Why AI-Written Code Rots — and Takes Developer Judgment Down With It

The author pushes back on the fashionable claim that reading and writing code are becoming obsolete. The core problem with AI-generated code, they argue, is maintainability: bad architecture and unmaintainable code have no immediate measurable signal, so their costs only surface over months or years. That makes them invisible to the reinforcement-learning reward loops used to train models, which optimize for what can be measured now. Compounding this, models learn from beginner-oriented rulebooks and from a public corpus that is mostly mediocre code, and there is no fitness function for maintainability precise enough to have been encoded in linters.

The piece points to a concrete symptom: even state-of-the-art models are poor at simplifying code, often shattering functions into smaller pieces that aren’t genuinely reusable and that force a reader to trace through their implementations to understand the original. Defining clarifying, reusable abstractions is treated as an expert skill built from hard-won, context-dependent intuition — the kind of tacit judgment that can’t be reduced to fixed rules, and that most developers (and current AI) haven’t reached.

The deeper worry is human, not technical. As people lean on AI to both write and read code, they stop making decisions, owning mistakes, and learning from them — the exact feedback loop that produces expertise. The AI doesn’t learn from its errors either, so mastery erodes on both sides. The author, who uses LLMs daily and rejects the Luddite label, frames them as a useful but fading tool and predicts that some companies will eventually market explicit ‘no-AI’ policies as a competitive advantage.