Sleight of Mind 2

Ten Moves to Keep AI From Stealing Productive Struggle (Part 2 of 2)

Colleen Ferguson · Executive Director · 19 August 2026

This is the second of a two-part series that offers some concrete strategies to design for learning in the age of AI. Part 1 argued that the right classroom culture strips away AI's cover to perform a sleight of mind. Get the trust, transparency, and visible process right, and AI can't steal the learning.

The first five moves were:

1. Value Process over Product
2. Prioritise Live Learning
3. Provide Clarity about When and How to Use AI
4. Eliminate Gotcha Culture around AI Use
5. Expect and Model AI Transparency

For more on these strategies catch up on Part 1 of Sleight of Mind.

These next five moves turn from the climate around the task, to the task itself. These are moves that protect the “Aha!” moment so that the learning can’t be snatched whether AI is in the room or not.

AI is built to make everything easy, but learning requires friction. Our minds need to be challenged to think in new ways in order to carry a skill or concept from one context into another. Let AI be productive in our place, and we can admire the result but we won’t have been changed by the experience. We won’t have learned. And the research keeps telling us that this kind of production doesn’t stick.

So how do we ensure students will have some opportunity to wrestle with a challenge whether AI is in the room or not? Here are five more moves to try this year.

Move 6: Reflect Before, During and After Learning

Reflection has a bad reputation among students because it’s usually bolted onto the end of an assignment. This reduces it to an irrelevant add-on, not the main event, and something to be rushed through in order to complete the assignment. Students write the paragraph you asked for, they tell you what they think you want to hear, most of it is descriptive (I did this and then I did that) and all of it is irrelevant to the current assignment because by then the work is already over.

Weave reflection through the process instead, and you open the door to change while there’s still something left to change. Three reflective opportunities to build in: a statement of intention before the work begins (What do I want to achieve with this? What am I curious about? What am I excited to create?), a halfway check-in (What’s working? What am I unsure of? How do I move this to the next level?), and a final reflection on how the student can transfer what they’ve learned to whatever comes next. In the age of AI, this is no longer a nice-to-have. It’s what keeps the student invested enough in the work to protect it, not just produce it

A new study by Ates found that students who evaluated their own writing first by naming what was working well and what still needed development, retained the ideas and arguments weeks later. However, students who used the draft stage as a brain dump and then skipped straight to AI feedback, improved faster in the moment, but that improvement evaporated the moment the AI wasn’t there to lean on.

The key takeaway is that not all AI feedback is equal. Students need to build out their ideas, wrestle with the initial creative mess, and reflect deeply on their own progress before AI gets a say.

Metacognition helps students frame what they want from an assignment, gauge how far they have gotten in achieving that ambition, and at the end, carry what they’ve learned to the next challenge.

Move 7: Put Students in the AI Driver’s Seat

The problem with how many students use AI is that the technology is built to do all the work and users slide into being passive accepters of whatever it hands them. That’s the road to offloading and deskilling. But teachers can train students to put the friction back in by building struggle into the task itself.

Building in an opportunity to struggle is a part of the Pedagogy of Productive Struggle, a framework developed by Lincoln Smith, Damion Walker and myself. The struggle verbs below are based on Bloom’s Taxonomy, adapted for AI interactions and designed to keep student agency. They can’t be used until a student has produced something to start with and the interaction that follows should be visible, shared with the teacher, and reflected on: did the AI help, or did it get in the way?

Once a student has built an initial draft of a project and they’ve reflected on what they’ve done well and what they need to improve, provide a menu of options for how they can engage with AI to iterate their work. Students should choose the provocation based on the next steps identified in their reflection.

Justify | Explain, defend, or critique the process, investigation, or deduction. “Ask the AI to disagree with your idea. Then explain why you still think you’re right or how the AI helped you change your mind.”

Modify | Revise, rethink, refine. “Ask the AI to identify weaknesses in your draft. Revise your work based on one of the suggestions, then explain what has improved.”

Classify | Evaluate, judge, critique, and prioritise. “Ask the AI to point out the strongest and weakest parts of your three ideas. Decide whether you agree, and explain which idea you think is actually best, and why.”

Unify | Find connections between disparate things. “Give the AI two contrasting ideas or points of view. Prompt it to propose three ways they might relate. Then build on its suggestions to create a fourth, original link.”

Complexify | Move beyond initial thinking; explore tension and ambiguity. “Ask the AI to find a tricky question or problem hidden inside your idea or argument. Use that challenge to make your thinking more interesting or detailed.”

Amplify | Imagine how to scale the idea up in a real-world context. “Ask the AI how your idea could become bigger or reach more people. Pick one suggestion and sketch what that scaled-up version could look like.”

The verb does the work. In each case the AI is still in the room, but the student is the one steering, and the thinking that matters stays firmly on their side of the desk.

Move 8: Condition Productive Failure

We know that student-driven discovery fuels learning. It’s the whole rationale behind inquiry-based approaches. We also know that when students reach for answers before they’ve built mastery, they often arrive at the wrong ones, which is why we lean on constructivist strategies to help them build meaning. But maybe what we need to do more of is design learning tasks so complex and so challenging that they will inevitably produce some wrong answers, so that being wrong is the point.

This is hard for us, because the Zone of Proximal Development tells us to provide enough support that a student can bridge the gap between what she can do alone and what she can do with help. The support part is our jam. We’re wired to engineer success, not failure.

In 2008, Manu Kapur argued that engaging students in complex, ill-structured problems without support could result in something he called productive failure. Struggling primes the mind to recognise why the correct method works once it’s finally taught. It isn’t an unfortunate side effect to be minimised. Struggle is the mechanism that makes the instruction that follows resonate.

Productive failure only works, though, if the student can tolerate the failure part long enough to reach the productive part, and that’s a Growth Mindset challenge. The two ideas need each other. So alongside our instinct to engineer success, we need to build in low-stakes room to fail and treat that failure as a milestone, not a mistake.

Move 9: Red Team Your Assessments

At the start of Part 1 of this two part post, I asked a question: could AI do the thinking that leads to the Aha? This move turns that one-off diagnostic into a habit.

Before a new tech product or model is rolled out to the public, it is often “red-teamed”. Red teaming is when an organization deliberately has a person or group try to attack, break, or exploit a system in order to find its vulnerabilities before real adversaries do.

Before you assign anything, instruct the AI to genuinely try to break your own assessment. Feed it your exact instructions. Hand it your rubric. Ask it for the fastest possible route to full marks. Don’t do the polite version of this; do the version a stressed, capable, time-poor student would do at eleven at night. If the AI hands you a credible product that skipped the struggle entirely, you’ve found the flaw before your students did, which is precisely when you want to find it.

Then fix it, and you already have the tools. Ground part of the task in something that happened in the classroom, so the AI can’t reach it (Move 2). Integrate an AI log so that its use is transparent (Move 5). Shift the weight of the grade onto the process steps rather than the polished product (Move 1). Red teaming is the quality-control pass that tells you which of the other moves this particular task actually needs. It doesn’t take long and it’s the difference between hoping a task is AI-resistant and knowing it.

Move 10: Create One Blue Sky Assessment a Year

Once a year, give your students a wicked problem: a real-world challenge, aligned with the concepts and skills you’ve built all year, that’s genuinely contested and unsolved: a problem where reasonable people disagree and there’s no single right answer waiting to be found.

Examples could include:

The cafeteria challenge: How do we reduce waste and cost while optimizing for taste and nutrition? Who are the actual clients of a cafeteria, and what is it for?

The phone policy: Should phones be banned, tiered, or contract-based by class? How do we balance focus and safety against the fact that a ban just starts a game of cat and mouse?

The AI-era school: Can AI teach? What does a modern school need to preserve, and what should it change?

Literacy, reading, and numeracy: Reading and math scores are in decline. What, if anything, should be done about it, and what value do these skills hold now and in an AI future?

There’s a second way to blue sky an assessment, and it doesn’t need a contested problem at all. Take a task you already run and blow the scope wide open. Instead of a short story, a book. Instead of a history essay, a series of podcasts. Instead of a lab report, a public exhibit. Instead of a persuasive essay, a real campaign aimed at a real audience.

The content and skills stay the same. What changes is scale, and scale is its own kind of friction. AI can write a competent five-paragraph essay in one pass; it cannot sustain a coherent book, or a dozen well-researched podcast episodes in a consistent voice, without a student doing the sustained planning and judgment-calling that holds a bigger project together.

Either way, the principle is the same: AI cannot steal a struggle it can’t shortcut in one pass. It can summarize, suggest, draft but it can’t hand a student the answer to a question that doesn’t have one, or hold sixty pages of judgment calls in its head on the student’s behalf. That work stays with the student, because there’s nowhere else for it to go.

A blue sky problem is a challenge to run and assess, and it only works once students have the foundational knowledge to engage with it seriously. It’s the reward at the end of a year of narrower, scaffolded work, not a replacement for it. But the effort is worth it. These types of assessments: ambitious, amorphous, student-driven and inquiry-driven are how students find meaning in school. How they can see the result of their learning efforts.

Eric Hudson put it well in a recent Learning on Purpose post: schools might do better modeling themselves on gyms rather than on workplaces because a gym is built on the belief that the unglamorous, repeated effort is the point, and the reward lives in the future. In the workplace, by contrast, no one much cares how you arrive at the goal as long as you deliver: promotions, raises, recognition, are all tied to hitting that mark. Swap “goals” for “grades” and you’ve got a pretty accurate description of the transactional school model that incentivizes unethical AI use.

As much as I agree with the analogy, it’s important to note that workplaces are great learning laboratories as well. Goals are set with no recipe for achieving them. Individuals and teams have to figure out how to use whatever resources, information, and tools they have. There’s real human effort in getting there, but also smart, resourceful use of people and technology.

The beauty of a blue sky assessment is that it is both gym and workplace at once: the one day a year you take the training wheels off and let them ride. It’s analogous to the piano recital, the playoff game, the product launch and the pitch that finally lands. It’s the moment that shows a student what all the earlier, narrower work was actually for.

Students have been finding ways around the hard part of learning for as long as there’s been a hard part to get around. AI just made the sleight of hand faster, smoother, and far easier to miss. This means we must design tasks worth struggling through, create conditions of trust and transparency, and be clear about our goals and methods. Then AI can’t snatch the struggle.

That’s the work ThinkSlant exists to do, and I’d love for you to do some of it with us. Slantwise isn’t going anywhere — I’ll keep thinking out loud here every week. But you’ll also find these essays cross-posted on the ThinkSlant blog, alongside classroom-ready tools, a lesson exchange, and a practitioner journal built by teachers, for teachers.

Sign up for the newsletter at thinkslant.org. And let’s make sure the struggle stays exactly where it belongs — with the students doing the learning.

© 2026 Colleen Ferguson. Republished from Slantwise.