What AI does to the minds of novice coders

Scientific concepts / discoveries / nature phenomena

Human learning & cognition effects of generative AI in novice programming

• Widening performance gap (“novice polarization”): Generative AI tools may increase differences between learners who find programming easy vs. those who don’t.
• Illusion of competence: Beginners may copy AI-generated code that works on tests while misperceiving their actual understanding, leading them to fail when asked conceptual or edge-case questions.
• “Not knowing what you don’t know”: Learners may lack metacognitive awareness of their knowledge gaps due to AI scaffolding.
• Different cognitive strategies with AI:
  • Struggling beginners may seek speed, skip explanations, and over-rely on AI output.
  • More experienced novices may structure solutions first, treat AI suggestions as minor assistants, and maintain ownership of reasoning.
• Ownership and satisfaction in problem solving:
  • With AI, people can feel less ownership of solutions.
  • This can reduce the intrinsic reward (the “rush, struggle, satisfaction” cycle), potentially affecting motivation and learning persistence.

Educational outcomes and course performance

• Higher failure rates after tool adoption: Other studies cited in the subtitles report increased course failures following introduction of these tools.
• Prerequisite knowledge matters: The effectiveness of AI help depends on having foundational concepts and prior experience; otherwise, it can impair learning and performance.

Methodological ideas (experiment described)

• Participants: Two of 21 students in an introductory programming class (no prior programming experience; similar assignment).
• Interventions/tools:
  • Permission/use of large language models (LLMs) and AI code completion software.
  • Students use AI during homework while being studied.
• Data collection:
  • Eye tracking to observe attention/behavior.
  • Verbalized thought process (think-aloud) while completing the same assignment.
• Observed behavioral patterns:
  • One student repeatedly consults AI but scolls past guidance and copies code leading to superficial success.
  • Another student reads the problem carefully, outlines variables/strategy (comments), and uses AI mostly for small predictable completions.

Practical “rule of thumb” for learning (normative guidance)

• Learn to code without AI first: “Anything you’d ask an AI to write, you should be able to write yourself.”
• Slow practice benefits learning: Doing tasks slowly is positioned as better for learning and memory/comprehension (comparison to handwriting notes).
• Awareness of shortcut-taking: Build the habit of noticing when prompts replace your own reasoning.

List of researchers or sources featured

The subtitles mention a paper and other studies but do not name specific researchers or cite external sources by person. No identifiable researcher names are provided in the text.