Programmers Who Use Screen Readers in the Vibe Coding Era: Adaptation, Empowerment, and New Accessibility Landscape

Paper Title

Programmers Who Use Screen Readers in the Vibe Coding Era: Adaptation, Empowerment, and New Accessibility Landscape

Publication Info

• Topic area: Accessibility and human-AI interaction in programming for screen reader users.
• Keywords: Screen readers, AI code assistants, vibe coding, accessibility, human-AI collaboration, programming tools, GitHub Copilot, longitudinal study, inclusive design, automation.

Background and Problem

• Problem / challenge: Advanced AI code assistants like GitHub Copilot are reshaping programming workflows, but their accessibility and usability for screen reader (SR) users remain underexplored. Challenges include conveying intent, reviewing AI outputs, managing multiple views, and maintaining situational awareness.
• Significance: Addressing these challenges is critical to ensuring that SR programmers can benefit from AI-driven tools, which have the potential to bridge accessibility gaps and enhance productivity.
• Motivation and related work: Prior studies have shown that AI tools can improve productivity but introduce barriers for SR users, such as inaccessible interfaces and increased cognitive load. Existing research has not comprehensively examined how SR programmers interact with advanced AI features or how these tools impact their workflows.

Solution

• Proposed approach: A two-week, three-phase longitudinal study investigating how SR programmers use GitHub Copilot, focusing on empowerment, challenges, and design recommendations for accessible AI-assisted coding.
• Novelty:
  1. Longitudinal investigation of SR programmers’ engagement with advanced AI code assistants.
  2. Identification of specific challenges in communication, review workflows, and situational awareness.
  3. Evidence-based design principles and recommendations for accessible human-AI collaboration.
  4. Insights into evolving user preferences for automation versus control in AI tools.
• Procedure and key techniques:
  • Phase 1: Initial study with 16 participants, including a tutorial, programming task, and semi-structured interview.
  • Phase 2: Two-week exploration of Copilot in real-world programming, with diary entries documenting experiences.
  • Phase 3: Follow-up interviews to reflect on usage patterns and challenges.
  • Analysis: Thematic coding of interactions, Likert-scale ratings, and paired t-tests to assess changes in perceptions.

Results

• Concrete findings:
  • Participants spent 43.12 minutes on programming tasks, with only 25.41% on manual coding, indicating a shift toward guiding and reviewing AI outputs.
  • Copilot improved efficiency and accessibility, enabling tasks like UI development and code comprehension.
  • Challenges included difficulty in crafting effective prompts, reviewing AI outputs, managing multi-view interfaces, and maintaining situational awareness.
  • Participants’ preferences shifted from highly automated modes (e.g., Agent mode) to safer, more controlled modes (e.g., Ask mode).
• Advantage over baselines: Copilot bridged accessibility gaps in tasks like UI development and interpreting visual content, areas traditionally challenging for SR users.
• Experiments / evaluation:
  • Participants: 16 SR programmers with diverse experience levels.
  • Tools: GitHub Copilot in Visual Studio Code.
  • Metrics: Time spent on activities, Likert-scale ratings, and qualitative feedback.
• Limitations and future work:
  • Limited sample size (16 participants) and regional focus (China).
  • Gender imbalance (2 female participants).
  • Reliance on self-reported data and lack of systematic long-term tracking.
  • Future work should explore more diverse user groups, investigate individual differences, and address learning barriers for SR users.

Summary

This study investigates how screen reader users engage with advanced AI code assistants like GitHub Copilot, revealing both empowering benefits and persistent challenges. While AI tools enhanced efficiency, accessibility, and learning, users faced difficulties in communication, review, and maintaining control, particularly with highly automated features. Participants’ preferences evolved toward safer, more controlled interactions, highlighting the need for transparent and customizable designs. The paper proposes actionable recommendations to improve accessibility and inclusivity in human-AI collaboration, paving the way for more equitable participation in the emerging era of vibe coding.