The 2025 Stack Overflow Survey Found a Trust Gap AI Vendors Don't Talk About

Usage of AI coding tools among developers climbed to 84% in the 2025 Stack Overflow Developer Survey. Trust in the accuracy of what those same tools produce fell to 29%, down from 43% the year before. Those two numbers moving in opposite directions, at the same time, from the same 49,000-plus respondents, is the actual story in this year’s data. Adoption isn’t the open question anymore. Trust is.

The gap, in numbers

Adoption has climbed every year the survey has tracked it: 70% in 2023, 76% in 2024, 84% in 2025. Sentiment moved the other way, from 77% positive in 2023, to 72% in 2024, down to 60% in 2025. The trust-in-accuracy number, a narrower and blunter question than general sentiment, fell even harder: 40% in 2023, 43% in 2024, then a drop to 29% in 2025. Professional developers hold a more favorable overall view (61%) than people still learning to code (53%), which on its own is a useful data point: the group with the most hands-on experience using these tools professionally is also the group least starry-eyed about them.

This isn’t a story about AI tools getting worse. It’s a story about developers getting a clearer picture of where the tools actually help and where they don’t, after a year or more of daily use rather than a few weeks of novelty.

What’s actually generating the frustration

The survey asked developers directly what frustrates them about AI tools, and the answers cluster tightly around one theme. 66%, the largest share by a wide margin, pointed to output that’s “almost right, but not quite.” 45% said debugging AI-generated code takes more time than they expected it to.

Read together, those two numbers describe a specific and unglamorous failure mode. It’s not that the tools produce obviously broken code that fails a build immediately; that would be cheap to catch. It’s plausible-looking code that passes a casual read and fails somewhere subtler, a wrong assumption about an edge case, a slightly incorrect API usage, a test that looks like it covers the right thing but doesn’t. That kind of near-miss is more expensive to find than an outright failure, because it requires the same level of scrutiny as reviewing unfamiliar code from a human, and developers report they’re spending closer to that level of time, not less.

Python keeps winning, and it’s not close

Away from the AI-specific questions, Python’s usage climbed 7 percentage points year over year, the largest single-year jump for any major language in recent survey history. The driver isn’t mysterious: Python sits at the center of AI, data science, and increasingly backend development, which means its usage curve is riding three separate growth trends simultaneously rather than one.

For teams making a language choice for a new service, or scoping what to look for when they hire a Python developer, this kind of broad-based, multi-driver growth is a stronger signal than a language spiking because of one trend, since it suggests staying power rather than a temporary fad.

The IDE category has new top-tier entrants

Cursor and Claude Code both appeared in the survey’s IDE rankings for the first time this year and immediately landed in the top tier, alongside VS Code, which continues to dominate on raw usage. That’s a fast climb for tools that didn’t have enough of a user base to register a year or two ago, and it lines up with the adoption number: developers are reaching for AI-native tooling, not just AI features bolted onto existing editors.

On the model side, Anthropic’s Claude Sonnet came out as the most admired language model among respondents who evaluated it, with roughly 67.5% expressing admiration, and it ranked near the top on desirability too. That’s one data point among many in a survey covering 314 technologies, not a verdict on the entire model landscape, but it’s a notable result given how competitive the frontier model space has been this year.

What this actually means for how teams should operate

The practical takeaway isn’t “use AI tools less” or “use them more.” It’s to calibrate review discipline to the actual failure mode the survey describes. Code from an AI assistant deserves the same scrutiny you’d give a fast but unfamiliar contributor: someone who can produce a lot of plausible-looking work quickly, and who you haven’t yet built a track record with. That’s a different posture than either blind trust or blanket rejection, and it’s closer to what the survey data suggests developers are already doing instinctively, which may be exactly why trust is dropping even as usage climbs: people are using the tools more and learning their specific failure modes at the same time.

For teams building AI review practices into their workflow, that means treating “the code compiles and passes tests” as a floor, not a ceiling, and putting real weight on the same code review habits that caught bad human-written code before any of this tooling existed. The tools changed. The discipline that catches their mistakes shouldn’t have to.