Source-linked AI summary
How Developers Interact with AI: A Taxonomy of Human-AI Collaboration in Software Engineering
Christoph Treude, Marco A. Gerosa
TL;DR
Research has paid limited attention to the specific interaction types between developers and AI tools, despite their diversity and relevance to software development. This paper proposes an 11-type taxonomy and research agenda; it presents the taxonomy as a structured foundation for studying and designing developer-AI interactions.
Problem
Research has given limited attention to how developers interact with AI-powered tools, while the diversity of interaction types complicates their design and evaluation.
Source: §I · p. 1Method
The paper proposes a taxonomy of 11 developer-AI interaction types and outlines a research agenda addressing interaction styles, developer control, trust, and usability.
Source: §I · p. 1, §II · p. 1Results
The paper presents a taxonomy of 11 interaction types and a research agenda for further study of developer-AI interactions.
Source: §II · p. 1, §I · p. 1Takeaways & Limitations
The taxonomy offers a structured basis for empirical study, systematic tool evaluation, and development of AI assistance features.
Source: §I · p. 1
Abstract
from arXiv · showhide
Artificial intelligence (AI), including large language models 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.
I. INTRODUCTION
Research on AI in software development has focused less on how developers interact with AI tools than on AI tools themselves. The paper proposes a taxonomy and research agenda to give these interactions a structured basis for study and design.
I. INTRODUCTION
- Research on AI-powered tools has given limited attention to the specific ways developers interact with them.Prior work on traditional developer tools has established the importance of usability and integration.Source: §I · p. 1
- Interaction diversity complicates design and evaluation, while the lack of a shared framework leaves researchers and designers without common vocabulary for developer-AI interfaces.The paper argues that a taxonomy can support empirical studies, systematic tool evaluation, and development of AI assistance features.Source: §I · p. 1
- The paper proposes a taxonomy of developer-AI interaction types and a research agenda addressing interaction styles, developer control, trust, and usability.The paper presents the taxonomy as a foundation for further research and discussion on more effective, adaptive AI tools.Source: §I · p. 1
II. INTERACTION TYPES
The section organizes developer-AI interaction into 11 types, characterized by how each interaction begins, what the AI produces, how developers respond, and the resulting output. The types span automatic suggestions, deliberate requests, and workflow-triggered assistance.
II. INTERACTION TYPES
- Table I characterizes 11 interaction types by trigger, AI response, developer response, output, and concrete examples.Triggers range from typing context to explicit commands or workflow events; outputs include code, documentation, and quality reports.Source: §II · p. 1, Table I · p. 2
- Typing-triggered auto-complete supplies real-time code suggestions, while contextual and file-aware recommendations use broader project cues or the active file type.Contextual recommendations address project- or file-level guidance; file-aware suggestions are tailored to the file or directory being created or edited.Source: §II · pp. 2–3
- Explicit instructions, conversational exchanges, selected-code assistance, deliberate UI actions, and comment-embedded prompts let developers request targeted outputs in different ways.Developers can review, refine, adjust, or verify the resulting suggestions and generated material.Source: §II · pp. 2–3
- Workflow automation can start at commits or pull requests, use keyboard shortcuts, or respond to API calls and webhooks with reports or other outputs.Event-based triggers are tied to workflow milestones; automated API responses operate through system-to-system calls or webhooks.Source: §II · p. 3
III. RESEARCH AGENDA
The research agenda builds on the interaction taxonomy to explore ways to tailor and optimize AI assistance. It considers distinct questions raised by passive and active interactions, including control, cognitive load, trust, usability, and adaptability.
Effectiveness of Interaction Types
- Comparative studies could test how interaction types affect productivity, code quality, and task completion across activities and developer experience levels.The proposed measures include completion time, accuracy, and satisfaction.Source: §III · pp. 3–4
Developer Trust and Adoption
- Transparency, reliability, and control may shape developer trust, especially for high-stakes tasks such as security scans and AI code modifications.Surveys, interviews, and behavioral experiments tracking acceptance or override rates are proposed to study trust.Source: §III · p. 4
Context-aware AI Interactions
- Project structure, dependencies, design patterns, and past interactions are candidate context signals for making AI recommendations more relevant and actionable.The agenda proposes comparing acceptance rates for context-aware models and generic suggestions.Source: §III · p. 4
Optimizing for Developer Control and Customization
- Customizable suggestion frequency or intrusiveness could balance automation with developer control without adding unnecessary complexity.The agenda proposes testing customizable interfaces and surveying developer preferences.Source: §III · p. 4
Reducing Cognitive Load
- Frequent or intrusive AI interactions can disrupt focus, so studies could compare interaction types using task switching, eye tracking, and mental-load assessments.The agenda asks which interaction best balances productivity and reduced cognitive burden.Source: §III · p. 4
Ethics and Bias in AI Interactions
- Bias research could examine outputs across diverse developers and codebases, with developer issue reports helping refine AI behavior over time.AI tools may reinforce stereotypes or biases present in their training data.Source: §III · p. 4
Privacy and Data Protection
- Because tools may send code snippets or project metadata to external servers, research could assess privacy-preserving approaches and their effectiveness–privacy trade-offs.Federated learning and on-device processing are proposed approaches.Source: §III · p. 4
Hallucination and Damage Control
- Research could identify causes of hallucinations and evaluate feedback loops, confidence scores, or explicit review before high-risk changes to limit potential damage.The agenda frames incorrect or misleading outputs integrated into critical systems as a significant challenge.Source: §III · p. 4
IV. CONCLUSION
The conclusion presents the interaction taxonomy as a framework for understanding developer-AI engagement and guiding research and tool development. Its validation remains future work, including real-world observation and systematic evaluation of existing tools.
IV. CONCLUSION
- The taxonomy offers a structured framework for understanding how developers engage with AI-powered software engineering tools.The paper suggests it could guide research and tool development, including orchestration of interaction types across workflows.Source: §IV · p. 4
- The accompanying research agenda identifies trust, cognitive load, and customization as areas warranting further exploration.Source: §IV · p. 4
- Validation is outside the paper’s scope; proposed next steps include observing real-world workflows and evaluating existing tools’ interaction-type usage.The taxonomy could also be refined by studying how interaction types align with developer needs in specific contexts.Source: §IV · p. 4