Abstract
This undergraduate thesis portfolio brings together two projects united by a shared
concern about the persistent exclusion of people with disabilities from full participation in the
digital world. Despite decades of legal mandates and technical standards, millions of Americans
with visual and motor disabilities cannot independently navigate websites that connect them to
essential goods, services, and civic participation. The technical capstone project, Dream,
confronts this problem from the user side by building an open-source, AI-powered browser that
automates complex web tasks through natural language and voice commands, enabling users to
accomplish goals like "order my grocery list on Instacart" without relying on inaccessible
interfaces. The STS research paper confronts the same problem from a structural angle, asking
why developer-side compliance with web accessibility law has failed so comprehensively across
three decades despite clear legal mandates. Together, these projects form a sociotechnical
argument. The tools to make the web accessible already exist, and yet structural forces
consistently prevent them from being applied. Understanding both the technical possibilities and
the structural barriers is necessary for making progress on either front.
Dream is an open-source, privacy-first AI browser built on the Chromium foundation and
specifically designed to improve web accessibility for users with visual and motor disabilities.
The core insight motivating Dream is that user-side automation can bypass the compliance
failures of developer-side accessibility standards. Rather than waiting for organizations to fix
their inaccessible websites, Dream navigates those websites on the user's behalf. Users interact
with Dream through a conversational side panel or entirely through voice commands, describing
multi-step tasks in natural language. Dream then plans and executes those tasks autonomously
using large language model computer-use capabilities. Unlike proprietary alternatives such as
OpenAI's ChatGPT Atlas or Perplexity's Comet, Dream runs locally, keeping personal data on
the user's own machine and avoiding the security vulnerabilities that have been documented in
cloud-based AI browser products. Dream also avoids the key limitation of
ARIA-label-dependent approaches by using visual recognition methods that analyze what is
actually displayed on screen, making it functional even on websites that have not implemented
proper semantic markup. Critical safety features include website blacklisting for sensitive pages,
emergency stop controls accessible via voice, and configurable autonomy levels ranging from
planning assistance to fully autonomous execution. The project's success was evaluated through
a two-week study with ten participants representative of the target population, measuring task
completion rates without keyboard or mouse input, time-on-task improvements, and daily
adoption patterns.
The STS research paper, "The Social Construction of Inaccessible Websites: Power
Asymmetries and the Failure of WCAG Compliance," applies Klein and Kleinman's structural
Social Construction of Technology framework to explain why 95.9% of the most-visited
websites still contain detectable accessibility violations despite the Americans with Disabilities
Act having mandated equal digital access since 1990. The paper demonstrates that web
inaccessibility is not a technical problem, as the Web Content Accessibility Guidelines have
provided comprehensive standards since 1999, but a structural one produced by four interlocking
mechanisms. First, rules of access are too fragmented to provide users with disabilities reliable
legal standing. Congress has never issued binding website accessibility regulations, the
Department of Justice withdrew its only proposed rulemaking in 2017, and federal courts remain
split on whether the ADA applies to websites without a nexus to a physical location. Second,
resource asymmetries concentrate economic and cultural power among developers and paying
clients, while users with disabilities, dispersed across sixteen percent of the global population,
have no organized economic leverage to demand accessible design. Third, development practice
structurally deprioritizes accessibility. A systematic review of 83 studies on software task
prioritization found that accessibility-related bugs do not appear as a recognized priority category
anywhere in the literature. Fourth, rhetorical closure, achieved through quiet legal settlement and
regulatory silence, stabilizes the status quo in favor of minimal compliance without resolving the
underlying exclusion. Together, these mechanisms form a self-reinforcing system in which
inaccessibility is the predictable output of how the development ecosystem is organized, not the
result of any individual actor's deliberate choice to discriminate.
Pursuing both projects simultaneously sharpened my understanding of each. While
building Dream and working through the limitations of ARIA-label-dependent tools, the
arguments in the STS paper became significantly more concrete. The claim that accessibility is
structurally deprioritized in development practice was easier to accept as an analytical finding
once I had directly encountered it as a practical constraint. The influence also ran in the other
direction. Conducting the STS research pushed me to think more carefully about why certain
design decisions in Dream mattered beyond their immediate practical benefits. Understanding
how rhetorical closure allows companies to manage the appearance of compliance without
achieving it reframed the case for local execution and open-source transparency as a structural
response rather than simply a privacy preference. The most important insight from working on
both projects together is that building a technical workaround and analyzing the structural failure
that necessitated it are not independent activities. Dream exists because the compliance system
has failed, and the STS paper explains the mechanisms behind that failure. That connection
would have been much harder to see from either project alone.
