ChatGPT Friend or Foe? Cognitive Study of LLM-Assisted Code Comprehension
arXiv·medium signal
Exploratory lab study of 10 advanced developers (5 with AI, 5 without) making non-trivial extensions to unfamiliar code. Using Polya's problem-solving framework, researchers found AI-assisted developers explored fewer solution paths and developed shallower mental models of the codebase, even when producing correct code faster. Raises questions about long-term skill development when relying on LLM coding assistants.