Stack Overflow’s own 2025 Developer Survey, drawn from 49,009 responses across 166 countries, found that 84 percent of developers now use or plan to use AI tools in their workflow, up from 76 percent the year before. In the same survey, trust in the accuracy of that output fell to 29 percent, down from 40. More developers now actively distrust AI’s accuracy than trust it. That’s not a contradiction. It’s the actual shape of how the tools are being adopted.
The top complaint isn’t that AI can’t write code. It’s that it writes code that’s close enough to be tempting and wrong enough to be costly, what Stack Overflow’s own reporting calls “almost right.” Sixty six percent of developers named it their top frustration, and 45 percent said debugging AI generated code takes more time than writing it themselves would have. Almost right is a specific, expensive kind of wrong, because it looks finished right up until someone has to trust it.
That mismatch shows up clearly in how developers are actually using the tools day to day. Full autonomous agents, the kind that carry out multi step tasks with minimal supervision, see daily use from only 14.1 percent of developers, and 37.9 percent say they don’t plan to use them at all. The overwhelming preference is for AI that proposes code inside familiar tooling rather than acting independently across a system, copilot and autocomplete rather than agent. Developers want the suggestion, not the decision.
Google’s own DORA research backs this up from a different angle. Thirty nine percent of developers surveyed there reported little to no trust in AI generated code, and separate industry analysis has tied rising AI use to measurable increases in code duplication, the kind of technical debt that surfaces months after a project ships rather than on the day it’s written.
None of this is an argument against using the tools. It’s an argument for being specific about which parts of a job they’re actually good at. The adoption number tells you the tools are useful enough that almost everyone’s reaching for them. The trust number tells you nobody’s handing over judgement yet, and the survey data suggests they’re right not to. What’s actually sticking, based on how developers say they’re using these tools rather than how vendors say they should, is AI as a fast first draft inside an existing workflow, checked by someone who still understands the problem well enough to catch the almost right answer before it ships. The agencies dropping tools aren’t dropping AI itself, they’re dropping the ones that ask for more trust than the current generation of models has actually earned.
Almost right is a specific, expensive kind of wrong.
