The AI tools are here, the students are using them, and most classroom guidelines boil down to “be careful” and “don’t cheat.” This is not pedagogy – this is hope.
The real challenge isn’t teaching students to use AI safely. It teaches them to think rigorously in a world where cognitive shortcuts are free and instant. This requires understanding what AI actually disrupts in learning, then designing instruction that addresses that.
What AI is changing in learning
The generation effect—one of the most robust findings in cognitive science—tells us that actively producing information creates stronger memory traces than passively acquiring it. Struggling to retrieve an answer, even if unsuccessful, improves later learning more than providing the answer immediately.
AI reverses this. It removes the productive struggle by default. When a student asked ChatGPT to explain the symbolism in The Great Gatsbythey get a fluent, confident response without doing the interpretive work that builds literary thinking. The result looks like understanding.
It isn’t.
This does not make AI useless in classrooms. This makes the design question more acute: How do we position AI so that it augments cognitive work rather than replacing it?
A framework for positioning learning
Consider AI tools on a continuum based on when students encounter them in the learning process:
AI after thinking — Students first develop their own analysis, argument, or solution. They then consult the AI to compare, challenge or expand their thinking. This preserves the generation effect while adding a feedback mechanism.
AI as foil — Students rate, critique, or improve AI-generated content. This works because critical analysis requires understanding—you cannot identify what is wrong or weak without knowing what is right and strong.
AI as a collaborator — Students work iteratively with AI, but with explicit metacognitive checkpoints: What did I contribute? What has AI contributed? What do I really understand now? This requires complex facilitation and works best with students who have already developed knowledge of the area.
AI as a substitute — Students delegate thinking entirely to AI. This has legitimate uses (accessibility, efficiency for low-stakes tasks), but does not lead to learning. Be honest with students about when it is and when it is not appropriate.
Progression matters. Students need experience in the first two modes before they can use AI as a true collaborator rather than a crutch.
Three protocols that actually work
Protocol 1: Prognosis before consultation
Before students ask the AI, require a written prediction: What do you think the answer is? why Rate your confidence 1-5.
After consulting the AI, they return to their prediction: What did you get right? what did you miss If your confidence was high and you were wrong, what does that tell you?
This exploits the hypercorrection effect—high-certainty errors, once corrected, are better remembered than low-certainty errors. It also builds calibration, the metacognitive skill of knowing what you know.
Performance Note: This works for factual and conceptual questions, not open-ended creative tasks. Keep predictions short – one to two sentences. The goal is to activate prior knowledge, not create busy work.
Protocol 2: The revision stack
Students write a first draft without access to AI. They then prompt the AI for feedback on a specific dimension (argument structure, use of evidence, clarity). They revise based on this feedback, documenting what they changed and why.
The main limitation: students must be able to explain and defend each edit. If they can’t articulate why a change improves the piece, they don’t make it.
This constructs revision as a thinking skill rather than a compliance task. It also exposes students to the difference between surface editing (AI is good at this) and substantive revision (AI suggestions often smooth out voice and homogenize arguments).
Performance Note: Limit AI consultation to one dimension per revision cycle. “Make this better” produces a common polish. “Identify where my argument assumes something I haven’t proven” is thought provoking.
Protocol 3: The Competition Note
Set position. Students explore and develop their argument without AI. They then have the AI generate the strongest possible counterarguments to their position.
Their final task: to respond in writing to these counterarguments. Who have merit? Who can refute? Which ones require them to change their original position?
This works because it’s really hard to generate strong counterarguments for your own position – motivated reasoning gets in the way. AI does not have this bias. This will create challenges that students would not have thought of on their own.
Performance Note: This requires students to first have a developed position. Using it too early just causes whiplash, as students jump between AI-generated viewpoints without developing their own.
The Harder Conversation
Most guidelines for AI in education avoid an uncomfortable reality: these tools will render some traditional assessments meaningless. The five-paragraph essay assigned on Monday and Friday is now dead; we just haven’t buried it yet.
That doesn’t mean writing is dead. That’s what it means unobserved, product-focused written assessment is dead The answer is not to disable AI or install detection software (which doesn’t work reliably anyway). The answer is to shift to:
- Process documentation that makes thinking visible
- In-class writing where you can observe students’ actual composition choices
- Oral exam and defense of written work
- Assessments where access to AI is assumed and the task is designed accordingly
The goal was never the essay. The goal was the thinking that the essay was supposed to develop and demonstrate. If AI breaks this proxy, we need better proxies or we need to assess thinking directly.
What students really need to understand
Forget “AI can be wrong”. Students hear this and think it means random factual errors that they can check with Google. The actual problems are more subtle:
AI is convincingly wrong in ways that are hard to detect without experience. This is not a signal of insecurity. It will explain a concept incorrectly, using all the correct vocabulary, and the novice cannot tell the difference between this and the correct explanation. This is an argument for building knowledge before relying on AI for a given topic and not after.
AI results reflect training data patterns, including outliers and omissions. Ask him about well-documented topics and you’ll get a reasonable synthesis. Ask for anything specialist, recent or contested and the quality plummets. Students should develop an intuition about which queries are likely to produce reliable results.
Fluency is not understanding. That’s the most important thing. Students can read an AI explanation, feel they understand, and be completely unable to reconstruct that understanding without the AI’s help. The feeling of learning is not the same as learning. The only way to know if you’ve learned anything is to test yourself without the tool available.
The measure of equity
Access to AI at home is unevenly distributed – not only from access to the device, but also from the knowledge required to use these tools effectively. Students whose parents can teach fast engineering have an advantage over students whose parents don’t know about the existence of ChatGPT.
If AI literacy matters, it should be taught in school. If working with artificial intelligence becomes standard, students need practice time in class, not just at home. This is not optional work in fairness related to the real curriculum. Of primary importance is whether the curriculum serves all students.
