One of the fields in which artificial intelligence tools are having a greater impact is on the programming. Companies see an opportunity to theory, accelerate work while reducing costsand they are integrating them into their work dynamics. Satya NadellaCEO of Microsofthe already said in 2023 that Copilot had become ‘an integral part of the day -to -day developers’ of the company.
However, as time passes, they also become more evident His limitations against work done by humans. If last month an American laboratory study Metr With developers he concluded that using these tools, even if they expedite some parts of the programming, ends up costing more time (19 % more) than savenow one Survey to 49,000 professional developers points out that His confidence in the generative AI to program is being reducedreports Venture Beat.
The study has been carried out by Stack Overflowa popular question and answering platform specialized in programming and development of software that operates since 2008. The results reveal that Four out of five developers use AI tools in their daily work in 2025a figure that has grown rapidly in recent years. However, The confidence in the accuracy of AI has decreased from 40 % in years before 29 % in 2025, 27.5% less.
This difference between use and trust illustrates the impact they are having on the profession like Github co -ilot either Cursor. Few developers discuss that these solutions have utility, but they are still defining What are your best applications and its limits.
Almost correct solutions, but not quite
To the question of what is its greatest frustration with the tools of AI, 45 % pointed out the ‘almost correct solutions, but not quite’ as the main problem. And, unlike obvious errors, those almost correct answers can introduce unwanted failures or behaviors difficult to detect and expensive to debugespecially for Junior developers who trust AI too much.
As a consequence, more than a third of respondents claim that His visits to Stack Overflow are due to problems caused by AI. That is, they accepted code suggestions generated by an AI and then had to resort to others to solve the resulting problems.
Although more recent models have improved their abilities, that ‘almost’ reliability but not at all will continue to be A problem inherent in how this predictive technology works.
Against Vibe Coding
That’s why, 72 % of participants claim that ‘Vibe Coding’ is not part of their professional work. This is an upward trend driven by AI tools, in which replaces the deep understanding of the code with the intuition and sensations of the programmer When working. It is an ‘looks ok’ instead of ‘I have verified that everything is fine’. This practice is considered unreliable by many programmers and prone to generate difficult mistakes to trackinappropriate for professional environments.
Stack Overflow Consider that developers should trust automatic suggestions such as Copilot, and see them as starting point rather than a definitive solution. These tools are more effective as limited programming assistantscapable of indicating errors or suggesting solutions that the developer must then assessnot as generators of complete methods that are accepted without review.