AI tools are widely used by software developers, but these developers and their managers are always struggling with determination in the best way to use the tools to be used, increasing pain emerging along the way.
This is to remember from the latest survey of 49,000 professional developers by the community and information HUB Stackoverflow, which has itself been strongly affected by the addition of large-language models (LLM) to the workflows of developers.
The survey revealed that four out of five developers are using AI tools in their workflow in 2025 – a game that has increased rapidly in recent years. That said, “confidence in AI accuracy increased from 40% in previous years to only 29% this year.”
The disparity between these two measures illustrates the evolution and complex impact of AI tools such as the co -pilot or the Github cursor on the profession. There are relatively few debates among the developers that the tools are or should be useful, but people always determine the best applications (and limits).
When they were asked what their best frustration was with AI tools, 45% of respondents said they had difficulties with “IA solutions that are almost correct, but not quite” – the biggest problem reported. Indeed, unlike the outings that are clearly wrong, these can introduce insidious bugs or other problems which are difficult to identify immediately and relatively time to help out, in particular for junior developers who have addressed work with a false feeling of confidence thanks to their dependence on AI.
Consequently, more than a third of the developers of the survey “report that some of their visits to stacking the overflow are the result of AI -related problems”. That is to say, the code suggestions that they accepted from an LLM-based tool introduced problems that they then had to turn to other people to solve.
Even if major improvements have recently come via models optimized for reasoning, this lack of near but not very premensual reliability will probably not disappear completely; It is endemic to the very nature of the functioning of predictive technology.
