Evidence note

Human-Written vs. AI-Generated Code: A Large-Scale Study of Defects, Vulnerabilities, and Complexity

Peer-Reviewed Research2025IEEE ISSRE 2025

What was examined?

A large comparison of human-written and AI-generated Python and Java samples using defect, vulnerability and complexity analysis.

Comparison of human-written code with samples generated by ChatGPT, DeepSeek-Coder and Qwen-Coder.

More than 500,000 Python and Java code samples.

Key findings

01AI-generated code was generally structurally simpler and more repetitive in the studied samples.
02Human-written code showed greater structural complexity and more maintainability-related issues.
03AI-generated samples contained more high-risk security vulnerabilities in the study.
04The results indicate different defect profiles rather than one universal quality ranking.

Why it matters. Lornets interpretation.

Code provenance is relevant context but not a technical verdict. Different development approaches can create different quality profiles, so production assessment should evaluate the actual properties that matter for the system.

This is the Lornets reading of the source, not a finding of the source itself.

What it does not establish

01The result does not establish that all AI-generated code is less secure.
02It does not establish that all human-written code is harder to maintain.
03Results are specific to the studied languages, models and evaluation methodology.
04Code-sample analysis is not equivalent to assessment of a complete production system.