Programmers are utilizing AI greater than ever, however they do not like or trust the instruments very a lot, based on the 2025 Stack Overflow Developer Survey.
The survey of just about 50,000 builders discovered that 84% now use or plan to make use of AI instruments of their workflow, up from final yr (76%). Over half {of professional} builders (51%) use these instruments each day.
Additionally: The perfect AI for coding in 2025 (and what to not use)
Such figures would possibly recommend that programmers should love AI. Nonetheless, solely 60% expressed constructive sentiment towards AI instruments, a proportion down from over 70% in each 2023 and 2024.
Mistrust is a defining theme of the survey. In 2024, 43% of builders felt good about AI accuracy, and solely 31% had been skeptical. By 2025, 33% of builders trusted AI software outputs, 46% expressed energetic mistrust, and a mere 3% mentioned they extremely trusted the outcomes. Amongst seasoned professionals, the “extremely trust” determine dropped to only 2.6%, with 20% reporting sturdy skepticism.
Additionally: Dangerous vibes: How an AI agent coded its technique to catastrophe
In brief, builders actually use AI, however trusting the know-how to get the job proper by itself is one other matter. And after IT chief Jason Lemkin’s expertise with a Vibe programming undertaking that disintegrated, taking with it a manufacturing database, who can blame them?
The Stack Overflow research additionally discovered that the most important single frustration, cited by 66% of builders, is coping with “AI options which can be virtually proper, however not fairly,” which regularly results in the second-biggest frustration, at 45%, being “Debugging AI-generated code is extra time-consuming.”
As Invoice Harding, CEO of Amplenote and GitClear, famous in GitClear’s AI Copilot Code Quality research, which analyzed 211 million traces of code, “builders trust the present technology of AI assistants about as a lot as we trusted the earlier technology, i.e., not a lot.”
It isn’t simply programmers who do not trust AI. A latest survey of greater than 1,100 Individuals revealed that solely 8.5% mentioned they “all the time trust” the data they obtain from Google’s AI Overviews, and 21% mentioned they’ve zero trust within the function’s capacity to floor correct info. A brand new KPMG research discovered that, worldwide, 66% of people use AI, but only 46% trust AI systems.
Additionally: Find out how to eliminate AI Overviews in Google Search: 4 simple methods
Junior builders who trust AI probably the most could possibly be making an enormous mistake. A preferred weblog put up by Namanyay Goel, an impartial developer, warned: “We’re trading deep understanding for quick fixes, and whereas it feels nice within the second, we’ll pay for this later.”
Worries over AI-created technical debt are rising. Harding warned that if firms proceed to measure developer productiveness by the simple-minded metric of variety of commits or traces of code written, AI-driven technical debt will spiral uncontrolled. “Leaders need to recognize that more code is often worse,” he mentioned, suggesting that copying and pasting code results in larger defect charges. Certainly, GitClear discovered a direct connection between the rising defect charge and AI adoption.
All that mentioned, Stack Overflow additionally revealed in its survey that OpenAI’s GPT fashions are the preferred giant language fashions, with 82% of builders who use AI indicating that they used them for improvement work previously yr. Anthropic’s Claude Sonnet fashions got here second, adopted by Google’s Gemini Flash.
Regardless of AI’s rise, with regards to built-in improvement environments (IDEs), programmers nonetheless desire Visible Studio (75%) and Visible Studio Code (29%) over AI-first programming IDEs. The old-school, easy code editors, Vim and Notepad++, stay widespread, even amongst programmers who use AI. That mentioned, Microsoft’s incorporation of Copilot into its instruments has confirmed to be a wise transfer.
Additionally: Coding with AI? My prime 5 ideas for vetting its output – and staying out of bother
As earlier than, JavaScript, HTML/CSS, and Python preserve their standing as probably the most extensively used programming languages. Python, and doubtless because of the recognition of Python-based generative AI libraries TensorFlow and PyTorch, is particularly wanted by builders adopting a brand new language. Nonetheless, Rust, with an 83% approval mark, stays probably the most admired language.
Trying forward, whereas AI is being adopted rapidly, builders are, if something, rising cautious about handing off important duties to brokers. A overwhelming majority (75%) mentioned human recommendation continues to be irreplaceable in situations the place they do not trust AI’s output.
As for AI brokers, they’ve but to achieve mainstream acceptance. Over half of the survey’s respondents use easier AI instruments, and 38% don’t have any plans to undertake brokers quickly.