Study reveals how developers converse with AI coding assistants in real-world projects

Researchers have completed the first large-scale empirical examination of how developers actually use conversational AI tools within their integrated development environments—revealing substantial shifts in how programming work gets organized and executed.

The Study's Scope#

The analysis drew from 74,998 developer messages across 11,579 chat sessions, spanning 1,300 repositories and 899 developers who used Cursor and GitHub Copilot in real-world conditions. Crucially, the chats examined were committed to public repositories as part of normal development workflows, capturing authentic behavior rather than controlled laboratory settings. This scale and methodology represent a departure from prior research that relied on small samples or studied general-purpose chatbots disconnected from project context.

Three Shifts in How Coding Happens#

The findings identify three major changes in the structure of development work:

Progressive specification: Rather than articulating complete task requirements upfront, developers engage in iterative back-and-forth exchanges with their AI assistants. They refine and adjust outputs gradually, treating the conversation itself as a mechanism for progressively clarifying what they need.

Cognitive redistribution: Developers are delegating substantial intellectual work to AI—not just code generation, but diagnosis of problems, comprehension of existing code, and validation of results. This offloads tasks that developers might traditionally handle themselves through direct code inspection and testing.

Active collaboration management: Rather than passively receiving AI outputs, developers strategically manage their partnership by externalizing plans into shareable artifacts, controlling what information the AI sees through context injection, and imposing constraints on AI autonomy.

Why It Matters#

The emergence of codebase-aware, context-sensitive AI assistants embedded directly in development workflows represents a genuine departure from earlier tool paradigms. Understanding how developers actually work with these systems—their strategies, trade-offs, and behavioral patterns—is essential for designing the next generation of programming environments that can support this human-AI partnership effectively.