Overseeing AI-generated code
What a systematic review of 9,502 records says about deciding which AI-generated changes need a human, and six practical steps.
The study
How organizations keep a human in the loop
The driving question: how do software organizations scale human oversight of AI-generated code before it is deployed? The study asks practitioners which oversight gates their organizations keep and why, where those gates strain as AI-generated code grows, which human and automated mechanisms they use, and which approaches they find effective, and compares the answers across organizations of different sizes.
It is part of a Doctor of Technology at Purdue University, expected in 2028.
The literature
What is known, and what isn't
A systematic review following PRISMA 2020 screened 9,502 records from peer-reviewed databases, preprint servers and citation snowballing, and retained 72 studies. AI models assisted with screening and coding; a person made every inclusion decision.
The evidence that AI-generated code carries more risk and gets less review is growing. Most of it comes from open-source repositories. Very little describes how organizations actually staff and run their oversight gates, or what they find works. That gap is what the study addresses.
Open source
Human Oversight System (HOS)
HOS is a working framework for oversight of AI-generated code, built alongside the research (Thurlow, 2026): independent, adversarial risk assessment, cross-vendor agents that review adversarially, risk-tiered human gates, and an append-only audit trail. The approach generalizes the same model beyond code.
Writing
· LinkedIn
· LinkedIn
Code is outpacing our ability to check it
Verification, not generation, is now the bottleneck. Why reviewing everything and trusting the model both fail.
Follow the research
New findings and posts appear on LinkedIn first.