The Evolution of Information Seeking in Software Development: Understanding the Role and Impact of AI Assistants

1 Introduction

• Information seeking can contribute to or deter developer productivity, and developers’ ability to meet information needs contributes to building expertise.The study examines these concerns in the context of AI-assisted information seeking.Source: §1 · p. 1
• AI-assisted information seeking is understudied relative to AI use for understanding, generating, or modifying code, motivating questions about when, why, and how developers seek information and its impacts.The questions also address effects on productivity and skill development.Source: §1 · p. 1
• A survey of 128 developers, followed by 17 interviews, found AI tools can increase efficiency when seeking information, with caveats including avoiding over-reliance and building expertise to use AI effectively.Source: §1 · p. 1

2.1 Information Needs of Developers

• Earlier studies examined developer information needs through workplace observation, interviews, surveys, and focused analyses of bug reports, APIs, and release decisions.Examples include Ko et al., Fritz et al., and LaToza et al..Source: §2.1 · pp. 1–2

2.2 AI Tool Use in Software Engineering

• Prior AI-tool research addressed adoption, interaction modes, trust, usability, and productivity, often in the context of programming assistants and code completion.Examples include Barke et al. on acceleration and exploration modes and Bird et al. on code validation workload.Source: §2.2 · p. 2
• This study extends prior work by examining how AI interactions shape information-seeking behavior as well as task completion, productivity, and skill development.Source: §2.2 · p. 2

3.1 The Survey

• A 20-minute survey asked about developer backgrounds, AI-tool use, information seeking with and without AI, task impacts, and AI’s role in learning new technologies.The survey was piloted with three respondents, whose data were excluded.Source: §3.1 · p. 2
• After filtering incomplete, very rapid, or irrelevant responses, the researchers analyzed 128 valid responses using descriptive statistics, corrected Fisher’s exact tests, frequencies, and thematic summaries.Experience levels were grouped into five categories; open-ended responses received thematic summaries because of the small number of valid responses.Source: §3.1 · pp. 2–3

3.2 The Interviews

• Interviews explored participants’ AI-assisted information seeking, productivity, skill development, and work practices; analysis used iterative coding, theoretical saturation, and an external audit.All four authors independently coded an initial transcript, then agreed on an expanded codebook before coding the remaining transcripts.Source: §3.2 · pp. 3–4

4.1 Information Seeking with AI tools (RQ1)

• Most developers reported using AI for information seeking at least half the time, especially for best practices, new libraries, implementation trade-offs, and code or documentation questions.Participants also used AI to identify search keywords and validate information.Source: §4.1 · p. 4
• AI answers can be incomplete, inconsistent, or poorly calibrated in detail, so participants cross-check them and turn to documentation, forums, or colleagues when needed.Participants described validation demands as partly offsetting efficiency gains; traditional sources were also preferred for specialized, context-specific, or GUI-related tasks.Source: §4.1 · pp. 4–5

4.2 Impact on Perceived Productivity (RQ2)

• Participants described productivity through time saved, output quantity, and solution quality, while reporting faster task completion and possible costs to team knowledge-sharing and collaboration.AI assistance was also reported to help parse documentation, identify edge cases, and support robust solutions; some participants described reduced face-to-face interaction and organic discussion.Source: §4.2 · pp. 5–6

4.3 Impact on Skill Development (RQ3)

• Participants said AI can fill knowledge gaps, provide practice, reinforce concepts, and expose developers to alternative approaches that support learning and creativity.Some participants described solving practice problems before comparing their answers with AI-generated solutions.Source: §4.3 · pp. 6–7
• Participants also raised risks of overreliance, knowledge gaps, weaker independent problem-solving, and superficial understanding, particularly when using AI to integrate unfamiliar technologies.They noted that question-bounded answers may omit broader context, and that AI responses may lack sufficient pedagogical detail for learning.Source: §4.3 · pp. 6–7

5.1 The Evolution of Developer Information Seeking

• Developers shift toward task-oriented, on-demand AI learning and reactive implementation, while individually validating generated solutions shifts verification from community to developer.The individual verification burden partly offsets efficiency gains.Source: §5.1 · p. 7

5.2 Developer Productivity and Learning Trade-offs

• Participants report faster task completion and less blocked time alongside potential losses in problem solving, deeper understanding, confidence, and coherent mental models.Fragmented AI-assisted learning may impede integrating low-level knowledge into concepts transferable to other scenarios.Source: §5.2 · pp. 7–8

5.3 Directions for AI-Assisted Developer Tools

• Retrieval-augmented systems could ground AI assistance in project documentation, code history, forums, and specifications, connecting knowledge and supporting validation through trusted sources and provenance.Source: §5.3 · p. 8
• Adaptive systems could adjust assistance depth and style as developers progress from novice to expert, offering fuller explanations to newcomers and advanced suggestions to experienced developers.Source: §5.3 · p. 8
• Task-specific models could be fine-tuned for software engineering work such as code review, architectural decisions, security analysis, and learning assistance.Source: §5.3 · p. 8
• AI tools could support team and community knowledge sharing by documenting assisted solutions, sharing verified responses, and building team-specific knowledge bases.The proposed bases combine traditional documentation with validated AI-generated content.Source: §5.3 · p. 8

• Survey and interview self-reports may be affected by memory bias, which the researchers sought to mitigate by asking participants to ground responses in past experiences.Source: §6 · p. 8
• Recruitment through online platforms and mailing lists may not represent developers outside those networks; the study therefore aims for transferability rather than generalizability.Local community engagement and snowball sampling supplemented recruitment.Source: §6 · p. 8
• Participants could misinterpret survey or interview questions, so the researchers piloted and refined protocols and provided definitions for key terms.Source: §6 · p. 8

• The study explores how developer information-seeking behavior has evolved with AI-assisted tools, using survey and interview data.Source: §7 · p. 8
• It reports what information developers seek with AI, the challenges they encounter, and perceived effects on productivity and skill development.Source: §7 · p. 8
• The findings emphasize foundational knowledge, potential productivity gains, and practices for using AI tools as learning aids in software development.Source: §7 · p. 8