The Impact of Generative AI Coding Assistants on Developers Who Are Visually Impaired
Claudia Flores-Saviaga, Benjamin V. Hanrahan, Kashif Imteyaz, Steven Clarke, Saiph Savage · 2025
Read The Impact of Generative AI Coding Assistants on Developers Who Are Visually Impaired by Claudia Flores-Saviaga, Benjamin V. Hanrahan, Kashif Imteyaz, Steven Clarke, Saiph Savage (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-16491.txt.
Introduction
Claudia Flores-Saviaga and four co-authors studied how blind and low-vision programmers use AI coding helpers in a 2025 paper. These tools have spread fast through software work, yet little was known about how they affect developers who cannot see the screen well. The team had such developers carry out coding tasks with one of the helpers.
The users found real gains and said the tools could change their work for the better. But the helpers also made some old access problems worse and added new ones. They offered too many suggestions at once, which led users to ask for a way to pause the tool. Moving between their own code and the tool's output was also hard. The authors urge design that starts from the user's activity, not just the tool's power. Access has to be built in, not added later.
Catalog record
| CREATOR | Claudia Flores-Saviaga, Benjamin V. Hanrahan, Kashif Imteyaz, Steven Clarke, Saiph Savage |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 80 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-16491
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
The Impact of Generative AI Coding Assistants on Developers Who Are Visually Impaired Claudia Flores-Saviaga, Benjamin V. Hanrahan, Kashif Imteyaz, Steven Clarke, Saiph Savage Abstract . The rapid adoption of generative AI in software development has impacted the industry, yet its effects on developers with visual impairments remain largely unexplored. To address this gap, we used an Activity Theory framework to examine how developers with visual impairments interact with AI coding assistants. For this purpose, we conducted a study where developers who are visually impaired completed a series of programming tasks using a generative AI coding assistant. We uncovered that, while participants found the AI assistant beneficial and reported significant advantages, they also highlighted accessibility challenges. Specifically, the AI coding assistant often exacerbated existing accessibility barriers and introduced new challenges. For example, it overwhelmed users with an excessive number of suggestions, leading developers who are visually impaired to express a desire for “AI timeouts.” Additionally, the generative AI coding assistant made it more difficult for developers to switch contexts between the AI-generated content and their own code. Despite these challenges, participants were optimistic about the potential of AI coding assistants to transform the coding experience for developers with visual impairments. Our findings emphasize the need to apply activity-centered design principles to generative AI assistants, ensuring they better align with user behaviors and address specific accessibility needs. This approach can enable the assistants to provide more intuitive, inclusive, and effective experiences, while also contributing to the broader goal of enhancing accessibility in software development. Figure 1. Overview of our research to analyze how developers who are visually impaired interact with AI coding assistants. 1. Introduction The integration of generative artificial intelligence (AI) into software development is fundamentally transforming coding practices ( Liang et al., 2024 ) . AI coding assistants, such as GitHub Copilot ( GitHub, 2023 ) , provide functionalities like auto-completion, code generation, and test generation ( Popa, 2023 ) , which have the potential to revolutionize how developers work ( Peng et al., 2023 ) . Amidst this surge of new capabilities driven by AI, an important question emerges: How do these generative AI coding assistants impact developers with visual impairments? The convergence of AI and accessibility in coding presents both tremendous opportunities and intricate challenges ( Liang et al., 2024 ; Nguyen et al., 2024 ) . On one hand, AI has the potential to make coding more accessible for developers with visual impairments by reducing manual coding tasks and providing intelligent assistance ( Shinohara et al., 2022 ) . On the other hand, if these AI coding assistants are not designed with accessibility in mind, they may inadvertently create new barriers or worsen existing ones ( Nguyen et al., 2024 ; Wang et al., 2024 ) . This risk is especially high if the AI coding assistants depend heavily on visual cues or employ interaction methods incompatible with screen readers ( Sharif et al., 2021 ; Morris et al., 2018 ; Zong et al., 2022 ) . In this context, it is important to understand that the relationship between accessibility and usability in AI assistants mirrors longstanding challenges in web accessibility ( Henry et al., 2014 ; Lazar et al., 2004 ) , particularly for visually impaired users ( Bigham et al., 2010 ; Kane et al., 2008 ) . Similar to rich internet applications ( Antonelli et al., 2019 ; Påhlstorp and Gwardak, 2007 ) , AI assistants often introduce dynamic content and complex interactions that traditional accessibility approaches may not fully address ( Kelly et al., 2005 ; Hou et al., 2024 ) . For instance, Petrie and Kheir ( Petrie and Kheir, 2007 ) found that existing accessibility guidelines fail to capture many of the usability problems encountered by visually impaired users when interacting with dynamic content. This gap could be especially exacerbated in AI coding assistants ( Mowar et al., 2024 ; Oswal and Oswal, 2024 ; Adnin and Das, 2024 ) , where AI-generated suggestions, real-time code generation, and sophisticated user interfaces can present new hurdles ( Hou et al., 2024 ; Jalil, 2023 ) . Just as screen readers struggle with dynamic web content, AJAX updates, and automatic refreshes ( Borodin et al., 2010 ) , developers who are visually impaired may experience similar difficulties when interacting with rapidly changing AI-generated code suggestions and interface elements ( Li et al., 2025 ) . Consequently, simply following basic accessibility rules might not be enough to make AI coding assistants truly easy to use. We might need to rethink how to make these assistants more accessible for everyone. The challenge is that we do not fully understand the difficulties and benefits that visually impaired developers experience when using AI coding assistants. Closing this knowledge gap is important—not just for accessibility, but also for making AI-assisted coding tools more user-friendly and effective for all developers, including those with different levels of vision. Our study addresses this research gap by investigating what unique challenges and opportunities AI-assisted coding tools present for developers with visual impairments. Our research is driven by the following research question: • RQ1: What challenges do AI-assisted coding tools pose for developers who are visually impaired, and what opportunities do they offer to empower and enhance their work? To study this question, we use the Activity Theory framework ( Kaptelinin and Nardi, 2009 ) , which helps analyze how tools influence human activity and identifies contradictions that occur when a tool’s design does not fit well with users’ workflows and needs ( Baldwin et al., 2020 ) . We applied this framework to examine a qualitative study we conducted with 10 visually impaired developers of varying experience levels. These developers used an AI coding assistant, GitHub Copilot, to complete a coding task. During the study, we observed their real-time interactions with the AI assistant, noting both challenges and opportunities. After the task, we conducted interviews to gain deeper insights into their experiences and the difficulties they faced while using the AI coding assistant. Our findings reveal a complex landscape where AI coding assistants can both improve coding efficiency and introduce new accessibility challenges. Based on our findings, we outline a roadmap for the next generation of accessible AI coding assistants. By analyzing the experiences of visually impaired developers using AI-assisted coding tools, we derive key design insights that can reshape how accessibility is approached in AI-driven software development. The contributions of this paper are twofold: • We conduct a comprehensive analysis of the accessibility challenges and benefits of AI coding assistants, using real-world insights from developers who are visually impaired. • We propose a set of design recommendations to make AI coding assistants more accessible and inclusive for developers with visual impairments. 2. Related Work 2.1. Accessibility and Tools for Developers who are Visually Impaired Accessibility in general has been a subject of ongoing research ( Bigham et al., 2017b ; Gleason et al., 2018 ; Gleason et al., 2017 ; Savage et al., 2022 ) . Several studies have shed light on the significant challenges developers who are visually impaired face with coding and the techniques these programmers employ to overcome these challenges. Mountapmbeme et al. ( Mountapmbeme et al.,