September 2026 Post Claims LLMs Cannot Learn Maintainable Code Due to Missing Long-Term Reward Signals

Nedelcu documents the shift toward AI-generated code where developers stop reading or writing implementations. He lists concrete failure modes: non-deterministic breakage after changes, scattered feature additions, and fragile tests that lock in implementation details. These effects surface only after months or years, outside the immediate reward windows used in reinforcement learning from human feedback.

Dreyfus and Dreyfus 1986 model shows experts operate beyond rule lists by recognizing context-dependent patterns acquired through repeated failure. Nedelcu notes current models train on public code whose median quality is low and on beginner-oriented rulebooks. No linter or benchmark yet encodes long-horizon maintainability, matching SWE-bench observations that functional correctness scores do not correlate with later refactoring cost.

The result is an expertise erosion loop: humans defer choices to systems that never receive delayed feedback, while accumulated technical debt grows undetected. Projects that adopt this pattern at scale will face rising inconsistency rates once original authors depart, because invariants were never explicitly recorded.

Teams that retain mandatory human review of generated diffs and track architectural drift metrics over 12-month windows will preserve coherence; those that do not will encounter compounding maintenance overhead exceeding initial productivity gains.