Human and Machine-Written Code: A Multi-Dimensional Characterization
MSc Thesis Proposal by:
Mohaiminul Hoque
Date: Friday, September 11, 2026
Time: 1:00 PM
Location: Dillon Hall DH264
Abstract:
The rapid integration of AI coding tools raises urgent questions about the quality and security of the code produced. Existing work confirms that AI-generated code exhibits higher error rates than human-written code, but leaves key questions open: which violation types dominate, and how complexity and defects relate. We address these gaps through a large-scale comparison of human-written and AI-generated code. Using existing benchmark corpora spanning multiple programming languages, we apply three different analyses covering code linters, complexity metrics, and security vulnerabilities. Differences are assessed with significance tests, effect sizes, and correction for multiple comparisons. Preliminary analysis indicates that AI-generated code is generally simpler and more repetitive, yet more likely to leave unused variables and debugging statements behind, while human-written code shows greater structural complexity and more maintainability issues. AI-generated code also contains more high-risk security vulnerabilities. These differences point to properties an AI-code detector could use, though it is unclear how much they reflect authorship rather than style. This motivates a further line of work: paraphrasing code to attack existing AI-code detectors, classifying which edits most reliably cause failures, and using those findings to design a more robust one.
Keywords: AI-generated code, Code quality, Static analysis, AI code detection
Thesis Committee:
Internal Reader: Dr. Ikjot Saini
Internal Reader: Dr. Pooya Moradian Zadeh
Advisor: Dr. Muhammad Asaduzzaman
