If coding is solved, what now?: Measuring the sloppiness of code
SMRTR summary
Somewhere in the flood of AI-generated code, something quietly went wrong. Large language models can now write code that is technically correct at remarkable speed, but "correct" turns out to be a low bar. The real problem is slop: bloated, repetitive, poorly structured code that balloons project size and quietly erodes human understanding.
Measuring that sloppiness, though, is surprisingly hard. Using AI to judge its own code barely works. Human review doesn't scale. Even promising metrics like "verbosity" and "erosion," borrowed from the research paper SlopCodeBench, reveal a troubling gap: AI-generated code tends to be roughly twice as verbose and structurally degraded as code written by humans.
Perhaps most alarming is what happens when agents are tested in realistic, iterative conditions. Even state-of-the-art models achieve a zero percent strict pass rate when bad decisions are allowed to accumulate over time.
For anyone happily adding hundreds of thousands of lines of AI-generated code each day, that number deserves a moment of pause.
SMRTR provides this summary for quick context. The original article belongs to Hacker News.
Read the original article