LLMs can churn out code faster than any human, but a multivocal literature review of 104 sources (31 formal, 73 grey) finds they're also inventing new forms of technical debt that existing tools don't track.
Five New Debt Categories That Come With LLM Adoption
The paper identifies traditional debt amplified by LLMs - code, design, documentation - plus five fresh categories. Fast-integration debt is the standout: code that works today but was generated so quickly that quality took a back seat. That speed triggers a domino effect into governance debt (loss of control over the codebase) and balloons maintenance costs. Prompt debt, ethical debt, data debt, and provenance debt round out the list, each reflecting unique failure modes of LLM-assisted development.
What Detection Tools Exist And Where They Fall Short
Practitioners lean on SonarQube to catch conventional technical debt indicators. Research prototypes like CodeSmellEval are being built to assess how LLMs themselves contribute to debt. The review notes human-in-the-loop frameworks, prompt engineering, and data quality alignment as common mitigation strategies. But none of these are standardized.
The Benchmark Gap That Keeps Things Messy
No LLM-specific benchmarks or metrics exist yet for measuring technical debt in generated code. That's a gap the authors call urgent. Until someone ships a repeatable, automated way to quantify fast-integration or prompt debt, teams are flying blind on long-term cost.
Closing the loop will require tooling that scores not just correctness and style, but the provenance and governance risk baked into every LLM-produced line.
Source: Faster Code, Deeper Debt? A Multivocal Literature Review on Technical Debt and Its Early Signs in LLM-Assisted Software Development
Domain: arxiv.org
Comments load interactively on the live page.