AI Student Teacher Training: Meet DAImian, the Avatar

Your child’s future math teacher may be learning the job by teaching a simulated child. On September 22, 2026, education majors at the University of Wisconsin–Whitewater were reported practising math instruction in VR headsets with DAImian, an AI-powered avatar playing a struggling sixth-grader, in place of the peer role-play that teacher-preparation programmes have used for generations. AI student teacher training of this kind fixes a genuine weakness in how teachers learn to teach. It also introduces a question nobody has answered: whose struggling sixth-grader is the model simulating?

Key Takeaways

• The practice replaces peer role-play, where education majors pretend to be students for each other. Peers are adults who already know the content and cooperate, which makes the rehearsal unrealistic in the exact way that matters.
• The real skill being practised is diagnosing a misconception and responding to it, which requires a partner who holds the wrong idea consistently. An avatar can be instructed to do that. A classmate cannot.
• The open question is representativeness. A model asked to play “a struggling sixth-grader” produces a statistical composite of text about struggling students, and nobody has published what that composite contains.
• VR sessions generate data about the teacher candidate: gaze, movement, speech. At a postsecondary institution, FERPA rights over those records belong to the student, not the student’s parents.
• There is no evidence yet linking avatar practice to better outcomes for real children. Judge it as a plausible training improvement, not a proven one.

What is actually happening at UW-Whitewater

A simulated classroom is a rehearsal environment in which a teacher candidate practises instruction on a scripted or modelled student rather than on real children.

The AI-in-education tracker entry dated September 22, 2026 describes education majors using VR headsets to teach DAImian, an AI-powered avatar playing a struggling sixth-grader, in place of peer role-play.

Simulated-classroom practice is not new to teacher preparation. Programmes have used scripted actors, video analysis and role-play for decades, precisely because putting a novice in front of thirty real children to learn basic moves is bad for the novice and worse for the children. What changed is the layer underneath. An actor improvises. A model can be configured, repeated and varied, and it does not get tired of being wrong in the same way forty times.

That last property is the whole argument.

Why peer role-play was always the weak link

Here is the mechanics of the problem, because the fix only makes sense once you see it.

Teaching a child who is stuck is not primarily about explaining better. It is about figuring out what specific wrong idea the child is holding, then choosing a move that confronts that idea. A sixth grader who says 1/3 is smaller than 1/4 “because 3 is smaller than 4” is not confused about fractions in general. She is applying whole-number reasoning to denominators. The teacher’s job is to notice that in about four seconds and pick a representation that breaks it.

Peer role-play cannot rehearse this. When a classmate plays the struggling student, she knows the right answer and has to invent a plausible wrong one on the fly. She will drift. She will accept the first explanation out of politeness. She will not hold a misconception under pressure, because holding a wrong belief consistently while someone argues with you is genuinely hard to fake.

An AI avatar can be instructed to hold exactly one misconception and defend it until the candidate addresses the underlying idea. That is rehearsal with the difficulty left in. UCLA’s Bjork Learning and Forgetting Lab describes this family of conditions as desirable difficulties, “difficult but effective training conditions” that impede immediate performance while producing better long-term retention. A rehearsal partner who is politely cooperative removes the difficulty and the learning with it.

Practice setting	What it rehearses well	What it cannot do	Cost
Peer role-play	Lesson sequence, timing, nerves	Hold a misconception consistently; surprise the candidate	Near zero
AI avatar (DAImian)	Diagnosing and confronting a specific wrong idea, repeatedly	Represent a real child’s history, home context or relationship	Headsets, licences, faculty time
Scripted human actor	Realistic resistance and emotion	Repeat identically; scale to every candidate	High per hour
Real classroom placement	Everything, badly, at the child’s expense	Be safely repeatable or focused on one skill	Borne by real students

The representativeness problem nobody has measured

This is the part worth your attention, and it is a technical point rather than a political one.

When a language model is asked to play a struggling sixth-grader, it does not consult a child. It generates text that is statistically typical of text about struggling sixth-graders in its training data. That corpus includes teacher forums, intervention manuals, case studies and a great deal of writing in which struggling students are described by adults. Whatever patterns exist in how adults write about struggling children, including patterns tied to race, class, language background and gender, are available to the model as material.

So the question is not whether the avatar is biased. It is which composite it produces and whether candidates notice. If DAImian’s frustration, vocabulary and persistence are drawn from a particular slice of writing about particular children, a candidate who practises forty hours on that composite is calibrating her instincts to it. Then she walks into a real classroom.

This is not an argument against the tool. It is an argument for one specific disclosure: ask teacher-preparation programmes what the avatar was configured to represent and whether candidates practise on more than one profile. A programme that can answer has thought about it. A programme that says “it’s just the AI” has not.

The underlying lesson matches what the experimental literature keeps finding about model configuration. Bastani and colleagues reported in PNAS in 2025 that the same model helped or hurt high school math students depending on whether it was configured to give answers or teacher-designed hints. Configuration is not a detail. It is the intervention.

Why this matters for the teacher your child gets

Teacher quality is the lever with the strongest evidence behind it in this whole area, which is why training changes deserve more attention than software purchases.

Nickow, Oreopoulos and Quan’s 2020 NBER meta-analysis of experimental tutoring studies found a pooled effect of 0.37 standard deviations, and the subgroup finding was that tutoring by teachers and paraprofessionals outperformed tutoring by nonprofessionals and parents. The thing that made tutoring work was trained adults. Anything that improves how those adults are trained compounds across every child they ever teach.

That is the honest case for DAImian: it is an attempt to improve the input that the evidence says matters most. There is also no published evidence yet that avatar practice produces better outcomes for real children, and a parent should hold both facts at once.

What to do with this at home

Ask new teachers what they practised on

At a back-to-school night, “where did you do your student teaching?” is a normal question. “Did your programme use simulated students?” is now equally normal and far more revealing. A first-year teacher who rehearsed misconception diagnosis forty times is differently prepared than one who rehearsed on classmates. Our look at how a teacher’s own thinking shapes a child’s learning covers why that preparation carries so much weight.

Learn one misconception in your child’s current math unit

This is the highest-value ten minutes a parent can spend. Find out the specific wrong idea that trips up most children in fractions, or negative numbers, or place value, and listen for it. You will hear it. Naming it to your child (“you’re treating the bottom number like a counting number”) is the same move the candidates are rehearsing.

Treat VR exposure as a separate question for your own child

Teacher training in headsets says nothing about whether children should use them. Those are different populations and different durations. Our overview of kids and VR brain development covers what is and is not known for children specifically.

If you have a college student in education, talk about their data

VR sessions capture gaze, head movement and speech. At a postsecondary institution those recordings are the student’s own records: under FERPA, rights transfer from parent to student when the student turns 18 or enters a postsecondary institution at any age. That means your college sophomore, not you, can ask what is retained and for how long. Tell them to ask.

What not to do

Do not assume a simulated student is a lower standard than a real placement. Done well, it is the opposite: a safe place to be bad at something forty times, which no real classroom should be. The failure mode is not that it is fake. It is that it might be narrow, which is a question about configuration, and configurations can be fixed once someone asks.

What to Watch For Over the Next 3 Months

• Week 4: Watch for other programmes adopting it. Teacher-preparation innovations spread through state consortia, and if a second Wisconsin campus or a nearby state system announces the same approach, this stops being a pilot.
• Month 2 red flags: A programme that cannot describe what the avatar represents. Claims that simulated practice replaces classroom placement rather than preceding it. Any outcome claim this early, since none has been published.
• Month 3 self-check: Can you name one misconception your child currently holds in math, in the child’s own words? That is the artefact teacher candidates are being trained to hear, and hearing it yourself is the fastest way to understand what this training is for.

Frequently Asked Questions

What is DAImian?

An AI-powered avatar that plays a struggling sixth-grader, used with VR headsets by education majors at the University of Wisconsin–Whitewater to practise teaching math, in place of peer role-play. It was reported on September 22, 2026.

Why not just practise on real children?

Novices learning basic instructional moves in front of thirty real students is costly for the students. Simulated practice lets a candidate attempt the same difficult moment repeatedly, which is what builds the skill, before any real child is involved.

Is there evidence this makes teachers better?

Not yet, in terms of outcomes for real children. The reasoning is sound and matches what research says about rehearsing under difficulty, but no published study links avatar practice to student results. Treat it as promising rather than proven.

Could the avatar teach candidates the wrong things about real students?

That is the open risk. A model asked to play a struggling student generates a composite of writing about struggling students, with whatever patterns that writing contains. The useful question for a programme is what the avatar was configured to represent and whether candidates practise on more than one profile.

Does this mean my child will be taught by AI?

No. This is training for human teachers. The AI is the practice partner, not the instructor, and the stated purpose is to improve what the human does in the room.

Who owns the recordings of these practice sessions?

At a postsecondary institution, FERPA rights over a student’s education records belong to the student once they turn 18 or enrol, so a teacher candidate can ask their own programme what data is captured and retained.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Pursuit. (2026). “AI in Education: News, Policies, Innovations” (entry dated September 22, 2026). https://www.pursuit.us/news/ai-in-education-news-policies-innovations
2. Bjork Learning and Forgetting Lab, University of California, Los Angeles. “Research: Desirable Difficulties.” https://bjorklab.psych.ucla.edu/research/
3. Nickow, A., Oreopoulos, P., & Quan, V. (2020, July). “The Impressive Effects of Tutoring on PreK-12 Learning: A Systematic Review and Meta-Analysis of the Experimental Evidence.” NBER Working Paper 27476. https://www.nber.org/papers/w27476
4. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). “Generative AI without guardrails can harm learning: Evidence from high school mathematics.” Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.2422633122
5. U.S. Department of Education, Student Privacy Policy Office. “What is FERPA?” 34 CFR Part 99. https://studentprivacy.ed.gov/faq/what-ferpa
6. UNESCO. (2024). “AI Competency Framework for Students.” https://www.unesco.org/en/articles/ai-competency-framework-students
7. Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). “AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting.” Scientific Reports. https://www.nature.com/articles/s41598-025-97652-6