My favorite part of teaching middle school wasn’t the grading. It was the chaos of experimentation.
I scrapped the book report. Dead on arrival. Instead, students used graphic organizers, limited to 100 words, for multi-modal responses. Oral reports had a hard two-minute cap. If I caught you reading during SSR? Bonus points. Simple.
Tests? Bring your notes. My slides. The textbook. Anything.
I didn’t care if you remembered dates. I wanted to see if you could synthesize, evaluate, and analyze.
In twelve years of teaching, I never forced a single student to write a five-paragraph essay.
Why?
I’ve been a professional writer since before most of these AI models were born. I detest the format. Structure matters. But structure that exists only inside a classroom? That’s a waste of time.
Recently, some Substack writers on AI in education reignited my interest in design.
I ran a test.
I asked the big LLMs one thing: What makes an assignment easy for you?
I knew the answer before I asked. Pattern recognition.
We blame the tech. That’s lazy.
The problem isn’t that AI is too powerful.
The problem is that we designed our assignments to be predictable, procedural, and boringly automatable.
Industry knows this: if you can automate it, it gets automated. Education is just late to the party.
The Formula AI Loves
We didn’t stumble into AI-friendly tasks. We built them.
LLMs thrive on structure. Not insight. Not life experience. Structure.
Here’s what kills student creativity—and feeds AI:
1. Rigid Templates
The five-paragraph essay is king. Intro. Three body paragraphs. Conclusion. Swap the topic. The shape stays.
AI doesn’t need to think. It just fills in the blanks.
2. Obvious Prompts
“Explain the Water Cycle.”
“List Causes of the Revolution.”
These aren’t questions. They are search commands.
The model maps the cue to the stored answer. It’s completion, not creation.
3. Shallow Thinking
Summarize. Define. List.
Cognitive scientists call this “surface” level. It’s where pattern density is highest. AI dominates surface. It barely tries on depth.
4. Checklist Rubrics
Include three examples.
Use five vocabulary words.
Helpful? Yes.
Algorithm-friendly? Terrifyingly so.
LLMs reverse-engineer your rubric. They optimize for points, not meaning.
5. No Context
When an assignment ignores who the student actually is, it becomes decontextualized knowledge.
I made kids link books to TV shows. To video games. To their lives.
AI has no life. When the task lacks context, AI has no disadvantage against you.
The Backward Design Trap
Irony? Our best frameworks helped this happen.
Backward design starts with outcomes. Then you build the test.
It’s coherent. It’s tidy.
And it’s terrible for thinking.
When you define “success” too clearly—thesis statement, three citations, 800 words—you’re just writing a prompt for an AI.
I’ve watched my nieces do it. First thing they do? Copy-paste the instructions and the rubric into ChatGPT.
Conformance is guaranteed.
A checklist isn’t enough. You need quality indicators. You need high-level domains from Bloom’s Taxonomy. Not just boxes to check.
Yes, new models mimic synthesis. Yes, they fake creativity.
It makes designing harder. It makes checking easier.
Where AI Stumbles
Don’t burn your curriculum. Just change the variables.
AI struggles with:
- Authentic ambiguity
- Local, specific context
- Visible process
- Human judgment
- Original synthesis under constraints
Constraints are key.
Talk to the city council? Hard.
Write a podcast? Weird for an LLM.
Use conflicting local data? Tricky.
Revise based on peer hate mail? Painful for a bot.
When a task moves from pattern execution to meaning construction, the bot falters.
We need “productive friction.”
Not chaos.
Friction.
Colleagues at AI Friction Labs are building this. They replace “helpful” AI with resistance. Socratic dialogue. Stakeholder negotiation. Defending your argument in real time.
They assess the thinking, not the PDF.
Project-based learning?
- Define the mess.
- Figure out what you don’t know.
- Test a solution.
- Fail. Iterate.
No template works here. No five-paragraph escape hatch.
So What Now?
We’re asking the wrong question.
Not “How do we stop cheating?”
But “Why was this assignment so easy to fake?”
That hurts. It should.
If a bot does it in seconds, were we measuring learning? Or were we measuring obedience to a formula?
The tasks that break AI are the ones that demand:
- Novel ideas.
- Specialized expertise.
- Lived experience no dataset has.
This isn’t a crisis. It’s a filter.
It shows us where our assignments failed.
We need to design experiences that require situated, iterative, human thinking.
That stuff is still hard for them.
Good luck.
