Master the Thinking Before You Hand It to AI

Summary for Educators

Based on Daniel T. Willingham
Students Should Only Use AI for Things They Already Know How to Do Well
Daniel Willingham, August 20, 2026

THE BIG IDEA

As schools wrestle with rules for student use of artificial intelligence, cognitive psychologist Daniel Willingham offers a remarkably simple principle: students should use AI only for things they already know how to do well.

His reasoning turns the usual conversation about AI upside down. In the workplace, the quality of the finished product often matters most. In school, however, the product is frequently just a vehicle for learning. Students write essays not because the world needs another ninth-grade essay, but because writing forces them to organize evidence, formulate arguments, reason, revise, and communicate precisely. They solve mathematics problems because the mental work builds mathematical understanding.

If AI performs that work before students have mastered it, the assignment may look better while the student learns less. Willingham therefore argues that AI should remove unproductive work, not the productive struggle through which learning occurs.

KEY TAKEAWAYS

• Learning requires mental work. When AI substitutes for the thinking an assignment was designed to elicit, students may complete the task without developing the intended knowledge or skill.
• School is different from the workplace. Workplace AI is generally used to improve efficiency and products. In classrooms, the process that produces the product is often the primary educational objective.
• The calculator analogy is useful. Once students have mastered arithmetic, calculators can eliminate tedious computation while allowing them to concentrate on higher-level mathematics. Using powerful mathematical tools before developing foundational understanding is another matter.
• Brainstorming and editing aren’t automatically appropriate AI tasks. Students learn by generating ideas, evaluating alternatives, revising language, and recognizing weaknesses. Turning those activities over to AI may surrender precisely the thinking teachers hoped students would practice.
• AI scaffolding presents a paradox. Novices may be the students least equipped to use unrestricted AI effectively because they don’t yet possess enough knowledge to judge its suggestions or formulate sophisticated questions.
• Even mastery doesn’t automatically mean delegation. Students who can perform a skill still need practice. Teachers must determine whether continued practice remains part of the learning objective.

WHY IT MATTERS

Many school AI policies focus on how much AI assistance is permissible: brainstorming is acceptable, editing may be acceptable, generating an entire paper isn’t. Willingham suggests that this may be the wrong organizing principle.

A more educationally useful question is:

What thinking is this assignment supposed to make the student do?

Only after answering that question should teachers decide whether AI belongs in the activity.

That distinction could significantly change classroom AI policies. Rather than labeling particular AI functions as universally “appropriate” or “inappropriate,” teachers would connect AI permissions directly to learning objectives and demonstrated mastery.

Consider a student writing an argumentative essay. If generating claims, selecting evidence, organizing an argument, and revising prose are the learning objectives, AI shouldn’t perform those functions. But if students have already demonstrated mastery of bibliography formatting, having AI assist with formatting citations might save time without sacrificing meaningful learning.

Willingham captures the distinction particularly well: for students, learning—not producing—is the thing.

LEADERSHIP ACTIONS

1. Replace “Can students use AI?” with “What thinking must students do?”
Make learning objectives the starting point for AI decisions.

2. Ask teachers to identify the cognitive work in major assignments.
For each task, determine which portions develop knowledge and skills and which are merely logistical.

3. Establish a “mastery before automation” principle.
Before students delegate a task to AI, they should demonstrate that they can perform that task independently and competently.

4. Build AI expectations into assignment design.
Instead of relying solely on a districtwide acceptable-use policy, teachers should explicitly state what AI assistance is and isn’t appropriate for individual assignments.

5. Preserve productive struggle.
Difficulty isn’t necessarily evidence that an assignment needs AI support. Sometimes difficulty is the mechanism through which learning occurs.

6. Reconsider assessment.
Schools increasingly need ways to determine what students can actually understand, explain, write, and solve independently—not merely what they can produce with AI assistance.

LEADER REFLECTION

• Do our current AI guidelines protect student thinking, or primarily regulate student behavior?
• Which classroom tasks should students master before AI assistance is permitted?
• How will teachers determine when a student knows how to do something “well”?
• Are we inadvertently allowing AI to eliminate productive struggle in the name of efficiency?
• Could every major assignment include a simple statement identifying what students must do themselves and why?

Perhaps the most useful AI policy for schools isn’t a complicated chart of permitted and prohibited applications. It may begin with one question:

“If AI does this part for the student, what opportunity to learn disappears?”

If the answer is something important, the student probably isn’t ready to hand that work to AI.

Read Daniel Willingham's original article⁠

Original Article——————————

Prepared with the assistance of AI software OpenAI. (2026). ChatGPT (5.2) [Large language model]. https://chat.openai.com