Why AI Coding Assistants Repeat Mistakes and How 'Painful' Training Can Fix It

AI coding assistants trained on vast public internet data tend to generate code optimized for ideal 'happy path' scenarios, leaving them ill-equipped for real-world edge cases like race conditions, memory leaks, and network failures. This gap — dubbed the 'Straight-A Intern Paradox' — occurs because most training data comes from tutorials and classroom assignments that rarely model production-level complexity. Engineering teams often respond by writing rules-based prompt instructions forbidding specific bad patterns, but this approach frequently backfires as language models fixate on the very patterns they are told to avoid. The proposed alternative is to embed experiential, consequence-driven context into AI guidance — essentially teaching models through simulated 'scars' rather than abstract prohibitions. The argument draws on the idea that wisdom in engineering, as in human experience, comes not from rules alone but from internalizing the cost of failure.