I Don’t Want AI To Make My Students Average

Why I spent 70 pages teaching AI how to treat learners differently.

When I began using AI more seriously for lesson preparation, I expected to be frustrated by bad material.

That would have been easier.

Instead, AI produced useful material quickly.

It created lesson objectives, discussion questions, vocabulary exercises, writing prompts, model answers, homework, presentations, and feedback. A task that might have taken me an hour could happen in minutes.

That is pretty damn useful.

Then something started bothering me.

The Problem Wasn’t That AI Was Bad

The output was usually competent.

And increasingly, I realized competent was the problem.

The lessons looked like something a capable teacher could use with almost any student at approximately the right level.

Approximately.

That is a dangerous word in education.

The problem wasn’t that AI couldn’t create a good lesson. The problem was that it could create a good lesson without really understanding who the lesson was for.

It knew what a B1 lesson looked like.

It did not automatically know whether the learner in front of me needed more confidence, more challenge, more structure, more time, or a completely different example.

A reasonable lesson for a hypothetical learner is not necessarily a useful lesson for a real person.

That became more obvious the more I looked at actual learner work, recent notes, earlier reviews, lesson evidence, and current goals.

Those things tell the truth more reliably than labels do.

Age helps a little.

Grade helps a little.

A CEFR band helps a little.

A generic prompt helps less than we would like to admit.

The source of truth is the evidence about the person in front of the teacher: what they produced, what they understood, where they hesitated, what improved, what stayed shaky, and what they are trying to do next.

No Average Student Is Sitting In Front Of Me

Consider three learners I have taught. The details are intentionally general because their identities are not the point.

The first is an academically strong writer preparing for a more demanding academic environment.

Their English is already good. Giving them another list of vocabulary words may create activity, but not much growth.

They need to develop more nuanced claims, choose meaningful reasons, explain why evidence matters, connect ideas logically, and move beyond their first answer.

The challenge is not simply better English.

It is better thinking expressed through English.

The second is a younger learner whose current English skills do not yet match the academic expectations they are moving toward.

Here, personalization means something different.

Making everything easier might make today’s lesson more comfortable, but it could quietly preserve the gap we are trying to close.

The learner needs scaffolding without having the destination lowered.

That distinction is surprisingly difficult.

It also forces a distinction teachers need to guard carefully: language difficulty is not the same thing as thinking difficulty.

One learner may understand a sophisticated idea, see the pattern, and know exactly what they want to argue, but lack the English to express it clearly.

Another learner may speak smoothly, confidently, and at length, yet still need real support with evidence, analysis, organization, reasoning, or academic independence.

Weak English is not the same as weak thinking.

Fluent English is not the same as strong reasoning.

The third is an adult professional who already knows plenty of English.

They do not need another lesson explaining the past tense.

They need to handle the moment when someone challenges their proposal in a meeting. Or when they disagree but do not want to sound confrontational. Or when they do not understand what someone just said and need another ten seconds to think.

Their problem is not knowing English.

Their problem is using English when real life refuses to follow the lesson plan.

All three learners could receive technically correct English lessons.

They should not receive remotely similar ones.

Personalization is not putting someone’s name at the top of a worksheet.

It is understanding where they are, where they are trying to go, and what they need next.

And that understanding should begin with evidence, not assumption.

Current learner work.

Recent teacher observations.

Prior reviews.

What happened in the previous lesson.

The learner’s current goal.

The learner’s demonstrated needs.

That is the preparation starting point.

Not a vague idea of what students “at this age” or “at this level” usually need.

AI Has A Gravitational Pull Toward Average

Generative AI is very good at recognizing patterns and producing plausible responses.

Ask it to create a Grade 6 English lesson, and you will probably get something recognizable as a Grade 6 English lesson.

Ask for a B1 conversation lesson, and you will receive something that looks like a B1 conversation lesson.

That is a useful starting point.

But a learner’s level describes only part of a person.

Two B1 adults might differ enormously in confidence, vocabulary, listening comprehension, professional needs, interests, willingness to make mistakes, previous education, cultural communication habits, and ability to explain an idea.

The label is helpful.

The label is not the learner.

If I am not paying attention, AI can make lesson creation so easy that I stop asking the harder question:

What does this particular learner need from me today?

That is where efficiency can become dangerous.

The teacher gets faster.

The materials get prettier.

The workflow becomes smoother.

And the learner gets something increasingly generic.

That is especially risky now because AI can complete more of the visible task than it used to.

If a system can generate the worksheet, propose the paragraph, suggest the transition, or produce the model answer, then the remaining value shifts upward.

Judgment matters more.

Reasoning matters more.

Explanation matters more.

Evaluation matters more.

Adaptation matters more.

The human work is increasingly in deciding what fits, what does not, what the learner still does not understand, and what practice would move them forward.

My Instructions Started Getting Longer

At first, my instructions were simple:

Create a personalized lesson plan based on the learner’s level, goals, and previous lesson.

Sounds reasonable.

It was not enough.

AI would occasionally reduce academic difficulty because a learner struggled. So I added a requirement:

Scaffold difficulty without automatically lowering the learner’s target level.

It would create an activity without giving me enough understanding to teach the underlying concept confidently. Another requirement:

Explain the teaching purpose and concept to the teacher before providing the student-facing explanation.

Feedback could become a catalog of mistakes. So I added:

Recognize what the learner is doing successfully before identifying the next improvement.

Homework sometimes became generic extra practice. Another requirement:

Homework must reinforce the specific skills and thinking practiced during the lesson.

Adult learners were occasionally treated like school students. That required another reminder:

Adapt instruction according to proficiency and adult context rather than translating school-grade expectations onto adults.

And on it went.

Every failure became a requirement.

Every useful distinction needed a place.

At some point, I realized I had written dozens of pages telling one of the world’s most sophisticated AI systems things like, “Please do not make the student worse at English.”

That is a slight exaggeration.

Not much of one.

What began as a prompt became a specification. Then a framework. Then something resembling a small teaching operating system.

Eventually, it reached roughly 70 pages.

Why Does Personalization Need 70 Pages?

The answer is not really AI.

Teaching contains an enormous number of small judgments that experienced teachers make almost unconsciously.

Is the learner confused, or merely searching for a word?

Should I correct this mistake now, or let them finish?

Is the exercise too difficult, or is productive struggle happening?

Does the learner understand the concept but lack the language to express it?

Have I explained enough?

Am I talking too much?

Should we practice again or move forward?

Does this learner need correction, encouragement, challenge, or simply a few seconds?

Those decisions are easy to overlook because they often happen in the space between one sentence and the next.

They also shape what kind of difficulty we are actually seeing.

Sometimes the problem is vocabulary.

Sometimes it is sentence control.

Sometimes it is not language at all. It is reasoning.

A learner might write:

The character moved away from home. This shows change.

That is not exactly wrong.

It is just incomplete in the way many school answers are incomplete.

The real issue is often not “say more words.”

It is “make the reasoning visible.”

What changed?

Why did it change?

What caused it?

What does that result suggest?

Why does that support the claim?

A stronger response usually has a chain inside it: claim, reason, example, explanation, and connection.

Not because every paragraph needs to sound mechanical, but because readers need help crossing the gap between evidence and conclusion.

One diagnostic question I keep finding useful is:

What does the reader still need to understand?

If the reader still has to guess why the example matters, the reasoning is still doing half the job.

Another useful check is whether each sentence answers a question created by the sentence before it.

If a learner says something happened, the next sentence should help the reader ask why.

If they give a reason, the next sentence should help the reader ask how we know.

If they provide an example, the next sentence should explain what follows from it.

That small habit changes a lot.

It moves writing from reporting toward thinking.

And in an AI-heavy environment, that matters even more.

AI can now produce a passable example, complete a sentence, suggest a structure, or draft half a paragraph before a learner has finished adjusting their chair.

So the higher-value skill is increasingly not just producing text.

It is judging whether the text makes sense, explaining why one idea supports another, evaluating whether evidence is strong enough, adapting when the first answer is weak, and making the reasoning chain explicit rather than simply leaping from example to conclusion.

You cannot fully encode that judgment.

I do not think we should try.

The 70 pages are not instructions for replacing the teacher. They are instructions for making the preparation system behave more like a thoughtful assistant to the teacher.

That is a fundamentally different goal.

Then I Realized I Was Automating The Wrong Thing

The original attraction was obvious:

AI saves preparation time.

True.

But that increasingly feels like the least interesting benefit.

I am happy to automate:

• Reformatting lesson structures
• Recreating repeated exercise patterns
• Converting lessons into presentations
• Producing homework from material we have already covered
• Rebuilding rubrics
• Maintaining formatting consistency
• Organizing learner evidence, previous notes, and recurring patterns so I can review them faster
• Remembering every small requirement across multiple documents

Machines are quite welcome to those jobs.

I have spent much of my career doing similar work elsewhere: finding repetitive tasks, understanding why they exist, and then making systems handle as much of the repetition as reasonably possible.

That is my version of productive laziness.

What I do not want automated is:

• Listening
• Noticing hesitation
• Asking the unexpected follow-up
• Laughing when something goes sideways
• Changing direction because a learner suddenly becomes curious
• Recognizing frustration
• Challenging someone who can do more
• Backing off when someone needs another attempt
• Identifying the highest-value next ability the learner actually needs
• Asking questions that expose missing reasoning rather than merely missing words
• Interpreting whether the problem is language, thinking, confidence, or independence
• Remembering what mattered to them last week
• Being present

Those are not inefficiencies.

That’s teaching.

Productive Laziness Is Really About Attention

I have written before about the art of productive laziness.

The idea is not to avoid effort. It is to invest effort where it reduces unnecessary repetition later.

Work hard on the right system once, then let that system make everyday work lighter.

In organizational settings, that might mean automating reporting so people can spend more time having meaningful conversations. In teaching, it means automating the repetitive preparation so I can spend more attention on the learner.

The scarce resource is not necessarily time.

Initially, automation appears to save me 30 minutes. But if I respond by cramming another three tasks into those 30 minutes, I have not necessarily improved anything.

The resource I actually want to protect is attention.

Attention to what happened last time.

Attention to what the learner is struggling with.

Attention to whether the lesson I planned is still the lesson they need.

Attention to the difference between a learner who is stuck and a learner who is thinking.

Attention to the evidence that should drive preparation in the first place.

AI can help me collect, sort, summarize, and format that evidence.

It can help me prepare cleaner materials from it.

What it cannot do on its own is determine the learner’s actual next step.

That still requires human interpretation.

Productive laziness isn’t really about doing less work. It’s about refusing to waste attention on work that doesn’t deserve it.

Good Personalization Requires Boundaries

AI personalization should not mean infinitely changing everything for everybody.

That becomes chaos with better formatting.

Some things need to remain consistent:

• Learning objectives
• Appropriate academic expectations
• Lesson structure
• Quality standards
• Feedback principles
• Progression expectations

Other things should adapt:

• Examples
• Scaffolding
• Language complexity
• Discussion questions
• Practice quantity
• Teacher explanations
• Homework
• Pace

This is similar to the idea of Sustainable Ownership. Ownership does not come from removing all structure. It comes from providing enough purpose, context, expectations, and freedom for people to make meaningful decisions.

Personalized learning works similarly.

The framework cheat sheet describes this through purpose, principles, and values.

The purpose gives direction.

The principles shape what we practice.

The values guide how we show up.

In a lesson, structure provides direction. Personalization provides room to move.

Too little structure creates confusion.

Too much structure treats everyone the same.

The aim is not to make every lesson different for the sake of difference. It is to make the next useful step different when the learner requires it.

The Teacher Still Has To Own The Lesson

A 70-page skill can still generate something wrong.

Possibly beautifully wrong, which is one of AI’s more polished talents.

That is why I remain responsible for asking:

• Does this make sense for this learner?
• Is the difficulty appropriate?
• Is this actually worth learning?
• Does the activity serve the objective?
• Am I teaching something I understand?
• What happened last lesson that should change this one?
• What am I seeing today that the system could not possibly have known?

AI can prepare the lesson.

It cannot take responsibility for the learner.

That is the teacher’s job.

The system can help me remember patterns. It can suggest options. It can produce a first draft far faster than I can.

It can also help organize the evidence I already have so preparation starts from something more solid than guesswork.

But it cannot notice the small change in someone’s face when an explanation finally lands.

It cannot know that the question a learner asked is more valuable than the exercise I prepared.

It cannot decide whether a pause means fear, confusion, fatigue, or careful thought.

It cannot reliably determine the highest-value next ability without a teacher interpreting the evidence in context.

Those moments still require a person who is paying attention.

I Don’t Want AI To Make My Students Average

My concern is not that AI will suddenly destroy education.

The risk is subtler.

AI makes it incredibly easy to produce acceptable educational material.

And “acceptable” can become a comfortable place to stop.

That is exactly what I do not want.

I want the academically strong learner challenged.

I want the developing learner supported without lowering the destination.

I want the professional practicing situations they will actually encounter.

I want homework responding to what happened in class rather than what a textbook says should have happened.

I want preparation to begin with the actual source of truth: learner work, recent observations, prior evidence, current goals, and demonstrated need.

And sometimes I want to abandon half the lesson because the learner asks a much better question.

Seventy pages probably sounds ridiculous for something that is supposed to make my work easier.

Maybe it is.

But those pages are not really about telling AI how to teach.

They are about repeatedly asking myself:

What does good teaching look like for the person sitting in front of me?

Then I make the boring parts easier, so I have more capacity actually to do it.

That is the trade I want from AI: not more generic output, more frantic efficiency, or a classroom full of people performing competence.

I want a preparation assistant, not a substitute for judgment.

I don’t want AI to help me teach more students. I want it to help me pay better attention to the students I already have.