Programmers Who Use Screen Readers in the Vibe Coding Era: Adaptation, Empowerment, and New Accessibility Landscape
Nan Chen, Luna K. Qiu, Arran Zeyu Wang, Zilong Wang, Yuqing Yang · 2026 · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26) · doi:10.1145/3772318.3790726
Summary
This two-week, three-phase longitudinal study investigates how 16 blind and low-vision (BLV) programmers who rely on screen readers engage with advanced AI code assistants, specifically GitHub Copilot in Visual Studio Code. The study was motivated by the rapid shift from direct code authorship toward supervising AI-generated output — an interaction paradigm Karpathy dubbed 'vibe coding.' The authors designed an initial session (background interview, tailored Copilot tutorial, one-hour programming task, Likert-scale survey, and semi-structured interview), a two-week exploration phase where participants used Copilot in their regular work and maintained structured diaries, and a follow-up interview to capture how practices and perceptions had evolved. Participants completed one of four programming tasks (chat server, chat client, data analysis, calculator debugging) adapted from prior benchmarks and validated by professional sighted engineers to ensure Copilot alone could not solve them. Data from screencasts, chat logs, diary entries, and interviews were analysed using thematic analysis structured through Activity Theory, with dual-coder consensus and a paired t-test on Likert scores to measure shifts from pre- to post-exploration. Beyond empirical findings, the paper contributes design principles and concrete recommendations for accessible human-AI collaboration in programming.
Key findings
Copilot delivered broad empowerment: all 16 participants reported efficiency gains, 13 of 15 felt it improved their programming skills, and participants spent only 10.69 of 43.12 task minutes (25.4%) on manual coding, with the rest going to prompt engineering, reviewing, validation, and fixing. Code assistants bridged long-standing accessibility gaps, particularly for UI development — a task many BLV programmers had previously avoided. However, four persistent challenge areas emerged: (1) communicating intent, where keyboard-shortcut conflicts, arrow-key cycling through prompt history, and uncertainty about model selection produced friction; (2) reviewing AI output, where tracking changed code blocks, diff navigation, and hallucinated terminal output were recurring pain points; (3) managing multiple views, especially under Agent mode; and (4) maintaining situational awareness, where status notifications were too brief or ambiguous. Feature preferences shifted over two weeks: initial enthusiasm for fully-automated Agent mode gave way to safer Ask and Edit modes, with Ask becoming the most commonly preferred feature in the follow-up. Participants also weighed the Accessible View (simple but less expressive) against the structured Message List (richer but harder to navigate). Statistically significant improvement was observed for resistance to overreliance (p=0.02).
Relevance
This paper is directly relevant to anyone building, procuring, or teaching with AI-assisted development tools. It provides concrete evidence that 'vibe coding' workflows can substantially lower barriers for screen reader users — including domains historically excluded from BLV developers' work, such as front-end and UI — while also surfacing specific design failures (diff navigation, multi-view focus management, status notification quality, keyboard-shortcut collisions) that vendors can action immediately. The design recommendations — consistent shortcuts, structured and grouped message content, accessible change-tracking, Do-Not-Disturb notification modes, proactive clarification of user intent, and accessible learning resources — form a practical checklist for teams auditing AI code assistants against accessibility baselines. Limitations worth noting: the sample was majority male, China-based, and self-reported via diaries rather than instrumented telemetry, so generalisation across regions and across evolving Copilot releases should be treated with caution.
Tags: screen readers · blind and low vision · AI code assistants · GitHub Copilot · vibe coding · generative AI · human-AI interaction · developer tools · programming accessibility