AI usage generates a fragile codebase

AI usage for coding is being blamed for creating fragile codebases that break when changed and increase token consumption for analysis. The author, a 20‑year engineer now using AI for all code, says the problem shows up only after the code appears to work. The symptom is a surge in bugs and higher costs to read the code.

The fragility stems from several maintainability dimensions: lack of DRY, rising cyclomatic complexity, unclear naming, and poorly architected flows. Early AI tools produced shallow tests, which amplified these issues, though newer models like Astra and Fable 5.1 claim better test generation. Overly complex code appears when developers let the AI over‑engineer without constraints.

The author questions whether existing articles capture these dimensions rigorously. Without deterministic gates or adversarial reviews, AI output can slip through quality checks. The claim is that programming remains a high‑level mathematical activity, and AI must be guided to respect abstraction boundaries.

For anyone building software with AI, the takeaway is to enforce strict review pipelines, limit over‑engineering, and apply Pareto‑style limits, aim for 80 % of the result with 20 % of the effort. Keep code DRY, monitor cyclomatic complexity, and require clear naming conventions to prevent the hidden decay that AI can introduce.

TakeawayAdd deterministic reviews and guardrails to keep AI‑generated code from becoming hard to maintain.