Regulating the AI Tutor: Intentions, Help-Seeking, and Self-Regulated Learning in Adolescent GenAI Use

Technical Summary: Regulating the AI Tutor

Problem Statement
Generative AI (GenAI) tools have become common learning companions for adolescents, yet the mechanisms by which students regulate their use during authentic learning tasks remain poorly understood. While Self-Regulated Learning (SRL) and high-level Help-Seeking (HS) are theoretically proposed as safeguards against passive or shortcut-oriented use, existing empirical research predominantly focuses on aggregate learning outcomes rather than the moment-to-moment processes of AI-supported learning. There is a specific risk of "cognitive surrender," where students adopt AI-generated reasoning with minimal scrutiny, suppressing rather than supporting their own thinking. Current studies often lack a process-sensitive account grounded in SRL and HS theory to examine how students actually regulate learning with AI as it unfolds.

Methodology
This work-in-progress presents a classroom study involving 98 Grade-9 students (aged 14–15) from three German Gymnasium schools. The study utilized a web-based Mistral-Large tutor for a curriculum-aligned mathematics skill preparation task prior to an exam.

• Procedure: The study followed a five-phase protocol: (1) Pre-questionnaire (AI literacy and learning strategies); (2) Timed math pre-test; (3) Chat interaction with the AI tutor on a real-world mathematical modeling task; (4) Post-chat cognitive load assessment (intrinsic, extraneous, germane); and (5) Post-test on mathematical modeling.
• Data Collection: The dataset comprises 1,616 chat turns (808 student turns). Prior to the chat, students selected learning goals from eight options (e.g., concept explanations, step-by-step examples, final solutions) and rated their confidence.
• Analysis Framework: The authors propose a hybrid turn-level coding framework combining:
  • Deductive Codes: Based on Zimmerman's cyclical SRL model (PLAN, MONITOR, REQUEST, EVALUATE) and HS theory (instrumental vs. executive; conceptual, procedural, verification, answer-seeking). A mathematical modeling code based on Blum's framework distinguishes surface-level computation from deeper modeling activities.
  • Inductive LLM-Specific Codes: Derived via exploratory analysis, these include Epistemic Vigilance (detecting student reactions to inaccurate AI turns) and Agency over the AI (attempts to shape the AI's role or scaffolding).
• Coding Process: Preliminary results rely on AI-based coding (Gemini 2.5 Pro) of the full dataset. A stratified 30% human-validation pass by experts in mathematics didactics and learning sciences is ongoing.

Key Results
Preliminary analysis reveals significant misalignments between student intentions, enacted behaviors, and learning outcomes:

1. Forethought vs. Enactment: Students reported strong learning-oriented intentions, selecting goals like step-by-step examples (82.9%) and concept explanations (69.7%). However, their enacted behavior was dominated by REQUEST turns (72.9% of task-relevant turns), with PLAN being the only other frequent function (18.1%). Crucially, MONITOR and EVALUATE behaviors were nearly absent (medians of 0%).
2. Help-Seeking Patterns: While requests were predominantly instrumental (75.1%) rather than executive, the content was heavily procedural (37.4%) and conceptual (25.1%). Despite 69.7% of students stating they wanted to check their understanding, the actual rate of verification requests was low (16.3%).
3. Intent-Behavior Gap: There was a marked disconnect between stated intentions and enacted behaviors. Common learning-oriented intentions (e.g., seeking verification) did not significantly predict corresponding interaction codes. The only significant alignment was found for students who selected "just give me the final solutions," who subsequently engaged in higher rates of executive and answer-seeking requests.
4. Learning Outcomes: Contrary to expectations, post-test performance (56.9%) was significantly lower than pre-test performance (67.5%, p = .014). In an OLS model controlling for prior knowledge, higher extraneous cognitive load was the only significant predictor of lower post-test scores (b = −.034, p = .025).

Significance and Claims
The paper positions itself as a work-in-progress providing an initial, process-sensitive account of adolescent GenAI regulation. Its primary significance lies in demonstrating that:

• Instrumental help-seeking can remain shallow: Even when students request support rather than answers, the lack of explicit monitoring and evaluation prevents deep regulation of understanding.
• Intentions do not guarantee regulation: Students may possess the metacognitive knowledge of desirable AI uses but lack the interactional strategies to enact them in real-time.
• Interaction processes matter more than aggregate gains: The decline in performance and the link to extraneous load suggest that AI tutors may introduce new cognitive demands (e.g., prompt formulation, interaction management) that compromise meaningful learning if not properly regulated.

The authors argue that the core challenge is not whether students ask for help, but whether they can regulate that help-seeking to preserve agency, monitoring, and mathematical reasoning. They propose that future scaffolds must actively support the monitoring, evaluation, and "epistemic vigilance" behaviors that adolescents do not yet bring to the interaction autonomously. The paper concludes by noting that final reliability metrics and full validation of the codebook are pending.