Teaching Teachers to Question AI: Ethics Training Beats Tool-Only Workshops

Generative artificial intelligence has swept into classrooms faster than almost any technology in the history of education, and schools are still scrambling to decide how teachers should be prepared for it. A new randomized controlled study from researchers at the University of Oulu in Finland and partner institutions in Vietnam suggests that the answer is not simply more training on how to use AI tools. Instead, the study finds that adding a mindset component—covering AI ethics, human-centered pedagogy, and critical reflection—to technical training produces teachers who think far more carefully about the risks and responsibilities that come with AI, even if it temporarily tempers their confidence.

The research, published in the Journal of New Approaches in Educational Research, compared two approaches to professional development among 57 preservice teachers enrolled in a teacher education program in Vietnam, whose average age was about 21. Participants were randomly assigned to one of two groups. The first group received a combined workshop that included both hands-on practice with generative AI tools and a mindset module on AI ethics, growth mindset, and human-centered instructional design. The second group received only the tools-focused training, covering prompt engineering and demonstrations of how generative AI can produce lesson plans, videos, infographics, songs, rubrics, and slide decks. To keep the comparison fair, the tools-only group was given access to the mindset module only after the study’s post-intervention questionnaires were completed.

The design of the combined workshop was deliberate. The mindset module, titled “Teachers’ New Mindsets for the AI Era,” came first, immersing participants in real-world problems such as algorithmic bias, AI hallucinations, misconceptions about what AI can do, and discrimination. Only after wrestling with these issues did participants move on to technical practice. The researchers grounded this sequencing in Merrill’s First Principles of Instruction, which hold that learning is most effective when it begins with authentic problems, activates prior knowledge, and then moves to guided application. By confronting ethical and pedagogical dilemmas before touching the tools, the teachers were primed to use technology reflectively rather than automatically.

To measure outcomes, the team used a mixed factorial design with pre- and post-intervention questionnaires. Anxiety was assessed with an adapted version of the Abbreviated Technology Anxiety Scale, while competence self-efficacy was measured with the Teacher AI Competence Self-efficacy (TAICS) instrument, chosen because it aligns closely with UNESCO’s AI Competency Framework for Teachers and the AI-TPACK model. TAICS captures six dimensions of teacher AI competence: AI knowledge, AI pedagogy, AI assessment, AI ethics, human-centered AI, and AI-related professional engagement. The researchers analyzed the data using linear mixed-effects models, controlling for each participant’s motivation to adopt AI and their frequency of AI use, and supplemented the quantitative results with a directed content analysis of participants’ written reflections.

The quantitative results revealed a striking asymmetry. Both groups showed an overall reduction in AI anxiety over the course of the training, but only the tools-only group’s anxiety reduction reached statistical significance. More tellingly, the tools-only group reported significant gains in self-efficacy across five of the six competence domains, including AI knowledge, pedagogy, assessment, human-centered education, and professional engagement. The mindset-inclusive group, despite receiving exactly the same technical instruction, showed a significant improvement in only one domain: human-centered AI competence, which measures the capacity to critically evaluate both the benefits and the risks of AI in education. Motivation to adopt AI emerged as a consistent and significant predictor of self-efficacy across all models, with medium effect sizes, suggesting that intrinsic interest in the technology is a central driver of teachers’ confidence.

At first glance, the tools-only group’s confidence surge might look like success. The researchers argue it may instead be a warning sign. They draw a parallel to the Dunning–Kruger effect, the well-documented tendency for people with limited expertise to overestimate their own competence. Teachers who feel highly confident about AI without having grappled with its limitations may over-rely on AI outputs, overlook data privacy and bias concerns, and underestimate the need for human oversight. A teacher who assumes AI is always accurate, for example, might skip fact-checking and inadvertently pass incorrect information on to students. The study’s authors suggest that structured ethical checkpoints, such as collaborative reviews of AI-generated content or bias audits, could help counteract this unwarranted confidence.

The qualitative data reinforced this interpretation. Reflections from the mindset-inclusive group were markedly more nuanced. Participants in that group frequently acknowledged AI’s limitations, stressed the need for human supervision, and worried about over-dependence. One wrote that educators should focus on “understanding AI limitations, using AI as a supporting tool without reliance on it,” while another cautioned peers to “avoid abusing AI or AI will replace you.” Several described the training as transformative for their professional growth, with one participant saying they felt they had “upgraded” themselves in applying AI to teaching. By contrast, the tools-only group’s reflections were enthusiastic but largely focused on practical utility and novelty—learning to write complete prompts, create teaching materials, and make lessons more lively—with little explicit consideration of ethics, risks, or pedagogical responsibility.

The researchers interpret the mindset group’s more modest confidence gains not as a failure but as a natural stage in professional learning. When educators engage in deep critical reflection on complex ethical and pedagogical issues, they often experience a temporary dip in confidence before reconstructing a more robust and realistic sense of competence. Moderate, task-relevant anxiety, the authors note, has been shown in prior research to promote deeper learning and more conscientious teaching behavior. In this framing, the heightened ethical awareness of the mindset group may have produced a kind of productive discomfort—a reflective tension that fosters pedagogical intentionality and guards against the technosolutionist assumption that every classroom problem has a technological fix. Future professional development, they suggest, should pair brief hands-on tool practice with guided reflection exercises such as journals or peer-review tasks to support both immediate skill-building and sustained growth.

The findings carry weight because they arrive amid a global policy push to define what teachers actually need to know about AI. UNESCO’s AI Competency Framework for Teachers, introduced in 2024, identifies a human-centered mindset and ethical awareness as core domains alongside technical knowledge, pedagogy, and professional learning. Yet most teacher training on AI remains predominantly tool-centric, and most empirical studies have treated professional development programs as undifferentiated wholes, making it difficult to isolate the contribution of ethics and pedagogy content. By experimentally separating these components, the new study provides rare direct evidence that mindset-oriented content does something tools training alone cannot: it cultivates ethical judgment and critical reflection, even at the cost of a short-term confidence boost.

The authors also caution that their results come with limitations. The study relied on self-report measures and written reflections rather than objective indicators such as scored lesson plans or classroom observations, the intervention lasted only a single day with immediate post-testing, and the sample consisted entirely of preservice teachers from one institution in Vietnam—a context marked by significant disparities in digital infrastructure between urban and rural regions. Longitudinal follow-up and replication across culturally and technologically diverse settings will be needed to confirm that the effects endure. Still, the central message is clear: preparing teachers for the age of generative AI is not just about technical proficiency. Teacher preparation programs and policymakers, the researchers argue, should place ethical reflection and human-centered pedagogy on equal footing with hands-on skills, so that future educators can teach with, about, and even critically against AI in socially responsible ways.

Subject of Research: The effects of mindset-oriented versus tool-focused professional development on preservice teachers' responsible use of generative AI

Article Title: Mindset matters: Fostering teachers’ responsible AI use through professional development

Article References: Huynh, L., Le, T. H., Dang, B., An, B. T., Vu, C. T., & Nguyen, A. (2025). Mindset matters: Fostering teachers’ responsible AI use through professional development. Journal of New Approaches in Educational Research, 14(1), Article 23. https://doi.org/10.1007/s44322-025-00043-y

Image Credits: AI Generated

DOI: 10.1007/s44322-025-00043-y

Keywords: generative AI, teacher professional development, AI ethics, preservice teachers, AI competence, human-centered education, self-efficacy, AI anxiety, UNESCO AI Competency Framework, critical reflection, educational technology, randomized controlled trial