Principle 01
[26:09]Use dynamic, non-obvious task structures
The shape of the task is the biggest design lever. Some structures invite offloading; others block it intrinsically.
A working reference guide · Capsim · April 2026
When Student Work Looks Right, but No One Is Thinking
Dr. Erich Dierdorff's framework on AI, critical thinking, and learning design — indexed and deep-linked to the recording.
This page is the framework from the webinar, in scannable form, with every element clickable to the moment in the recording where Erich explains it. Use the nav to jump to what you need: the user typology matrix, the four principles, a 12-question audit you can run against an assignment, where simulations fit, or the recording and references.
§1 The mechanism
What's actually happening when AI shows up in a course
AI use isn't all-or-nothing. Faculty alarm is real: a College Board survey of 3,000+ faculty found more than 95% concerned that AI is damaging how higher ed delivers learning [3:32]. But the headline binary ("ban or embrace") obscures what's actually happening. Use varies along two dimensions: frequency and agency [8:22]. That produces four very different students. The naive user, not the power user, is where cognitive offloading does the damage [10:57].
Low freq · High agency
Precise user
Uses AI selectively, for tasks they understand well enough to evaluate the output.
High freq · High agency
Power user
Prompt-engineers, iterates, scrutinizes. Knows what AI is and isn't good at.
Low freq · Low agency
Non-user
Takes whatever AI happens to give them and treats it as ground truth.
High freq · Low agency
Naive user
Hammers AI with single prompts and accepts whatever comes back.
The one to worry about
Why it matters: friction is essential to learning
Generative AI is built, deliberately and at the UX level, to remove friction. Friction is what learning requires. The same property that makes AI useful makes it corrosive to learning: students stop doing the cognitive work the assignment was designed to develop. That mechanism has a name. Cognitive offloading [19:38].
Friction is absolutely essential to learning. We have over 70 years of learning science that shows you have to create friction for deeper-level learning to happen.
Dr. Erich Dierdorff [16:11]
§2 A framework
Four principles for AI-resistant learning
Critical thinking is the load-bearing skill cognitive offloading threatens — and it's what the four principles below protect. Erich operationalizes it as three skills working in sequence: discern, diagnose, deploy. Cognitive offloading attacks all three [21:34].
Step 1 · The first D
Discern [21:42]
Comprehend the information in front of you.
AI reads it for the student. The comprehension habit never gets built.
Step 2 · The second D
Diagnose [21:48]
Evaluate the quality of that information.
AI evaluates for the student. The critical filter never develops.
Step 3 · The third D
Deploy [21:54]
Act on it appropriately given that quality.
AI acts for the student. Judgment under stakes never gets practiced.
Tactics travel poorly. Principles travel — they hold across institution, course, and the particular AI tool of the moment. The specific assignments that implement them are the practitioner's call.
Principle 02
[28:05]Create friction in the learning process
Where task structure alone doesn't block offloading, design moves put friction back in.
- Document and reflect on prompts — the reflection is the lever [28:23]
- Increase distrust of AI output — trust correlates with offloading [28:54]
- Deliberately induce errors — error-management training has decades of evidence [29:50]
- Throw curveballs — change specs mid-task; AI handles a static spec better than a moving one [30:24]
Principle 03
[30:58]Facilitate critical thinking directly
The 3Ds are skills. Like any skill, they develop through repetition under feedback.
Principle 04
[33:49]Expose the value of what students are learning
The least-discussed principle, and the most consequential when it lands. Students offload when they quietly believe the work doesn't matter.
We have to really move beyond this idea of best practices. The problem with best practices is you just end up copying a tactic and not really understanding what the underlying principle is. Which is what we call best principles.
Dr. Erich Dierdorff [23:40]
§3 Apply it
Audit a course or program
Pick what you can speak to — an assignment, a course, a program, or your curriculum as a whole. The questions ask about the learning design itself, so they work at any of those levels. Answer twelve yes/no questions for the unit you picked. The lowest-scoring principle is your highest-leverage redesign target.
If you can't confidently say yes, leave the box unchecked. Uncertainty is itself a signal — it usually means the design doesn't currently require that move.
§4 Where simulations fit
One modality already satisfies most of these principles
Case studies generally hit complexity and multiple viewpoints but miss equifinality, dynamism, and error inducement unless multi-staged. Simulations, on the other hand, hit the full set by design [37:39].
| Principle-1 task characteristic | Case studies | Simulations |
|---|---|---|
| Equifinality multiple valid paths to success | ||
| Complex / open-ended | ||
| Ill-structured no clean answer key | ||
| Multiple viewpoints | ||
| Dynamism the situation changes over time | ||
| Error inducement participants live with consequences |
Simulations pretty much touch every single task characteristic we've talked about. It's really hard to cognitively offload a lot of this work. There's a lot of ways to compete in simulations. There's a lot of paths to success. By design open-ended and complex. There's changes in dynamism over time. Usually it takes multiple viewpoints. Simulations create errors — we know people make bad choices and you have to live with those choices.
Dr. Erich Dierdorff [38:54]
The reason the mapping is so clean isn't accidental. A business simulation built well will force the conditions Principle 1 describes:
- Equifinality — multiple strategies can win; copying a competitor's playbook usually loses.
- Open-endedness and ill-structuredness — the round doesn't have an answer key; the market does.
- Multiple viewpoints — functional decisions integrate across marketing, finance, operations; peer teams compete in real time.
- Dynamism — the situation changes round to round in response to participant choices.
- Error inducement — participants make consequential decisions and have to live with the results into the next round.
None of those features were designed for the AI era. They predate it by decades. They happen to be exactly the features the AI era requires.
What about AI inside the simulation itself?
If a student can consult an AI tutor during a simulation, doesn't that reintroduce offloading? It depends entirely on what the tutor is designed to do. A tutor that produces answers undermines the experience. A tutor that asks the right question protects it. Capsim's built-in tutor does the second one explicitly [40:20] — Dierdorff describes it:
It tells them things that you would say as a faculty member: okay, so what is your strategy? Do you understand what your competitors are trying to do? Like all the things that you would do anyway if they came to your office hours.
Dr. Erich Dierdorff, on Capsim's built-in AI tutor [49:08]
That's the design intent: office hours, programmatically embedded. The work — discern, diagnose, deploy — stays with the student. Dierdorff himself stops short of "AI-proof" — "I hesitate to say AI-proof because maybe nothing is 100% AI-proof, but it makes it much more difficult to be a naive user" [40:03]. The calibration is worth keeping: naive-user-resistant is the honest claim, and the design target every higher-ed learning experience now has to aim at.
§5 Resources & references
Watch, share, cite
The webinar recording
The full 50-minute talk with Dr. Erich Dierdorff. Hosted by Capsim, recorded April 2026. All timestamps on this page deep-link into this video.
Share this page
If a colleague would benefit from the framework or the audit, send them the page directly.
Cite this page
For syllabi, faculty meetings, or anywhere it's useful to reference the framework directly.
- Capsim. When Student Work Looks Right, but No One Is Thinking: A Working Reference on AI & Learning Design. 2026.
- capsim.com/vault/ai-critical-thinking
References
The webinar this page distills, plus the scholarship behind the framework.
- Abdulnour, R. E. E., Gin, B., & Boscardin, C. K. (2025). Educational strategies for clinical supervision of artificial intelligence use. New England Journal of Medicine, 393(8), 786–797. https://doi.org/10.1056/NEJMra2503232
- College Board. (2025, February). New College Board research: Faculty express near-universal concern that student AI use undermines original writing and critical thinking. newsroom.collegeboard.org
- Dierdorff, E. C. (2026, April 21). When student work looks ‘right’… but no one is thinking [Webinar]. Capsim Management Simulations. youtube.com/watch?v=uUB_8QJWzKw
- Gallup. (2026). Voices of Gen Z: The AI paradox. gallup.com/analytics/651674/gen-z-research.aspx
- Gerlich, M. (2025). AI tools in society: Impacts on cognitive offloading and the future of critical thinking. Societies, 15(9), 252. https://doi.org/10.3390/soc15010006
- Holmes, N. G., Wieman, C. E., & Bonn, D. A. (2015). Teaching critical thinking. Proceedings of the National Academy of Sciences, 112(36), 11199–11204.
- Lan, M., & Zhou, X. (2025). A qualitative systematic review on AI empowered self-regulated learning in higher education. npj Science of Learning, 10(1), 21. https://doi.org/10.1038/s41539-025-00319-0
- Li, F., Yan, X., Su, H., Shen, R., & Mao, G. (2025). An assessment of human–AI interaction capability in the generative AI era: The influence of critical thinking. Journal of Intelligence, 13(6), 62. https://doi.org/10.3390/jintelligence13060062
- Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002