A survey of 360 Java developers published today finds that nearly all (98%) are now using artificial intelligence (AI) tools to write code, compared to 54% who also rely on AI to review and test code.
Conducted by BellSoft at a Devoxx Belgium 2026 conference, the survey also finds that only 25% of respondents, however, allow AI to review pull requests before a human.
Instead, developers are pasting stack traces/logs into a chat assistant (49%), allowing an AI agent in their integrated development environment (IDE) to read code and logs to identify issues (46%) and explain unfamiliar errors or code (29%). A total of 44% are also open to the idea that AI might be used to approve PRs, but right now only 1% do so, the survey finds.
Additionally, the survey finds that while 93% of respondents are using AI to debug code, less than a quarter (24%) give AI tools access to debugging tools.
Dmitry Chuyko, a performance architect for BellSoft, said the survey suggests that while adoption of AI is widespread, most developers are reluctant to allow AI to complete tasks on their behalf.
The paradox is that developers who do give the AI tools and platform access to debugging tools are nearly twice as likely (40% vs 23%) to save half their time.
In total, 61% of the developers surveyed said AI saves them more than a quarter of their coding time. At the same time, 60% said reproducing bugs locally is their biggest debugging challenge in terms of the amount of time consumed.
Another 36% named incorrect code and hallucinations as their top concern, followed by security (27%). However, developers most worried about AI security are the least likely to use AI in code review (32% vs 20%) and are more likely to have no regular code review at all (18% vs 7%).
The most widely used AI coding tools are Claude / Claude Code (39%) from Anthropic, followed closely by GitHub Copilot (35%) and then ChatGPT / Codex (11%) from OpenAI.
At this point, it’s apparent that software engineering teams need to embed AI more deeply across the software development lifecycle (SDLC), said Chuyko. Right now, the rate at which code is being generated is so great that much of it is not being reviewed, he added. The only way to review that code will be to rely more on AI agents, noted Chuyko.
The challenge is putting the right guardrails in place to ensure that the code being deployed in a production environment meets quality standards, he added. Otherwise, software engineers will soon find themselves troubleshooting more performance and security issues than ever, noted Chuyko.
Ultimately, most organizations are going to re-engineer DevOps workflows for the agentic AI era. The issue, of course, is how quickly that goal can be achieved before the amount of code being generated by AI coding tools overwhelms existing workflows. In the meantime, however, DevOps teams would be well advised to make sure as much code as possible is reviewed before being deployed in production environments, where it may very easily wind up doing more harm than good.

