How developers actually respond to AI code review feedback

Artificial intelligence coding assistants are generating thousands of code review comments on pull requests, yet researchers had little data on whether developers actually value or act on this feedback. A new empirical study offers the first comprehensive look at how developers respond to AI-generated review comments across major platforms and coding agents.

The study scope#

Researchers examined over 54,000 code review comments produced by five widely-adopted AI agents—including Copilot, Cursor, Codex, Devin, and Claude—across 342 Python repositories on GitHub. The analysis focused on three key dimensions: whether comments were resolved, the experience level of developers who addressed them, and the specific characteristics that made feedback more actionable.

Which agents get noticed#

Resolution rates varied dramatically across different AI tools, with Copilot generating by far the most addressed feedback (72.9% of all resolved comments). This concentration suggests that established tools with broader adoption may have more aligned expectations or better-understood conventions among developers. The study also revealed that core team members typically handle design and evolvability comments, while peripheral contributors tend to focus on resolving functional defect reports.

Why feedback gets ignored#

Through analysis of 470 unresolved comment threads, researchers identified ten distinct patterns explaining why developers choose not to act on AI suggestions. The most common reasons were incorrect or misguided suggestions from the agent and developers' intentional design choices that conflicted with the AI's recommendation. This suggests that many AI comments miss important context about project goals or architectural decisions.

What makes feedback stick#

The strongest predictor of whether a developer would act on a comment was the presence of an inline code suggestion—essentially, when the AI didn't just identify a problem but also proposed a fix. Conversely, lengthy and complex comments were significantly less likely to receive attention, pointing to the value of concise, focused feedback in developer workflows.

The findings provide actionable guidance for improving how AI-generated code review integrates into real development teams: agents should prioritize delivering concrete fixes alongside criticism, keep comments brief, and better account for intentional design decisions that might differ from their suggestions.