Unpacking Vibe Coding: Help-Seeking Processes in Student-AI Interactions While Programming

A study of 19,418 AI-student interactions reveals that top-performing programmers use generative AI as a tutor through exploratory questioning, while low performers delegate tasks passively. The research demonstrates that current AI systems passively mirror student intent rather than actively promoting learning, highlighting a critical gap in pedagogical design for educational AI tools.

This research exposes a fundamental limitation in how generative AI currently operates within educational contexts. The distinction between instrumental help-seeking (inquiry-driven) and executive help-seeking (task delegation) reveals that AI systems amplify existing student behaviors rather than correcting unproductive patterns. Top performers leverage AI as a collaborative thinking partner, asking clarifying questions and exploring problem spaces, while struggling students use AI as a shortcut to bypass cognitive engagement—a pattern the AI systems enable rather than discourage.

The findings emerge at a critical juncture for AI in education. As institutions rapidly adopt generative AI for programming instruction, the default behavior of these systems—compliance and task completion—creates a pedagogical liability. Most large language models are optimized for user satisfaction and task completion, not learning outcomes. This creates a perverse incentive structure where students who most need guided learning instead receive frictionless code generation, potentially widening achievement gaps.

For educational technology providers and institutions, this research signals that passive AI integration underperforms compared to pedagogically designed interventions. The market for AI tutoring systems assumes capability, but this study demonstrates that without deliberate design constraints and adaptive steering, AI becomes a tool for cognitive outsourcing rather than augmentation. Organizations investing in AI-powered education platforms must incorporate detection mechanisms for unproductive delegation and implement guardrails that redirect students toward inquiry-based engagement.

Looking forward, the next phase of educational AI requires systems that actively diagnose learning patterns and adapt responsiveness based on pedagogical principles. Success will depend on balancing assistance with productive struggle—providing support that maintains cognitive load rather than eliminating it.

• →Top students use AI for exploratory learning through questioning; low performers use it for task delegation without cognitive engagement
• →Current AI systems passively comply with user intent rather than actively promoting learning or detecting unproductive patterns
• →AI-assisted programming education risks widening achievement gaps by providing frictionless solutions to struggling students
• →Effective educational AI requires pedagogical design that detects and redirects delegatory behavior toward inquiry-based interactions
• →Educational institutions must implement adaptive guardrails in AI tools to ensure learning augmentation rather than cognitive outsourcing