AI-Generated Code is Twice as 'Sloppy' as Human-Written Software
New research reveals that while LLMs pass functional tests, they produce significantly more verbose and eroded code than humans.
Large language models can now generate code that is formally correct, but they are increasingly producing what researchers call 'slop.' This phenomenon involves the creation of unnecessary abstractions, duplicate logic, and poor architectural choices that compromise long-term maintainability.
According to research by Sebastian of Earendil, AI-generated code is on average roughly twice as verbose and eroded as code written by humans. Quantitative metrics highlight a stark divide: verbosity in established repositories sits at 0.15 ± 0.06, while agent-generated code jumps to 0.33 ± 0.10. Similarly, erosion levels in established repos are 0.31 ± 0.17, compared to 0.68 ± 0.20 in code produced by AI agents.
The Rise of Vibe-Coding
This trend emerges as the industry shifts toward 'vibe-coding' and the use of end-to-end agents. As LLMs become proficient at passing basic functional tests, there is a growing tendency to prioritize immediate results over structural integrity. This has led to an explosion in total lines of code (LOC) and a corresponding loss of human agency. As codebases expand rapidly, they often become too large for humans to manage effectively and too sloppy for the AI agents themselves to self-correct.
Sebastian notes that "just because the code is formally correct doesn’t mean that it is not introducing unnecessary abstractions, creating duplicates, or just making bad decisions overall."
The Technical Debt Trap
The implications for the software industry are significant. If AI is used to scale production without a mechanism to measure and mitigate this 'slop,' the result may be an ecosystem of unmaintainable software. The research suggests that current AI agents are fundamentally unable to handle their own technical debt.
This was evidenced in SlopCodeBench evaluations, which mimicked iterative human use. In these tests, state-of-the-art models achieved a 0% pass rate for strict solve rates across all checkpoints. This failure indicates that as a project evolves, AI agents struggle to maintain quality, leading to a compounding accumulation of architectural decay.
The Future of Architecture
These findings suggest that while the basic act of writing syntax may be 'solved,' human intuition and taste remain critical for software architecture. The industry faces a paradox where increasing the volume of code production through AI may actually decrease the overall health of the software.
Moving forward, the challenge lies in developing metrics that go beyond functional correctness. Until agents can recognize and prune their own verbosity, the role of the human developer will likely shift from writing code to acting as a critical editor and architectural guardian to prevent systemic erosion.