If you want to know teachers' real attitudes toward AI in the classroom, there are two ways: ask teachers directly, or ask an LLM—"Please predict how Japanese middle school teachers perceive the risks of AI-assisted teaching." The second method is extremely cheap, fast, and increasingly being used—by researchers, policymakers, and edtech companies alike.
Tao, Viberg, Dennison, Wu, and Kizilcec (Cornell University and KTH Royal Institute of Technology) ran a large-scale validation experiment. They used OECD's TALIS survey data—the most authoritative international teacher survey, covering 55 countries and regions. They then had 8 recent LLMs (from 4 model families) predict teachers' responses, under both generic prompts and country-specific prompts, and compared the results against the real data.
The gap was substantial.
- LLMs cannot reliably reflect cross-country differences in teacher perceptions. They compress differences between countries—the variation in educational systems, cultural values, and technological infrastructure gets smoothed away in LLM outputs.
- They overestimate both benefits and risks. LLMs tend to give more extreme judgments.
- Identity prompts ("you are a Japanese middle school teacher") and stronger reasoning capability (high-reasoning models like o3) bring only limited improvement.
- TALIS data was collected in 2024; 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 the model's own bias, and how much from the time lag in the data?
- Differences in attitudes across subjects (programming vs. math vs. language arts) were not analyzed.
The most fundamental problem: LLM outputs themselves are shaping how teachers perceive and discuss AI. If model outputs are inconsistent with real teacher perceptions, then LLM-based AI education policy recommendations may carry systematic bias. The good news: some models, such as Gemini 3 Fast, perform reasonably well at capturing cross-country ordering patterns—useful for assisting hypothesis generation, but no substitute for direct conversations with teachers.
What remains unclear:
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.