% — and of those, 68% use it specifically for lesson planning, according to TALIS 2024. Chile's national teacher professional development agency (CPEIP) leads a national AI training program, and in March 2025 the Ministry of Education published the guide "PotencIA el Aprendizaje" to support this integration. The practice has explicit institutional backing. What's missing is guidance on how to do it well.

The real risk isn't the one most people name

The most common objection is: "AI makes you lazy." There is something to it, but it points at the wrong target. The risk is not that the teacher becomes lazy — it's that the fluency of generated text can produce a feeling of comprehension without actual comprehension. A study with 117 university students (Fan et al., 2024, British Journal of Educational Technology) found that the group using ChatGPT to produce an essay scored higher grades but showed no improvement in knowledge transfer or self-regulation compared to groups that didn't use AI. The text looks good. The reasoning wasn't necessarily built.

The same applies to a teacher reviewing a generated draft: if you arrive at the plan without having thought through the objectives first, the smooth AI text will appear correct because you have no position of your own to compare it against. This is why the condition isn't "review the draft." The condition is arriving at the prompt with the objectives already articulated, the course context already read, and the differentiations already considered. If you meet that condition, the AI draft is exactly the same as using a template or asking a colleague to help organize what you already thought through. If you don't, the deskilling risk flagged by the OECD Digital Education Outlook 2026 — erosion of professional judgment through delegation to AI systems — is entirely legitimate.

There is another risk the evidence does document: AI-generated lesson plans are not neutral. An analysis of 310 civics lesson plans generated by AI (CITE Journal, 2025) found that only 2% included higher-order thinking skills according to Bloom's taxonomy, and that historical figures associated with civil rights were systematically absent. A separate preprint (arXiv:2511.19482) documents specific failures in Global South contexts. Under a supervised workflow, these biases are detectable and correctable. Without supervision, they are not.

What we don't know and haven't tested

Being honest about limits matters as much as citing benefits.

First: there is no RCT or peer-reviewed study on AI-assisted lesson planning conducted in Chilean schools. TALIS 2024 data on adoption describes usage and perception, not impact on lesson quality or student learning.

Second: the EEF/NFER RCT measures that 31% of planning time is saved, but does not measure where that recovered time goes. There is no evidence it is redistributed toward formative feedback or student interaction. It is equally possible it gets absorbed by other administrative tasks.

Third: studies on "metacognitive laziness" from AI use were conducted with students, not teachers. There is no longitudinal data on whether teachers who use AI over years lose the capacity for autonomous planning.

Fourth: Chile's CPEIP and Ministry of Education programs are underway but have not published outcome reports as of this article.

Fifth: if most teachers use the same models with similar prompts, there could be systemic curricular homogenization at scale — a risk that the literature has not yet measured.

What is actually worth doing tomorrow

Before opening any AI tool, write out — by hand or in a separate document, not in the chat — your answers to three questions:

1. What do I want my students to understand by the end of this lesson, not just be able to repeat?
2. What do I know about this specific class that changes how I need to frame this? (Assessments, what they struggled with in the previous unit, what activates them.)
3. Who is going to need a different version of the activity, and why?

With those three answers written down, open the chat and write a prompt that includes them. For example: "Prepare a 45-minute lesson on equivalent fractions for a mixed-ability grade 5 class. The central objective is that students understand why two different fractions can represent the same quantity — not just that they apply the procedure. Include a variant for students who struggle with abstract reasoning. Do not use activities requiring materials other than pencil, paper, and the board."

What comes back is a draft. Read it the way you'd read a lesson plan from a colleague: with judgment, not reverence. Change what doesn't fit your class. Cut what sounds generic. Add what only you know about that group.

A concrete note on privacy: do not paste identifiable student data into any free AI tool — learning support assessments, family situations, names associated with specific difficulties. "A student who struggles with abstract reasoning" is sufficient pedagogical context and does not compromise personal data. Chile's Law 21,719 — the country's GDPR equivalent, effective from December 2026 — requires explicit consent for the processing of identifiable data on minors. The risk profile of "plan a fractions lesson for grade 5" is radically different from including personal student information in the prompt.

57% of Chilean teachers identify administrative overload as their main source of stress, according to TALIS 2024. Some of that time can be recovered. The condition is that the thinking work is done by you first.