If you want to know teachers' real attitudes toward AI in the classroom, there are two options: ask teachers directly, or ask an LLM—"Please predict how Japanese middle school teachers perceive the risks of AI-assisted teaching." The second option is nearly free, fast, and increasingly used by researchers, policymakers, and edtech companies.
Tao, Viberg, Dennison, Wu, and Kizilcec (Cornell University and KTH Royal Institute of Technology) ran a large-scale validation experiment. They used OECD TALIS survey data—the most authoritative international teacher survey, covering 55 countries and regions—and had eight recent LLMs (from four model families) predict teachers' responses under two conditions: a generic prompt and a country-specific prompt. The predictions were then compared against the real data.
The gap was large. LLMs cannot reliably reflect cross-national differences in teacher perceptions. They compress variation between countries—differences in education systems, cultural values, and technology infrastructure get flattened in LLM outputs. At the same time, they overestimate both benefits and risks, tending toward more extreme judgments. Persona prompting ("You are a Japanese middle school teacher") and stronger reasoning ability (high-reasoning models like o3) brought only limited improvement.
The most fundamental problem: LLM outputs themselves are shaping how teachers perceive and discuss AI. If model outputs diverge from real teacher perceptions, then LLM-based AI education policy recommendations may carry systematic bias. The good news: some models, such as Gemini 3 Fast, performed reasonably well at capturing cross-country ranking patterns—useful for hypothesis generation, but not a substitute for talking to teachers directly.
Open questions: TALIS data was collected in 2024, and teacher perceptions may have shifted considerably between 2024 and 2026. The recency of LLM training data is unknown—how much of the gap between model predictions and real data comes from model bias versus temporal mismatch? Differences across subjects (coding vs. math vs. language arts) in teacher attitudes were not analyzed.
---
References
1. Tao, Y., Viberg, O., Dennison, D. V., Wu, Z., & Kizilcec, R. F. (2026). *Teachers' Perceived Benefits and Risks of AI Across Fifty-Five Countries: An Audit of LLM Alignment and Steerability*. arXiv:2605.08486 [cs.CY].
2. OECD. (2024). *TALIS 2024 Results: Teachers and School Leaders as Lifelong Learners*.
3. Kizilcec, R. F., et al. (2024). *Scaling up Behavioral Science Interventions in Online Education*. Proceedings of the National Academy of Sciences.