AI-generated code may be creating more problems than it solves, particularly in areas of security and correctness, according to a new analysis by CodeRabbit. While artificial intelligence tools are increasingly used to speed up software development, the findings suggest that they also introduce a higher volume of issues, placing additional pressure on human reviewers.
CodeRabbit’s data shows that pull requests created with the assistance of AI tools contained an average of 10.83 issues, compared to 6.45 issues in pull requests written entirely by human developers. This difference often results in longer review cycles and increases the risk of bugs slipping into production environments.
Overall, AI-assisted pull requests were found to have 1.7 times more issues, including 1.4 times more critical issues and 1.7 times more major issues, highlighting that the problems go beyond simple or cosmetic errors.
The analysis revealed that logic and correctness errors were 75 percent more common in AI-generated code. Issues related to code quality and maintainability rose by 64 percent, while security-related problems increased by 57 percent.
Performance-related issues were also 42 percent more frequent. Common vulnerabilities identified in AI-assisted code included improper password handling, insecure object references, cross-site scripting flaws, and insecure deserialization.
Despite these concerns, the report notes some benefits of AI-assisted development. AI-generated code showed fewer spelling errors and fewer testability issues, indicating improvements in certain aspects of code cleanliness and structure. These gains suggest that AI can be useful for handling repetitive tasks and improving consistency.
Rather than replacing human developers, the findings indicate that AI is reshaping their role. Developers are increasingly required to review, validate, and correct AI-produced code.
As AI-driven development accelerates overall code output, organizations may see a rise in absolute vulnerability numbers, even if the relative quality of code does not significantly decline. Ongoing improvements to AI models aim to address these weaknesses, but human oversight remains critical.
