Compassionate Feedback in AI-Assisted Assessment: A Comparative Analysis of Novice and Experienced Teachers' Feedback Patterns within the Integrated AI Triad (IAT) Framework

<title>Abstract</title> <p>The integration of generative AI into educational assessment has transformed feedback practices, yet the extent to which AI-generated feedback supports compassionate teaching—integrating emotional attunement with precise cognitive scaffolding— remains critically underexamined, a gap this study addresses. This study investigates how novice (n = 26, 0–3 years) and experienced (n = 31, 7 + years) teachers generate compassionate feedback for weak student responses within the Integrated AI Traiad (IAT) Framework. Employing a comparative qualitative content analysis of 57 feedback entries generated across a 16-session professional development workshop, we examined structural patterns, linguistic tone, IAT alignment, and the effects of AI tool selection, subject area, and session progression. Experienced teachers produced significantly more specific feedback (M = 4.0/5 vs. M = 2.8/5) with stronger cognitive scaffolding (M = 3.8/5 vs. M = 3.0/5), while novices demonstrated marginally higher but less targeted compassion (M = 4.0/5 vs. M = 3.7/5). DeepSeek generated more Socratic, layered questioning than ChatGPT or Copilot, though tool benefits were moderated by teacher experience. Feedback quality improved significantly across sessions 1–3, with a plateau in session 4—suggesting condensed three-session workshops may achieve comparable outcomes while reducing implementation costs by approximately 25%. Negative case analysis revealed 19% of novices outperformed the experienced mean and 16% of experienced teachers underperformed the novice mean, indicating that experience alone does not determine feedback quality. The findings extend the IAT framework by operationalising compassionate feedback as a measurable dual-construct practice, with implications for teacher professional development, AI tool selection, and the design of assessment systems that preserve human pedagogical judgment. We discuss limitations, including sample size and the overrepresentation of science topics, and propose directions for future research.</p>