The Formulaic Trap: Why AI Finds Your Assignments Easy

My favorite part of teaching middle school was the freedom to experiment with instructional design.

I scrapped the standard book report and replaced it with a graphic organizer that required students to present the key ideas in a multi-modal response using no more than 100 words. The related oral reports were governed by a two-minute time limit. A student earned bonus credit later if I spotted anyone during SSR reading a book they had recommended.

The rule for my project-ending summative exams was simple: You can use your notes, my slide decks, our textbook or any other resource you value to complete the test. I did not care about recall — only analysis, synthesis, and evaluation. 

Proudly, in 12 years of classroom teaching, I never assigned a five-paragraph essay. As a professional writer who published his first work exactly 50 years ago, I detest the format. Yes, structure matters, but not one that students will never encounter outside of the classroom. 

Recent pieces published by the crew who write about AI and education on Substack revived my interest in instructional design. This experience generated an idea: What if I asked all the major large-language models (LLMs) what features of an assignment make it so easy for them to produce high-scoring student work?

I suspected there were formulaic features LLMs could identify and exploit. After all, their core function is pattern recognition.

The standard response to cognitive offloading is to blame the technology. But that diagnosis misses the point. The real issue is not that large AI models are “too good.” It’s that many of our assignments were designed in ways that made them predictable, procedural, and ultimately automatable.

Education should learn a simple lesson from industry: If a task can be automated, it will be automated. 

Let’s examine the features of formulaic instructional design — and how to break them.

The Five-Feature Formula

We didn’t just accidentally create AI-friendly tasks. We engineered them. Let’s start with a hard truth: LLMs thrive on structure. Not deep understanding, not lived experience — structure.

Many traditional assignments contain a set of recognizable, repeatable elements:

1. Fixed Organizational Templates

The five-paragraph essay is the poster child here. Introduction with thesis, three body paragraphs, conclusion. Swap in a topic, and the structure holds.

LLMs excel because they’ve seen millions of variations of this exact format. They don’t need to think; they just instantiate the pattern.

2. Predictable Prompt-Response Pairings

Prompts like “Explain the causes of the American Revolution” or “Describe the water cycle” are not open questions. They are retrieval cues.

The model maps the prompt to a well-worn response pathway. This is less generation and more completion.

3. Surface-Level Cognitive Demand

Many assignments ask students to summarize, define, or list. These are what cognitive scientists would classify as surface-level tasks.

Effective instruction requires matching strategies to learning depth: surface, deep, and transfer. LLMs dominate the surface level because that’s where pattern density is highest.

4. Clear Success Criteria (Sometimes Too Clear)

Rubrics that specify required elements:

• Include three examples
• Use at least five vocabulary words
• Provide textual evidence

These are helpful for students, but they are also algorithmically friendly. LLMs can reverse-engineer the rubric and optimize output accordingly.

5. Decontextualized Knowledge

Assignments divorced from a student’s lived experience remove the one variable AI cannot replicate: personal context. That’s why my earlier graphic organizer required students to link via explanation the book they were reading to a favorite TV show, movie, song, or video game.

When the task is context-free, the AI has no disadvantage.

To understand why these assignments are so easy for AI, we need to revisit what LLMs actually do. They ingest vast amounts of text and learn the statistical relationships between words, phrases, and structures.

When a student is asked to write a persuasive essay, summarize a historical event, or generate a structured report, the model isn’t solving a problem — it’s completing a familiar pattern. This aligns with a broader truth in curriculum design: When tasks emphasize coverage over complexity, they become easier to replicate. This is true for both students and machines. LLMs are not cheating the system. They are revealing it.

The Backward Design Problem We Didn’t See Coming

Ironically, some of our best instructional frameworks contributed to this moment.

Backward design, for example, asks teachers to identify desired outcomes and then build assessments aligned to those outcomes. In theory, this creates coherence. In practice, it often leads to tightly scaffolded, highly structured tasks that signal exactly what success looks like. When both the task and the path to success are explicit, AI excels.

Keep in mind that backward design isn’t the problem. Poor implementation is. When outcomes are reduced to narrow, checklist-style criteria, the resulting tasks become highly predictable and thus highly automatable.

When the evaluation criteria are explicit and narrow — use a thesis statement, cite three sources, include a counterargument, stay between 500-800 words — LLMs can directly optimize for each item on the checklist. The rubric essentially becomes the prompt, and conformance is near-guaranteed. I’ve seen this firsthand sitting next to my nieces and nephews as they work on school assignments with an AI helper. The first thing they do is post the assignment directions followed by the rubric.

A checklist approach to assessment has been one of my pet peeves for decades. I argue that a long-term, complex assignment requires both checklists (required elements, genre, structure, and due dates) and quality indicators, best found in a rubric that focuses on the higher-level domains often associated with Bloom’s Taxonomy. 

Yes, the newer LLM models can mimic creativity, synthesis, and evaluation, which makes the tasks of curriculum and assessment design even more challenging. That reality requires us to understand the types of tasks that LLMs find difficult to complete. 

What AI Still Struggles With

Before we throw out every essay and worksheet, let’s be clear: LLMs are far less reliable when assignments require:

• Authentic ambiguity (no single right structure or answer)
• Local context (classroom discussions, community issues)
• Process visibility (drafting, iteration, feedback loops)
• Human judgment and decision-making
• Original synthesis under constraint

That last item deserves a quick explanation. Constraints are deliberate limitations that shape the task. They might include speaking to a specific audience (city council members), atypical format requirements (podcast, infographic, etc.), evidence requirements (conflicting sources or local data), point of view (argue from a stakeholder perspective), and process constraints (revise based on peer feedback and then defend your decisions.)

In short, AI falters when the task moves from pattern execution to meaning construction. If we accept that formulaic elements make assignments AI-friendly, the path forward becomes clearer: introduce productive friction. Not confusion. Not chaos. Friction.

A few weeks ago, I learned that Substack colleagues Mike Kentz and Doan Winkel have launched a new platform called AI Friction Labs. 

This platform replaces helpful AI with productive friction. Students engage in high-resistance simulations — Socratic dialogues, stakeholder negotiations — that force them to defend arguments, adapt under pressure, and refine reasoning in real time. Teachers get tools to design these experiences, along with a Cognitive Rubric that shifts assessment from final product to visible thinking.

This is where inquiry-based models — problem-based, project-based, and challenge-based learning — offer a useful contrast to formulaic curriculum. These constructivist approaches center on messy, real-world problems that resist templated responses.

As project-based learning demands, students must:

• Define the problem
• Identify what they know and need to know
• Generate and test solutions
• Iterate based on new information  

There is no five-paragraph shortcut for that.

Final Thoughts

The question isn’t, “How do we stop students from using AI?”

It’s, “Why were our assignments so easy for AI to complete in the first place?”

That’s a harder conversation, but a more productive one.

Because if an LLM can complete a task flawlessly in seconds, we should at least ask whether that task was ever measuring what we thought it was.

The assignments that trip up LLMs are the opposite — tasks that require genuinely novel ideas, specialized expertise with limited prior examples, or deeply personal lived experience unavailable in any training set.

For educators, this is less a crisis and more a moment of clarity. The goal isn’t to outsmart AI with more complex prompts. It’s to design learning experiences that require something AI cannot easily replicate: thinking that is situated, iterative, and deeply human.

And that, thankfully, is still largely beyond even the best model’s consistent reach.