787,562 Function Pairs Show AI Code Is Half the Size of Human Code With Different Defect Classes, Not Fewer

arXiv 2609.12708·high signalXBlueskyLinkedInCopy link

Each human function was mined from open source and paired with implementations generated from its docstring by GPT models, DeepSeek-Coder, and Qwen2.5-Coder across Python, Java, and C, then mapped onto Orthogonal Defect Classification and CWE. AI code is structurally compressed and stylistically templated at roughly half the size and branching of human code, human code concentrates the issues of mature codebases while AI code concentrates repetitive boilerplate, and security cuts both ways: LLMs produce more and more severe findings in Python and Java but fewer high-severity memory-safety findings than humans in C. Once size is controlled for, complexity metrics carry almost no signal while naturalness still separates authors, and the team releases CQBench with 27,346 issue-prone tasks.