Fake Education Might Be the Best Teacher
The entire education system is flying blind.
School leaders can’t test whether a scheduling reform will improve learning without disrupting actual students’ days. New teachers enter classrooms being told what to expect but without the experience of handling 25 real learners. Many wash out early because of the intense stress of learning the hard way. Researchers studying how to develop durable skills work largely from theory, not evidence, because longitudinal studies take years while student populations and contexts shift. Instead, they’re forced toward less important experiments.
When someone proposes a radical redesign, of course people reject it as too risky. It’s not “evidence-based.” The choice becomes implement it and hope for the best, or reject it because we can’t prove it works. No wonder there’s little reform. Hope is a poor substitute for evidence.
Education desperately needs what other complex fields have—a way to safely explore “what if” scenarios at every level of the system. We need simulations.
Learning from Other Fields
Other fields have begun exploring similar approaches. Marketing departments are experimenting with AI-generated personas to test campaigns. Economics has a longer history with simulations, using macro-level models to test policy interventions. Urban planners use data-driven simulations to test traffic interventions and zoning changes. Where simulations have taken hold, they’ve proven valuable for testing ideas before committing resources.
Education has dabbled in simulations for decades. SimSchool, developed in the early 2000s, created virtual classrooms where teacher candidates could practice instructional strategies without harming real children. Give students work that’s too challenging and virtual students displayed frustration. Make it too easy and they coasted. Research showed pre-service teachers using the simulator gained instructional self-efficacy more rapidly than those in traditional preparation programs.
These early simulations proved valuable but had fundamental limitations. The simulated students operated on simple rule sets. They could demonstrate frustration or confusion but lacked the nuanced responses of real children. Some programs used virtual students puppeteered by trained actors. Effective, but labor-intensive and impossible to scale.
Human behavior is extraordinarily complex. Earlier systems couldn’t capture that complexity reliably.
Today’s AI engages in fluid dialogue that closely mimics human conversation. An AI student can respond with contextually appropriate answers, make realistic mistakes, ask spontaneous questions, or exhibit confusion. Simulated humans can now be much more complex, nuanced, and multi-faceted.
This year, researchers created an AI student avatar named “Cecilia” to help teachers practice eliciting student thinking. The GenAI version prompted Cecilia to act as a virtual first-grader with specific knowledge gaps. Cecelia solved problems in childlike ways. Teacher participants found the experience realistic and valuable.
This is a single-agent simulation. One AI student helping an individual teacher practice specific skills. But the same technology enables multi-agent simulations—entire virtual classrooms populated with diverse AI students, each representing different backgrounds and learning profiles. Imagine a virtual class with an AI student who struggles with reading, another who excels in math but is shy, another displaying ADHD tendencies. Run a proposed teaching approach through this population and observe how each responds. Or hand the simulated class to a teaching student on day one as their “class” for the term.
(I should note these examples aren’t endorsements. I had AI do a deep research search for recent educational AI simulations and it didn’t come up with much, but of course it may have missed a lot. I’m not sure how unique Cecelia is.)
The Knowledge-Building Loop
The obvious question is whether we can describe human behavior well enough to reliably simulate various people in or related to education systems—teachers, students, leaders, payers, parents or guardians.
The honest answer is not perfectly, but well enough to start. And the process of trying has many huge benefits.
AI can generate plausible variations of students quickly. Plausible is sufficient for many uses. Teacher training doesn’t require perfect fidelity. It needs realistic enough interactions that teachers develop useful pattern recognition and response strategies. System-level explorations (will this scheduling change likely improve or degrade learning?) as variations of current paradigms should work with AI that’s infused with community knowledge and experiences.
It might be possible for teachers to simulate their exact class if AI had access to learning management systems and student data. Ethically, this raises concerns around privacy and consent, or inappropriate use of such emulations. Creating a simulation of an identifiable student without consent ventures into troubling territory, even with pedagogical intent. Better to use composite student types drawn from aggregated patterns rather than individual digital twins.
Other needs require tighter validation. Before making high-stakes policy decisions, you’d want to validate the model against documented prior interventions. Does your simulation correctly predict what happened when that disciplinary policy changed five years ago? The mismatches tell you where your model needs refinement.
If these simulations are controlled by the education research and teaching college community rather than locked in proprietary systems, they become a forcing mechanism for discussion about what really matters and what to work on next.
You can’t build a simulation by putting off decisions like we do in real life, or by discounting factors simply because you’re uncertain. With a simulation, you have to make the choices. What factors drive student engagement? How do teacher actions affect classroom dynamics? What makes some students more resilient to setbacks? In my experience observing game-based learning design, this forces choices that are often avoided in regular practice. The construction process drags tacit knowledge into the open, surfaces hidden assumptions, and forces clarity about what actually drives outcomes. Often simulations reveal something surprising or a “watch out for,” not necessarily accurate predictions.
The community building the simulation must articulate their assumptions, and those assumptions get tested every time someone runs it. When predictions fail, a big red arrow points at the next important field experiment.
This could help coalesce a field that’s frustratingly fragmented. Education research struggles to conduct scaled experiments. Replication studies are rare. Different sub-communities work in isolation. A shared simulation platform creates a common reference point. SimCity wasn’t community built and controlled, and far more accurate “digital city” models exist in the research world. But it had a powerful focusing effect on urban planning. Researchers debated its accuracy. It became a common teaching tool. And the populace learned about urban planning and could even demonstrate innovative approaches, effectively crowdsourcing solutions. The various city models became a centroid for research.
Watching virtual students respond to interventions sparks concrete discussions about practice rather than abstract debates about theory. Teachers, researchers, and administrators explore the same scenarios and compare interpretations. Over time, the simulation improves as accumulated knowledge gets encoded into better models.
Why Extrapolation Fails for Transformation
Most education research has studied traditional paradigms using variants of traditional methods. We know something about how students learn algebra through direct instruction in 45-minute periods. We know less about learning through extended project work with peer collaboration. And for an education model that’s extremely different from the norm, the accumulated evidence is see-through.
Change the paradigm substantially, and everyone involved will behave differently. The teacher who excels at lecture-based instruction might struggle with facilitating project-based learning. Students conditioned to wait for teacher direction need time to develop agency in student-centered environments.
The point is, radical changes trigger unanticipated reactions throughout the system. Switching to project-based learning affects pedagogy, scheduling, assessment, parent communication, and professional development. Some teachers embrace it, others resist. Some students thrive, others flounder.
You can’t extrapolate from “teaching the way we’ve taught it” to understand how a fundamentally different paradigm might work. You need knowledge at more fundamental levels—basic patterns of human motivation, learning, and behavior that hold across different contexts.
Simulations can surface these challenges before implementation. Testing block scheduling in a virtual school might reveal that students prone to attention difficulties struggle in longer class periods unless teaching strategies adjust accordingly. Noticing this in advance allows adding supports or teacher training to address the issue before rolling out the change. A simulation showing that a new grading system confuses parents in certain demographics prompts preemptive communication strategies.
The simulation won’t predict everything. Human systems are too complex. But it dramatically expands the range of scenarios we can explore and the speed at which we can explore them. It turns transformation from a terrifying leap into informed experimentation.
Time to Build the Laboratory
Games and simulations have always been valuable educational tools when done well, compressing time and providing safe spaces for experimentation. But they were brittle. Scripted scenarios couldn’t adapt. Simple AI produced shallow interactions. Building sophisticated simulations required enormous resources. The models missed crucial nuances.
AI’s ability to role-play changes everything. Creating a sophisticated AI student now requires careful prompting rather than months of coding. Scaling is straightforward.
Teacher preparation programs can give every candidate extensive practice with diverse virtual students. School leaders can test policy changes before committing resources. Researchers can run experiments that would be impractical or unethical in real schools.
The sophistication of behavior isn’t the limiting factor anymore. The limitation is our ability to articulate what we know about teaching and learning in forms that simulations can use.
The education system has been flying blind for too long. AI agent simulations are a new, powerful corrective lens.
©2025 Dasey Consulting LLC
This is brilliant. Airlines invest X in sophisticated flight simulators to train pilots to save hundreds of lives with real time practice. Education needs comparable investment.
Me to GPT 5.1: Give me a number for X or rewrite this statement accordingly.
GPT 5.1: Here are two (sic) strong, clean versions you can use. Each gives a concrete number and maintains the force of your point without exaggeration.
Option 1: Use a specific dollar figure (most defensible).
Airlines invest about 6 billion dollars in professional-grade flight simulators so pilots can rehearse life-and-death decisions in real time. Education needs an equivalent investment in simulation-based practice for teachers and students.
Option 2: Use a per-device estimate (even more tangible).
Airlines routinely spend 10 to 15 million dollars on a single full-flight simulator because the ability to practice complex decisions safely is worth it. Education deserves training systems built with the same seriousness.
Option 3: Keep it general but powerful.
Airlines pour billions into flight-simulation technology because practice under realistic conditions saves lives. Schools need the same mindset about investing in high-fidelity learning environments.
If you want, I can tailor these lines for LinkedIn, for a district memo, or for a speech.
Interesting idea! As a candidate for a job as principal, I recently had a performance task with an AI simulation of an angry parent. It was more realistic than I thought it would be, and I can see the potential benefit in teacher (and school leader!) training. However, given that I have lots of experience having meetings with angry parents, I did notice a few things-- my levels of adrenaline and cortisol were considerably lower when interacting with a screen than when getting the biofeedback of a human being with all their tone and body language-- so I was able to respond with much more calm thoughtfulness than I could in a real life scenario where I would be spending some of my cognitive resources just managing my adrenaline response. But yes, overall, promising idea!