Although AI tools are widely used by software developers, these developers and their managers are still working to understand how to best deploy the tools they use.
This is the point from the latest research of 49,000 professional developers by Community and Information Hub Stack Overflow. This has been greatly influenced by the addition of large-scale language models (LLMs) to developer workflows.
The survey found that four in five developers used AI tools in their workflows in 2025. This is a part of the process that has grown rapidly in recent years. “Reliability in AI accuracy has declined from the past 40% to just 29% this year,” he said.
The disparity between these two metrics demonstrates the evolutionary and complex impact of AI tools such as Github Copilot and Cursor on the job. There is little argument among developers that a tool should or should be useful, but people are still thinking about what the best application (and limitations) is.
When asked what their best complaints about AI tools were, 45% of respondents said they struggled with “almost correct but not perfect AI solutions.” This is unlike the obviously wrong output, as these can introduce other issues that are difficult to identify right away and relatively time-consuming, especially for junior developers who approached their work with false confidence thanks to their reliance on AI.
As a result, more than a third of research developers reported that “part of the visits to stack overflows are the result of AI-related issues.” In other words, code suggestions accepted from LLM-based tools introduced problems that had to rely on others to solve.
Even if large-scale improvements have been recently brought about through optimized models, it is unlikely that their inference, but unreliable properties will disappear completely. It is inherent to the essence of how prediction technology works.
