| Sumario: | Background: Despite growing integration of generative AI in educational settings, little is known about whether and to what extent AI‐generated real‐time feedback can support first‐grade students in solving multimodal mathematical tasks. Objectives: Our study aims to examine student–AI feedback dynamics from a process‐oriented perspective, focusing on how students interact with AI feedback after incorrectly solving multimodal mathematical tasks and which interaction patterns lead to successful re‐attempts. Methods: We analyze data from 13,000 first‐grade pupils across two countries who completed a single‐session assessment of 16 foundational numeracy skills via multimodal tasks, where incorrect and correct responses triggered real‐time feedback from a GPT‐4.1‐based system. All feedback instances were collected and qualitatively coded to investigate how student‐AI interaction sequences differ between successful and unsuccessful re‐attempts, and to what extent specific actions or interaction patterns predict successful attempts. Data were examined using Heterogeneous Transition Network Analysis and sequential pattern mining. Moreover, chi‐squared tests and regression models were used to test the association between successful and unsuccessful attempts and interaction types and sequences thereof. Results: Findings revealed that feedback effectiveness depended not on type alone but on how it was embedded within broader interaction sequences. Successful re‐attempts were more likely to follow clarification‐oriented feedback, whereas unsuccessful ones were more commonly preceded by question‐based prompts or direct orders, which reflect more highly anthropomorphic and authoritative forms of address. The absence of AI feedback was associated with poorer recovery. Conclusions: These findings are situated within the broader concern that misaligned generative AI scaffolding and feedback may deepen confusion or reinforce misconceptions instead of supporting learning. Lay Summary: What is currently known about this topic? ○Generative AI can provide instant, personalized feedback in math○Most studies focus on test scores, not learning processes○Multimodal tasks increase cognitive demands for young learners.○Little is known about AI feedback for early numeracy skills.What does this paper add? ○We analyze 13,000 pupils' real‐time interactions with AI feedback○We study how feedback sequences relate to success or failure.○Clarifying feedback supports more successful retries○Feedback sequencing matters more than type.Implications for practice or policy ○Design AI feedback to clarify concepts before prompting action○Avoid overreliance on directive or question‐only feedback○Any type of (AI) feedback is better than none
|
|---|