LLMs amplify code quality, they don't improve it
A critical reflection argues that LLMs do not inherently improve code quality but instead amplify existing patterns—producing clean code from clean input and compounding flaws from poor or "rotten" code, placing the onus for good input on the developer.
Background
- The author argues that large language models (LLMs) like ChatGPT, Claude, or GitHub Copilot are "lossy amplifiers" of existing code quality: they reproduce whatever patterns they're fed, for better or worse.
- This contradicts the common marketing claim that AI-assisted coding tools inherently improve code. The post asserts that if your codebase is messy, LLMs will generate equally messy (or messier) output.
- The underlying issue is that LLMs have no internal judgment ("taste") — they statistically replicate the style and quality of their training data, which in practice includes a huge amount of low-quality, copy-pasted, or obsolete code from the open web.
- This is a recurring debate in developer circles: tools like Copilot increase raw output speed, but critics argue they also accelerate the accumulation of technical debt unless developers maintain strict code review discipline.
- The post's "on you" conclusion reflects a broader industry concern that AI assistance shifts responsibility for code quality from the tool to the human, who must still curate inputs carefully.