Developers share a “trust paradox” when using AI tools for coding, suggests Google’s new DORA 2025 “State of AI-assisted Software Development” research report. On the one hand, they find AI outputs useful and valuable. On the other hand, they stop short of fully trusting them.
The findings of the report, released on September 23, suggest AI is being integrated less as a replacement for human judgment and more as a supportive tool—accelerating workflows, boosting productivity, and improving efficiency, while developers retain the final say.
In this context, it’s important to note that generative AI (GenAI) is rewriting how software gets built, making English the “hottest programming language”, as Andrej Karpathy, AI researcher and former Tesla AI director, put it. He coined the term “vibe coding” to explain the shift where developers no longer labour over every line of code but instead describe what they want in plain English, and large language models (LLMs) generate the rest.
Vibe coding marks a leap from no-code and low-code platforms like Bubble and FlutterFlow, which let non-technical users drag and drop elements to build apps. Those frameworks opened up software creation but imposed limits on customization. Vibe coding removes that ceiling: AI produces raw code that developers can freely adapt. It also goes further than AI-assisted coding. Tools like GitHub Copilot speed up tasks by auto-completing functions, suggesting optimizations, or helping with debugging. But the process is still structured and developer-driven.
Vibe coding flips that hierarchy. Here, the developer sets the intent, and AI becomes an adaptive co-creator, generating, refining, and reworking code dynamically. Consider the case of a weather app. Traditionally, a developer sets up the React UI and writes functions to fetch data, while AI offers suggestions.
In vibe coding, a simple prompt—“Build a React weather app with a clean UI showing temperature and humidity”—produces a full project, complete with application programming interface (API) integration. Iteration is equally natural: “Add voice search” or “Change the background based on weather”. Even debugging is transformed. Instead of manually parsing error logs, a developer can say: “Fix this Python script so it runs 30% faster without errors.” The AI debugs, optimizes, and even refactors code automatically.
AI-assisted coding, thus, accelerates human workflows while vibe coding reframes them, resulting in faster builds. But is the code generated thus, accurate, and fit for project if left unsupervised?
Google’s DORA report insists that AI is a transformative tool for developers, but realizing its full potential requires more than just adoption. It demands that organizations evolve their culture, processes and systems to support a new era of software development. According to Google, the “trust but verify” approach when using AI tools, is a sign of mature adoption.
AI adoption among software professionals has surged to 90%, up 14% from last year, according to DORA 2025. Developers, product managers, and others now weave AI into their daily workflows, typically spending about two hours a day with these tools.
The reliance is significant: 65% report using AI heavily, with 37% leaning on it a “moderate amount,” 20% “a lot,” and 8% “a great deal.” The benefits are clear. Over 80% of respondents say AI has boosted their productivity, while 59% credit it with improving code quality.
Google has also introduced a new blueprint, the DORA AI Capabilities Model, comprising seven essential capabilities to amplify AI’s impact. These include drafting a clear AI policy, a healthy data ecosystem, a quality internal platform, and having a user-centric focus.
