Why do developers keep using AI tools they don't trust?

Stack Overflow's 2025 Developer Survey reports that AI tool adoption keeps climbing — 80% of developers now use AI tools in their workflows — while trust in the accuracy of that output has fallen sharply, from 40% in previous years to 29% this year, and positive favorability toward AI dropped from 72% to 60% year over year. The survey attributes the decline to a specific complaint: 45% of respondents name "AI solutions that are almost right, but not quite" as their top frustration, and 66% say they are spending more time fixing "almost-right" AI-generated code than before. When trust breaks down on a complicated or high-stakes problem, 75% say they would still ask another person rather than rely on the AI's answer.

The survey's own framing is that adoption and confidence have decoupled: developers keep reaching for AI tools because they're useful for routine work, but the "almost right" failure mode — code that looks correct but contains a subtle wrongness — imposes a verification tax that erodes trust faster than usage grows. This is also why the survey finds AI agents and "vibe coding" have not become mainstream professional practice: 52% say agents have affected how they work, but the primary effect is personal productivity, not delegation, and nearly 72% say generating whole applications from prompts isn't part of their professional work. The survey also reports a compensating behavior: Stack Overflow is becoming what the post calls a "human-verified source of truth" for AI-generated code, with about 35% of visits now driven by AI-related issues, and developers ranking "reading comments" as their top on-site activity — a preference for human-to-human verification over AI output.

The "almost right" complaint gives a user-facing name to the same gap that Do AI coding tools actually speed up experienced developers? measured directly in completion times: expected acceleration converts into unexpected verification overhead. It also parallels Does AI assistance erode the skills needed to oversee it?, where self-reported productivity gains coexist with an admitted inability to hand off most work unsupervised — this survey's 75% who still ask a person when AI can't be trusted is the same caution from the other side of the keyboard. The finding that AI agents remain supplementary rather than autonomous (52% affected, vibe coding not professional for 72%) matches How are national lab staff actually using generative AI?, another setting where real deployment sits inside a "copilot" envelope rather than full delegation.

The excerpt gives no total respondent count, no sampling method, and no breakdown by seniority, language, or region for the headline trust figures — this is Stack Overflow's own characterization of its annual community survey, a platform with a direct commercial interest in developers continuing to treat human-written answers as authoritative. The post also does not separate self-reported time spent fixing AI code from any measured baseline, so the 66%-more-time figure is a perception, not a timed comparison like the METR or Google trials ran. What the survey does establish, at the strength self-report allows, is that usage and trust are moving in opposite directions for developers broadly, and that the mechanism they name for the mistrust is correctness failures that are hard to catch rather than outright refusals or crashes.