AI coding tools have become a crucial part of software development workflows, offering productivity gains to developers worldwide. However, their output is often deemed untrustworthy, leading to manual reviews and potential losses in productivity.
According to a survey conducted by Qodo, 82% of respondents use AI coding tools at least weekly, with 78% reporting improved productivity. Yet, the lack of confidence in these tools’ accuracy hampers some of these gains.
“It’s a big net positive, but the gains aren’t evenly distributed,” said Itamar Friedman, CEO and co-founder of Qodo. “We see massive gains from power users, but moderate gains for most developers. The rest are struggling to effectively leverage AI tools.”
The survey revealed that 60% of developers believe AI has improved or somewhat improved overall code quality, while 20% report a degradation. Power users tend to experience more benefits, with 10Xers (highly experienced developers) seeing significant gains.
Tech leads and reviewers often feel pressure due to increased code volume, leading to more review work, oversight, and stress. To address this, 76% of respondents refuse to ship AI-suggested code without human review, opting for manual rewriting or reviewing instead.
AI is particularly effective in code reviews, with 81% of developers using it for reviews reporting quality improvements. However, hallucinations (syntax errors or incorrect calls) are a significant concern, affecting about three-quarters of respondents.
Friedman suggests strategies to mitigate these issues, including prompting the AI agent to review codebase structure and documentation before actual development tasks. Providing clear specifications and generating tests that comply with them can also help improve accuracy.
The most requested improvement by devs was “improved contextual understanding,” indicating a need for better information input into AI models. Friedman emphasizes the importance of providing detailed context, including product requirements and coding styles, to avoid “garbage in, garbage out.”
Organizations offering these tools must ensure that input complies with corporate policies. By doing so, they can help flatten the learning curve for developers to effectively use AI models, unlocking their full potential.
Source: https://www.theregister.com/2025/06/12/devs_mostly_welcome_ai_coding