Lines of Code Are Back (and It's Worse Than Before)

3 months ago

The software industry historically agreed that lines of code (LOC) is a terrible metric for productivity, as it encourages verbose and inefficient code.
With the advent of AI, LOC has resurfaced as a key metric, with tech CEOs boasting about the percentage of AI-generated code in their projects.
AI-generated code is often measured by volume (e.g., 'Total Lines Suggested' and 'Total Lines Accepted'), ignoring quality, bugs, or whether the code was ever deployed.
Goodhart's Law is exacerbated by AI, as the cost of generating code drops to zero, making LOC an even more meaningless metric.
Studies show that AI-generated code leads to increased duplication, reduced refactoring, and higher churn rates, indicating lower quality and maintainability.
Developers using AI tools often take longer to complete tasks but perceive themselves as faster, highlighting a significant perception gap.
Security flaws are prevalent in AI-generated code, with reports showing 45% of such code contains vulnerabilities.
The industry has shifted to 'acceptance rate' as a new metric, but it suffers from the same flaws as LOC, conflating 'not rejected' with 'valuable.'
Alternative metrics like time-to-value, code half-life, defect origin rate, and comprehension coverage are proposed as better measures of productivity and quality.
The real bottleneck in software development is understanding, design, and judgment, not typing speed or code volume.

Hasty Briefsbeta