The Impact of Explainable AI on Teachers' Trust and Acceptance of AI EdTech Recommendations: The Power of Domain-specific Explanations

Synthesis: The Impact of Explainable AI on Teachers' Trust and Acceptance of AI EdTech Recommendations: The Power of Domain-specific Explanations — Feldman-Maggor, Cukurova, Kent, and Alexandron (2025) adapt Hoff and Bashir's "trust in automation" model to AI in education, proposing that explainable AI (XAI) builds Trust in AI EdTech recommendations indirectly by increasing their understandability. In a mixed-methods, within-subject experiment with 41 in-service chemistry teachers using the AI recommendation tool GrouPer, they find that understandability, trust, and acceptance of AI recommendations are positively correlated and that domain-driven explanations — framed in curricular/pedagogical language — foster greater understandability and trust than purely data-driven (feature-importance) explanations. The authors also surface two situational factors shaping acceptance beyond trust: pedagogical alignment and workload-reduction potential.

Key Findings

• Understandability is the bridge between XAI and trust. Adapting Hoff and Bashir's (2013, 2015) model, the paper positions explainable AI as a design feature whose effect on dynamic, learned Trust is mediated by users' understandability of the system's performance — explainability raises trust by letting teachers validate AI outputs against expectations. All correlations among understandability, trust, and acceptance were positive and mostly strong (r > 0.5, one moderate r = 0.439), strengthening after domain-driven explanations were added (Trust Calibration).

• Domain-driven explanations outperform data-driven ones. Moving teachers from feature-importance ("data-driven") explanations to semantic, curricular-language ("domain-driven") explanations significantly increased understandability (W = 80.5, p = 0.005), learned trust (W = 52, p = 0.002), and acceptance (W = 22.5, p = 0.003). All seven think-aloud teachers reported domain-driven explanations as more influential for building their trust than data-driven ones — evidence that explanations should "speak" the teacher's pedagogical language, not just expose model internals.

• Trust is dynamic and requires validation, not just explanation. Qualitative analysis showed that for some teachers, understandability alone was insufficient to establish trust — several stressed that real classroom experience with the tool was needed before they would fully rely on it, highlighting the situated, dynamic nature of learned Trust rather than a one-shot transparency effect.

• Acceptance also depends on situational factors beyond trust. Bottom-up analysis of think-aloud protocols revealed two additional factors influencing teachers' willingness to accept AI recommendations: pedagogical considerations (whether the tool aligns with their approach to differentiating instruction, reported by 8 of 11) and the workload-reduction potential of the tool (reported by 6 of 11), which the authors treat as acceptance drivers not directly tied to trust and grounded in exploratory qualitative evidence.

What this means for practice

• EdTech designers. Ship domain-driven explanations in curricular language, not only feature importance. Moving teachers from data-driven to domain-driven explanations significantly increased understandability (W = 80.5, p = 0.005), learned Trust (W = 52, p = 0.002), and acceptance (W = 22.5, p = 0.003).
• EdTech designers. Design explanations to support validation rather than persuasion: five of the eleven think-aloud teachers used them to check the tool's groupings against statistics or their own subject expertise.
• Instructors. Expect trust to accrue through classroom use, not from a single walkthrough. Three think-aloud teachers reported that understandability alone was insufficient and that real experience with the tool was required before they would rely on it.
• EdTech designers. Do not optimize for trust alone. Pedagogical alignment with the teacher's approach to differentiating instruction (reported by 8 of 11) and workload-reduction potential (reported by 6 of 11) emerged as separate acceptance drivers.
• Instructors. Pair AI training with data-fluency development so you can interrogate technical, data-driven explanations instead of relying only on the semantic layer.

Limitations

• The study is small and single-context: 41 in-service high-school chemistry teachers (37 female, 4 male) working with one tool (GrouPer) in one national setting, and the authors state that the small sample size limits the generalizability of the results.
• The within-subject staged design carries a carryover risk, since the second stage may be affected by the first; the authors explain that counterbalancing across groups was not feasible because the authenticity requirement limited the number of suitable subjects.
• Understandability, Trust, and acceptance were each measured with a single 5-point item — a deliberate trade-off to keep a protocol of roughly 45 minutes from fatiguing participants — and understandability was captured as teachers' perception of a system attribute rather than the attribute itself.
• The qualitative findings come from 11 of the 41 teachers who completed a semi-structured think-aloud protocol, and the authors report only insights observed in at least three teachers, noting that teachers who did not speak to an issue may still hold an opinion.

Citation

Feldman-Maggor, Y., Cukurova, M., Kent, C., & Alexandron, G. (2025). The Impact of Explainable AI on Teachers' Trust and Acceptance of AI EdTech Recommendations: The Power of Domain-specific Explanations. International Journal of Artificial Intelligence in Education, 35, 2889–2922.