How Developers Interact with AI: A Taxonomy of Human-AI Collaboration in Software Engineering (2501.08774v2)

Abstract: AI, including LLMs and generative AI, is emerging as a significant force in software development, offering developers powerful tools that span the entire development lifecycle. Although software engineering research has extensively studied AI tools in software development, the specific types of interactions between developers and these AI-powered tools have only recently begun to receive attention. Understanding and improving these interactions has the potential to enhance productivity, trust, and efficiency in AI-driven workflows. In this paper, we propose a taxonomy of interaction types between developers and AI tools, identifying eleven distinct interaction types, such as auto-complete code suggestions, command-driven actions, and conversational assistance. Building on this taxonomy, we outline a research agenda focused on optimizing AI interactions, improving developer control, and addressing trust and usability challenges in AI-assisted development. By establishing a structured foundation for studying developer-AI interactions, this paper aims to stimulate research on creating more effective, adaptive AI tools for software development.

• The paper introduces an eleven-type taxonomy that categorizes diverse interactions between developers and AI tools in software engineering.
• It systematically analyzes methods such as auto-complete, command-driven actions, conversational assistance, and contextual recommendations.
• The study outlines implications for improved AI tool design, greater developer trust, and targeted research on context-aware and ethical AI applications.

Taxonomy of Human-AI Collaboration in Software Engineering

The paper "How Developers Interact with AI: A Taxonomy of Human-AI Collaboration in Software Engineering" provides an intricate exploration of the interactions between software developers and AI tools, an area that is gaining increasing attention within the domain of software engineering. The authors, Christoph Treude and Marco A. Gerosa, propose a detailed taxonomy encompassing eleven distinct types of interactions developers have with AI tools. The proposed classification addresses a pressing need in the field for a structured framework to analyze and enhance developer-AI workflows.

Summary of Interaction Types

The taxonomy delineates interactions into categories such as auto-complete code suggestions, command-driven actions, conversational assistance, and contextual recommendations, among others. Each interaction type is characterized by initiation triggers, the AI response, developer reactions, and output generated, encapsulating a broad spectrum of developer-tool engagements:

• Auto-Complete Code Suggestions: Initiated by typing context, tools like GitHub Copilot provide real-time code recommendations, reducing repetitive coding effort.
• Command-Driven Actions: Developers use explicit inputs to command AI for specific outputs, enhancing direct control over AI functionalities.
• Conversational Assistance: This allows for natural language interaction, aiding developers with guidance and explanations in chat interfaces.
• Contextual Recommendations: AI tools employ contextual understanding to suggest improvements aligned with the developers’ current tasks and project structure.

This classification provides a comprehensive framework for understanding how AI tools support developers across various stages of software development, from code generation to project management and testing.

Research Agenda and Future Directions

The taxonomy introduced in this paper establishes a foundation for a dedicated research agenda aimed at optimizing human-AI collaboration in software engineering. The authors identify critical areas warranting investigation:

1. Effectiveness of Interaction Types: Evaluating which types lead to the most productive workflows, taking into account developer experience levels and task specificity.
2. Developer Trust and Adoption: Assessing how interaction types influence trust and how tool design can improve AI reliability perceptions.
3. Context-Aware Interactions: Exploring methods to enhance AI contextual awareness, resulting in more relevant suggestions that align with project-specific goals.
4. Balancing Automation with Control: Ensuring developers have the ability to customize AI interactions to prevent cognitive overload and maintain productivity.
5. Ethics and Bias: Investigating mechanisms to mitigate biases inherent in AI-generated suggestions and ensure inclusive development practices.

The outlined research agenda encourages the development of adaptive and efficient AI tools that align closely with the dynamic needs of software developers, enhancing productivity and trust without sacrificing transparency or control.

Implications

The implications of this research extend to both theoretical and practical dimensions of AI integration in software engineering. The taxonomy provides a conceptual framework that guides empirical studies and tool development efforts aimed at enhancing interaction design and usability. Moreover, the focus on trust, cognitive load, ethical concerns, and privacy from a developer's perspective underscores crucial areas for the ongoing evolution of AI tools in professional settings.

By systematically identifying interaction types and proposing a structured research path, this paper not only lays groundwork for future studies but also could aid developers in selecting and adapting AI tools best suited for their individual workflows. Future developments in this domain could see increased personalization and contextual understanding in AI-driven developer tools, fulfilling the dual objectives of heightened productivity and seamless integration into existing software development paradigms.

In conclusion, this paper serves as a critical scholarly contribution to understanding and refining AI's role in software engineering, pushing the boundaries of how developers engage with software-enhancing technologies.