Show a model your old code and it writes your old bugs: 32 runs, 0% reuse

A recent experiment by a developer demonstrates how AI models are heavily influenced by the existing code in a repository. By testing four different LLMs against two versions of a codebase—one containing outdated, duplicated components and another with a refactored shared component—the study found that models consistently mimicked the patterns they were shown. When exposed to old, buggy code, the models reproduced specific defects, such as incorrect keyboard handling and hardcoded values, rather than adopting the new, cleaner implementation. Even when models were instructed to use the new shared component, they often invented incorrect API interfaces unless the source code was explicitly provided. The findings suggest that AI-assisted development is highly context-dependent, and that maintaining a clean, well-documented repository is essential to prevent models from propagating legacy technical debt and outdated coding patterns.