IA coding tools are being widely adopted by developers, but As its use increases, trust in them decreasesaccording to a report by the community and Stack Overflow Information Center after surveying 49,000 professional developers.
The survey revealed that four out of five developers They use AI tools in their workflow in 2025, a percentage that has experienced rapid growth in recent years. That said, “The confidence in the accuracy of AI has decreased from 40% in years prior to only 29% this year”they indicate. The main reason: the hidden tax of the “almost correct” code, which Venturebeat speaks.
IA coding tools
The world of software development was one of the first to adopt artificial intelligence tools, but like other sectors (see agricultural AI) it still has a way to travel and improve for generalized implementation. Simply, they still need wide human supervision. Although these technologies are expected to serve as a powerful tool to improve productivity, allowing developers to work more efficiently, authorized voices that say they say IA will not replace programmers in the near future.
The disparity of results among the metrics analyzed by Stack Overflow illustrates the complex and evolutionary impact of AI tools such as Github Copilot or cursor in the profession. There is relatively little debate among developers on whether the tools are or should be useful, but You are still investigating what your best applications and its limitations are.
When asked what their main frustration was with AI tools, almost half of the respondents claimed to have difficulties with the “IA solutions that are almost correct, but not quite”the biggest problem reported. This is because, unlike clearly erroneous results, these can introduce insidious errors or other difficult problems to identify Immediately and whose resolution requires a lot of time, especially for lower experience developers who approached work with a false sense of trust due to their AI dependence.
As a result, more than a third of the developers surveyed claim that some of their visits to Stack Overflow are due to Ia -related problems. That is, the code suggestions that accepted a LLM -based tool They generated problems that then had to solve others.
Although important improvements have recently occurred through optimized models for reasoning, it is unlikely that this lack of reliability will disappear completely: it is endemic to the very nature of predictive technology. That is why 72 percent of the survey participants said that “vibration coding” is not part of their professional work; Some feel that it is too unreliable and that it can introduce difficult problems to purify that they are not appropriate for production.
Useful, but not totally reliable
Given all that skepticism and frustration, why do developers still use the tools? Well, in some cases, those responsible try to force them, but the most common is that the tools are used because they remain clearly useful: it is simply important to use them correctly and know their real scope.
It is important that individual managers and collaborators incorporate ia tools to the workflow together with solid training to guarantee a deep understanding of best practices, so that the tools are not incorrectly used in a way that Do not generate more problems than you solve or waste more time than you save.
Developers must distrust less elements such as co -surveyed self -reflected suggestions, considering them more as a starting point instead of simply pressing the tab and move on. Tools like these are more appropriate for a limited couples programming relationship: ask the LLM to find problems or suggest more elegant solutions that are considered critically, Not to suggest complete methods that are accepted without further ado.
They can also be useful for learning. The opportunity to be in constant learning, continuously familiarly with new languages, work frames or methodologies, is one of the attractions for some people, and the LLMs can reduce friction in that process when answering questions more specificly than with laborious searches in technical documentation often incomplete; Precisely the type of thing for which Stack Overflow has been used historically.
“Although we have observed a decrease in traffic, it is not as drastic as some suggest”said Jody Bailey, product director and technology at Stack Overflow, in a comment to Venturebeat. Stack Overflow plans to allocate part of its resources both to expand knowledge about AI tools and to encourage community debates that help solve specific problems of workflows that involve these tools.
IA coding tools are useful and will be even more in the future when artificial intelligence technologies improve. But for now, The trust in them is not up to the huge investment and marketing. IA tools promise productivity improvements, but in reality they could generate new categories of technical debt. The “almost perfect” code generated requires a significant intervention of the developer to be ready for production. This creates a particularly insidious problem, according to the Stack Overflow survey.
