90% of devs use AI coding tools. Only 3% highly trust them

a computer screen with a bunch of code on it

AI assistants write code faster than most developers can review it. The adoption numbers back that up: Google’s 2025 DORA report found that 90% of technology professionals use AI at work. The trust numbers tell a different story.

According to the 2025 Stack Overflow Developer Survey, 46% of developers distrust the accuracy of AI tools. Only 3% say they highly trust them. And the DORA report adds a specific sting: AI adoption still correlates with less stable software delivery, not more.

The Control Gap

Vaadin’s engineering team put a name to the problem: the AI control gap. When you use an assistant like GitHub Copilot, Cursor, Windsurf, or Claude, you control the prompt you write and the code you choose to accept. Everything in between, which libraries get pulled in, how the code is structured, what gets tested, is largely out of view.

For Java teams in particular, that gap carries compounding risk. Enterprise Java codebases have long dependency chains, compliance requirements, and type contracts that a general-purpose model has no reason to respect unless you build the constraints in.

The Structural Fix

The Vaadin post argues that Java teams get safer results when they put structure around the assistant rather than relying on prompt discipline alone. The practical levers: typed code that lets the compiler catch what the reviewer misses, current framework context fed to the model, fast test coverage via JUnit, and small reviewable changes rather than large accepted blocks.

Tools like CodeRabbit and MCP integrations are mentioned as part of the stack for teams trying to close the gap systematically.

The framing from Vaadin CEO Jurka Rahikkala is that the answer is control by construction: one language, one codebase, one compiler. Speed without the invisible risk.

Stay on top of AI & Automation with BizStack Newsletter
BizStack  —  Entrepreneur’s Business Stack
Logo