AI generated Code is 10% Bullshit (alexanderweichart.de)

A practical observation from a developer: AI-generated code reliably contains roughly "10% bullshit"—small, subtle errors, unnecessary complexity, or fragile logic that slips past cursory review. Because AI outputs are highly dependent on prompts and context, those small defects compound: accepted once, they propagate through the codebase and through future prompts, degrading quality over time. The post points to a vivid analogy — repeatedly asking a model to reproduce an image without change causes gradual drift — showing that iterative regeneration amplifies tiny deviations into meaningful corruption. For the AI/ML community this underscores that generated code should be treated as a draft, not a final artifact. Key technical mitigations: strict human review focused on edge cases, unit/regression tests, type checking and static analysis, deterministic model settings or seed control, more constrained prompting with explicit specs and examples, and CI gates that reject unchecked changes. Tooling to track provenance and automated test coverage can prevent the “10%” from scaling. In short: use AI to accelerate development, but enforce verification and guardrails so small model hallucinations don’t silently erode your codebase.