A student submits a beautifully structured response.
The introduction is clear. The explanation is fluent. The vocabulary is stronger than usual. The paragraphing is tidy, the conclusion works, and on the surface, the task appears complete.
Then you ask one question.
“Talk me through how you arrived at this.”
The answer is uncertain.
This is the moment many teachers are now worried about. Not because AI was used, but because the finished work may no longer show whether the student did the thinking the task was meant to develop.
AI can help students think more clearly. It can also help them move around the thinking altogether.
That is why this conversation needs to be about task design, not only tool use. If a task mainly rewards completion, AI will make completion easier. If a task requires students to reason, decide, explain, revise or create meaning, AI can become something students have to think with.
AI use becomes a learning concern when it does the part of the task where student thinking was supposed to happen.
To stop AI from replacing student thinking, teachers need to identify where the learning is meant to happen in the task and make sure that part still requires students to do the cognitive work.
This is a central part of Visible Agency. In an AI-rich classroom, teachers need to see more than a finished product. They need to see whether students still had to think. The question is not only whether AI was used. The better question is what thinking remained necessary because of the way the task was designed. Research on generative AI and agency suggests that GenAI may support learner agency through personalisation and assistance, while also raising concerns about learner autonomy, equitable access and changing notions of agency (Roe & Perkins, 2024).
For the broader framework, see Visible Agency: How to Design AI-Supported Learning Without Outsourcing Student Thinking.
AI has not created the problem of shallow learning. It has made shallow learning easier to see.
Many classroom tasks were already vulnerable before AI arrived. Some tasks rewarded neat completion more than careful thinking. Some asked students to produce an answer without showing how they arrived there. Some valued the final product more than the decisions, reasoning, questions or revisions behind it.
AI makes those weaknesses more visible because it can complete many product-focused tasks quickly. If the task asks for a summary, AI can summarise. If the task asks for a generic explanation, AI can explain. If the task asks for a polished paragraph, AI can produce one. If the task asks for a list of ideas, AI can generate it.
That does not mean those tasks have no value. It means they may need redesigning if the learning demand can be completed by the tool.
A task is fragile in an AI-rich environment when the required output can be produced without the student engaging in the thinking the task was meant to develop. The issue is not that AI can help. The issue is whether the task still requires the student to do the learning. Recent work on human agency in AI education identifies transparency, cognitive offloading and human oversight as central dilemmas for AI-supported learning (Viberg et al., 2026).
This shift asks teachers to look more carefully at task design. What is the task really asking students to do? Recall information? Apply a concept? Compare ideas? Interpret evidence? Make a judgement? Develop an argument? Revise for clarity? Create meaning?
Once that is clear, the next question becomes more practical:
Can AI do that part for the student?
If the answer is yes, the task needs a stronger design move. It may need a first attempt, a comparison, a justification, an evidence check, a revision explanation or a reflection on what changed. The goal is not to make the task harder for its own sake. The goal is to protect the part of the task where learning is meant to happen.
AI does not replace thinking in a well-designed task. It exposes where the task did not require enough thinking to begin with.
This is why AI belongs in a learning-design conversation, not only a policy conversation. Rules may tell students what is allowed. Task design determines whether students still have to think.
The central risk is completion without cognition.
Completion without cognition happens when students can produce the required output without engaging in the thinking the task was meant to develop.
A student may submit a completed response, but the completion itself may not show understanding. They may generate a paragraph without building the argument. They may produce a summary without deciding what matters. They may create a reflection without doing much reflecting. They may hand in a polished explanation without being able to explain the concept.
This is not new. Students have always found ways to complete tasks without fully engaging in the intended learning. AI simply makes the gap between completion and cognition larger, faster and less visible.
That is why final product quality is no longer enough evidence. A polished product might show strong learning. It might also show strong AI support. The teacher needs a way to see the cognitive work behind the product.
What did the student have to notice? What did they have to decide? What did they have to test? What did they have to explain? What did they have to revise? What did they have to understand well enough to own?
These questions matter because learning is not the same as output. The output is only valuable if it carries evidence of the thinking it was meant to develop. Research on assessment design in the age of generative AI argues that educators need ways to determine how vulnerable tasks are to AI completion, especially where tasks are intended to assess critical thinking (Zaphir et al., 2024).
A task that produces completion without cognition may look efficient, but it does not build capacity. It trains students to get the work done rather than become more capable through the work.
The design challenge is to make cognition necessary again.
Outsourcing thinking does not mean using support.
Students have always used support. They use teachers, peers, worked examples, sentence starters, success criteria, feedback, models, checklists, graphic organisers and now AI. Support is part of learning.
Outsourcing thinking happens when the meaningful cognitive demand of the task is transferred away from the student.
That may happen when AI generates the central idea, chooses the evidence, structures the argument, explains the reasoning, evaluates the quality, makes the revision decisions or produces the final judgement. The student may still be active in a surface-level way, but the learning demand has moved.
The danger is subtle because the student may still feel involved. They prompted. They copied. They edited. They formatted. They submitted. But if AI did the part of the task where the student was supposed to reason, interpret, evaluate or create meaning, the thinking has been outsourced.
This is the line teachers need to examine.
For example, if the learning purpose is to practise persuasive structure, AI can support students by generating examples to critique. But if AI writes the entire persuasive structure and the student only edits surface language, the learning demand may have shifted.
If the learning purpose is to understand a historical event, AI can support students by offering a starting summary to compare against sources. But if AI determines what mattered and the student does not verify, question or interpret, the task may no longer require historical thinking.
If the learning purpose is to develop a scientific explanation, AI can support students by identifying possible misconceptions in a draft. But if AI produces the explanation and the student cannot explain the causal relationship, the task has not protected the learning.
Outsourcing thinking is not about whether AI appears in the process. It is about whether AI has taken over the cognitive work the task was designed to develop.
AI assists learning when it supports the student’s thinking. It replaces thinking when it removes the need for that thinking.
This distinction is essential.
AI can assist when students use it to generate options they must compare. It can assist when students use it to identify gaps they must address. It can assist when students use it to receive feedback they must evaluate. It can assist when students use it to improve clarity after they have already formed the idea. It can assist when students use it to test their reasoning, challenge their assumptions or explore alternative explanations.
In each of these cases, the student still has to think. They still have to decide what is useful, what is accurate, what is relevant and what they are prepared to stand behind.
AI replaces thinking when it removes the need for the student to do the learning work. It replaces thinking when the student no longer has to form the idea, make the comparison, judge the evidence, explain the reasoning, revise the meaning or make the final decision.
The same tool can assist in one task and replace thinking in another. The difference is not the tool. The difference is the design of the task.
This is where the articles on judgement and ownership become important. Students need judgement so they can evaluate AI-supported ideas before using them. They need ownership so they can explain and stand behind the learning decisions behind the work.
For the judgement side of this work, see How to Design AI Tasks That Require Student Judgement.
For the ownership side, see How to Design AI-Rich Tasks That Still Require Student Ownership.
The practical question for teachers is simple:
What thinking must the student still do?
If the task answers that question clearly, AI can become a support. If the task does not answer that question, AI can quietly become a substitute.
A useful way to redesign AI-supported tasks is to begin with the learning demand.
Before deciding whether AI should be allowed, restricted or required, ask what thinking the task is meant to develop. The decision about AI should come after the learning demand is clear.
The Learning Demand Test asks five questions.
Every task has a point where the important learning is meant to occur.
It might be in forming a question, selecting evidence, explaining a process, analysing a text, solving a problem, comparing alternatives, revising a draft or making a judgement. Teachers need to name that point clearly.
If the learning is supposed to happen in the explanation, then students must still explain. If the learning is supposed to happen in the comparison, then students must still compare. If the learning is supposed to happen in the revision, then students must still revise with intention.
Without this clarity, AI decisions become vague. With this clarity, task design becomes sharper.
Once the learning demand is clear, teachers can identify the thinking that must remain with the student.
This does not mean students receive no support. It means the support should not remove the thinking the task is meant to develop.
If students are learning to interpret evidence, they might use AI to generate possible questions, but they still need to interpret the evidence themselves. If they are learning to improve an argument, they might use AI to identify weaknesses, but they still need to decide what changes to make and why. If they are learning to explain a concept, they might use AI to test clarity, but they still need to understand the explanation.
The point is to protect the thinking that matters.
This is the uncomfortable but useful question.
If AI can complete the central thinking demand, the task may need redesigning. A summary task may need a comparison layer. A writing task may need a reasoning explanation. A research task may need evidence verification. A design task may need a decision log. A reflection task may need a connection to personal experience or specific learning moments.
The goal is not to outsmart AI. The goal is to move the task towards learning that requires the student’s mind to be present.
A task becomes stronger when AI can support the process but cannot replace the student’s role in making meaning.
If the final product is no longer enough evidence, teachers need to decide what else they will look for.
Evidence of thinking might include a first attempt, annotated changes, comparison notes, an oral explanation, a revision rationale, evidence checks, a question trail, a decision log or a short statement about what changed in the student’s understanding.
The evidence does not need to be heavy. It simply needs to make the learning visible enough for feedback and accountability. The Education Endowment Foundation describes metacognition and self-regulation as approaches that help students think more explicitly about learning through planning, monitoring and evaluating, especially when embedded within curriculum and subject learning (Education Endowment Foundation, 2025).
This connects directly to How to Make Student Thinking Visible When AI Is Part of the Process [link URL]. If AI can polish the product, teachers need better evidence of the thinking behind it.
This is the most constructive question.
AI does not need to be excluded from learning for student thinking to remain strong. It needs to be placed carefully.
AI might be useful before the task, helping students generate questions or activate prior knowledge. It might be useful during the task, helping students identify gaps, compare alternatives or test clarity. It might be useful after a first attempt, helping students revise, refine or extend their thinking.
The placement matters.
If AI enters too early, it may set the direction before the student has thought. If AI enters at the right moment, it can give the student something valuable to examine. If AI enters after the student has made an initial attempt, it can support revision without replacing first thinking.
The task design should make that timing intentional.
Redesigning AI-supported tasks does not always require a complete rewrite. Often, it requires one stronger thinking move.
A task that asks students to produce an answer can become stronger when it asks them to explain how they arrived there. A task that asks students to generate ideas can become stronger when it asks them to compare and prioritise those ideas. A task that asks students to write a summary can become stronger when it asks them to verify the summary against a source. A task that asks students to revise a paragraph can become stronger when it asks them to explain which revision improved the meaning and why.
The product can still matter. It simply cannot be the only thing that matters.
Here are several practical redesign moves.
Ask students to make an initial attempt before using AI.
This might be a rough explanation, prediction, question, plan, sketch, argument, solution or list of what they already understand. The first attempt gives students a cognitive starting point. It also gives them something to compare later.
The point is not to catch students out. The point is to make their starting thinking visible.
Ask students to compare their thinking with AI support.
They might compare their explanation with an AI-generated explanation, their plan with AI’s suggested plan, or their draft with AI feedback. This turns AI into a surface for thinking rather than a replacement for it.
Comparison forces students to notice quality, difference and consequence.
Ask students to decide what is useful, accurate, relevant or worth changing.
This is where AI-supported work becomes more than tool use. Students have to evaluate the support they receive. They must decide what to accept, reject or revise.
This is the work developed more fully in How to Design AI Tasks That Require Student Judgement.
Ask students to verify important claims.
If AI provides information, students should check what matters. They may need to connect claims to sources, class materials, data, examples or criteria. This prevents fluency from being mistaken for truth.
Evidence checks are especially important in research, explanation, argument and inquiry tasks.
Ask students to explain what changed and why.
Revision is valuable only when students understand the improvement. If AI helps improve a draft, students should be able to identify which changes strengthened accuracy, clarity, structure, reasoning or audience connection.
A stronger final version should come with a stronger understanding of why it is stronger.
Ask students what remains theirs in the work.
This protects responsibility. Students should be able to explain what thinking they did, what support they used, what decisions they made and why the final work represents their learning.
This is the work developed more fully in How to Design AI-Rich Tasks That Still Require Student Ownership.
These redesign moves do not need to be used all at once. The best move depends on where the learning demand sits. If the learning is in the explanation, protect explanation. If the learning is in the evidence, protect evidence. If the learning is in the judgement, protect judgement. If the learning is in the revision, protect revision.
The aim is to keep the student intellectually present.
In AI-supported learning, teachers need to look beyond the final product because the final product may hide the process.
This does not mean every task needs extensive documentation. It means teachers need enough evidence to know whether the student did the thinking the task was meant to develop.
Look for signs of first thinking. Did the student attempt an idea before AI support arrived? Did they have a starting position, prediction, explanation or question?
Look for signs of comparison. Did the student notice differences between their thinking and AI support? Did they identify what improved, what was missing or what needed checking?
Look for signs of judgement. Did the student decide what to accept, reject or revise? Could they explain why?
Look for signs of evidence. Did the student check important claims? Did they connect ideas to sources, examples, data or criteria?
Look for signs of revision. Did the student make meaningful changes? Could they explain how those changes improved the work?
Look for signs of ownership. Could the student explain what remained theirs? Could they stand behind the final work?
These signs do not have to appear in a long written reflection. They can appear in a brief note, annotation, conference, draft comparison, oral explanation or exit response.
The important thing is that the task gives the teacher something to see.
When teachers can see the thinking, they can give better feedback. When students can see their own thinking, they can become more intentional learners. That is the heart of Visible Agency.
For school leaders, this article points to a larger issue.
Schools cannot solve AI’s impact on learning through detection alone. Detection may answer whether AI was used, but it does not answer whether students learned. It does not show where the cognitive demand sat inside the task. It does not tell teachers whether AI assisted the learning or replaced it.
Educational AI evaluation also needs to move beyond accuracy and task efficiency to include learner agency, context, ethics and human-centred outcomes (Ding & Magerko, 2025). That is why task design matters. Schools need a way to examine not only whether AI tools perform well, but whether learning tasks still protect the thinking they are meant to develop.
The stronger leadership question is:
How do we design tasks where students still have to think?
That question moves the conversation from control to design. It helps teachers examine learning demand, task structure, evidence of thinking and student responsibility. It also gives teams a shared language for improving AI-supported learning without turning every conversation into compliance.
AI makes this work more urgent, but the work itself is not new. Strong learning has always required more than task completion. Students need to reason, question, connect, evaluate, revise and explain. AI simply makes it easier to see when those demands are missing.
AI can support learning when it gives students something better to think with. It weakens learning when it removes the need to think.
That is the line teachers need to design around.
The task is not simply to decide whether AI is allowed. The task is to decide where the learning is supposed to happen and how to keep the student present at that point.
If AI helps students compare, question, verify, revise, explain or create meaning, it can support agency. If AI quietly performs the cognitive work the task was meant to develop, the task needs redesigning.
The goal is not to make learning immune to AI. The goal is to make learning strong enough that AI cannot replace the part that matters most.
This article is part of the Visible Agency series.
For the broader framework, read Visible Agency: How to Design AI-Supported Learning Without Outsourcing Student Thinking
For the evidence side of the work, read How to Make Student Thinking Visible When AI Is Part of the Process.
For the judgement side of the work, read How to Design AI Tasks That Require Student Judgement
For the ownership side of the work, read How to Design AI-Rich Tasks That Still Require Student Ownership