“I look at it as the king of knowledge”: How Blind People Use and

Understand Generative AI Tools

Rudaiba Adnin Maitraye Das
Northeastern University Northeastern University
Boston, Massachusetts, USA Boston, Massachusetts, USA
adnin.r@northeastern.edu ma.das@northeastern.edu
ABSTRACT promises of GenAI for people with disabilities [37, 38], we know
The proliferation of Generative Artificial Intelligence (GenAI) tools considerably less about how blind people use GenAI tools in their
has brought a critical shift in how people approach information regular work and how they make sense of these tools for their needs.
retrieval and content creation in diverse contexts. Yet, we have At this critical juncture of AI-driven world, understanding the use
limited understanding of how blind people use and make sense of GenAI among blind individuals—a large group that still remains
of GenAI systems. To bridge this gap, we report findings from in- underrepresented in many professions and higher education [7,
terviews with 19 blind individuals who incorporate mainstream 72]—is imperative to ensure they receive equitable opportunities to
GenAI tools like ChatGPT and Be My AI in their everyday practices. leverage the benefits of these emerging technologies.
Our findings reveal how blind users navigate accessibility issues, While the prospects of GenAI are immense, so are the potential
inaccuracies, hallucinations, and idiosyncracies associated with harms it can perpetuate, especially for people who are unaware of
GenAI and develop interesting (but often flawed) mental models of the risks associated with these technologies [13, 95]. An important
how these tools work. We discuss key considerations for rethinking way to mitigate these harms is to improve people’s mental models
access and information verification in GenAI tools, unpacking er- of GenAI [98] so that they can question its capabilities and limita-
roneous mental models among blind users, and reconciling harms tions and accordingly decide when and how to use these tools [55].
and benefits of GenAI from an accessibility perspective. Given the complex and opaque nature of GenAI [24, 55] and in the
absence of technical know-how, non-expert users run the risk of
CCS CONCEPTS developing erroneous mental models and unrealistic expectations
that are not consistent with the actual functionalities of these tools
• Human-centered computing → Empirical studies in Acces-
[98]. Although these risks apply to all users, the effect could be mag-
sibility.
nified for blind people, since an incorrect understanding of GenAI
KEYWORDS may reinforce and amplify the accessibility challenges [19, 52, 89],
misinformation propagation [81], undue trust in unverified infor-
Accessibility, blind, visual impairment, Generative AI, ChatGPT mation [61], ableist biases [25, 60] and other technology-related
ACM Reference Format: harms [18, 96] blind people already encounter. Thus, to better un-
Rudaiba Adnin and Maitraye Das. 2024. “I look at it as the king of knowl- derstand GenAI accessibility for blind people, we investigate: How
edge”: How Blind People Use and Understand Generative AI Tools. In The do blind people use GenAI tools and for what purposes? How do they
26th International ACM SIGACCESS Conference on Computers and Accessi- navigate challenges and biases, if at all, while using GenAI? What
bility (ASSETS ’24), October 27–30, 2024, St. John’s, NL, Canada. ACM, New
mental models do they develop to make sense of GenAI tools?
York, NY, USA, 14 pages. https://doi.org/10.1145/3663548.3675631
To this end, we present findings from interviews with 19 blind
1 INTRODUCTION individuals who have experience with GenAI chatbots such as Chat-
GPT, Copilot, Gemini, and Claude and GenAI-powered image de-
The surge in Generative Artificial Intelligence (GenAI) tools reflects scription tool Be My AI (a feature of the Be My Eyes app). Our analy-
a broader paradigm shift in workflow and productivity. Nowadays, sis shows that blind individuals use GenAI tools for various content
people are incorporating GenAI tools (e.g., ChatGPT [73], Google creation and information retrieval tasks, while navigating critical
Gemini [33], Microsoft Copilot [65], and Claude [5]) into a wide accessibility issues on GenAI interfaces and working through the
variety of domains, including education [36, 62, 83], programming inaccuracies, hallucinations, and idiosyncracies of GenAI responses.
[16, 54, 92], communication [14], and content creation [45, 56]. Blind We also detail the ways in which blind individuals form—at times
people are no exception to this. Accessibility technologies such as flawed and oversimplified—mental models of GenAI tools. Finally,
Be My Eyes and Envision have incorporated GenAI capabilities to we highlight how blind users grapple with concerns about ableist
assist blind users by answering visual questions [2, 87]. However, biases and other harms perpetuated by GenAI tools against the
despite significant commercial and public attention towards the benefits they receive from using these tools.
Permission to make digital or hard copies of part or all of this work for personal or Overall, our paper makes three key contributions. First, we
classroom use is granted without fee provided that copies are not made or distributed present rich empirical understandings of how blind people integrate
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored. GenAI tools to enhance productivity and access to information in
For all other uses, contact the owner/author(s). their regular work practices, extending prior research that inves-
ASSETS ’24, October 27–30, 2024, St. John’s, NL, Canada tigated disability representation and biases in GenAI [25, 29, 60]
© 2024 Copyright held by the owner/author(s).
ACM ISBN 979-8-4007-0677-6/24/10.
and potentials of GenAI to support access needs of people with
https://doi.org/10.1145/3663548.3675631
ASSETS ’24, October 27–30, 2024, St. John’s, NL, Canada Rudaiba Adnin and Maitraye Das

neurodivergence [32, 46, 86] and other disabilities like chronic ill- can save time and reduce physical and cognitive effort during com-
ness and aphantasia [30]. Second, we unpack how blind people munication, but these tools need to reflect users’ communication
negotiate the consequences of inaccurate GenAI content and the preferences [86]. Researchers also investigated disability represen-
effort needed to verify information by reasoning through compet- tation in GenAI and found ableist biases and stereotypes in GenAI
ing factors, such as context, stakes, verifiabililty, and believability. responses [25, 29, 60]. Focusing specifically on GenAI accessibility
Finally, we bring forth one of the first detailed accounts of blind for blind people, Das et al. [18] identified image provenance (i.e.,
individuals’ mental models of GenAI as shaped by their assump- information about image source) and aberrations (i.e., unrealistic
tions of these tools’ work processes, sources of information, and depictions in images) as important information desired by blind
response generation approach. We revisit the similarities and dif- users in the description of AI-generated images. Relatedly, Huh
ferences of these mental models with that of sighted users [98] to et al. [43] built a system to make text-to-image generation more
surface how (in)accessibility of GenAI tools uniquely shape blind accessible for blind users by providing detailed descriptions of the
people’s perceptions of GenAI capabilities and limitations, opening AI-generated images and allowing options to verify if generated
up future research directions in AI and accessibility. images follow their prompts.
While research on GenAI use among blind people remains nascent,
a larger body of work examines blind people’s experience with other
2 RELATED WORK
AI technologies. Researchers incorporated teachable AI to assist
We situate the present study within prior research on generative AI blind people in finding their personal objects [39, 68] and detailed
(GenAI) tools and practices, the intersection of AI and accessibility, what factors blind people assess when sharing their information for
and mental models of AI systems. AI datasets [47]. Others uncovered accessibility benefits and chal-
lenges blind users experience while using voice assistants [1, 76].
2.1 Generative AI Tools and Practices Collectively, prior work shows the many potential benefits, harms,
With the proliferation of large language models (LLMs), GenAI and considerations of AI for accessibility. Situated in this literature,
tools like ChatGPT, Copilot, and Gemini have gained immense pop- our study contributes to a detailed understanding of how blind peo-
ularity across various domains, including education [36, 49, 83, 99], ple use and understand GenAI tools in their regular work and the
programming [16, 54, 92], communication [14], and creative work challenges and opportunities these tools present for their practices.
[45, 56, 84]. Scholars have started exploring how to better sup-
2.3 Mental Models of AI Systems
port users’ interaction with GenAI to enhance their productivity
[45, 92]. To this end, researchers have built new systems e.g., col- Studying accessibility of complex and opaque systems like GenAI
laborative design application [88], conversational game [14], and requires unpacking how users form mental models i.e., conceptual
tools for writing better prompts [9, 93]. Others have investigated representations of the systems based on their experience interacting
how users incorporate GenAI into their workflows. For instance, with those systems [50, 71]. Without a clear conceptual understand-
prior work found that students are interested in using GenAI for ing, individuals often form their own simplified mental models
brainstorming new ideas [62] and addressing coursework-related of how a system works that do not always correspond to the sys-
queries [11]. These tools can also help prepare solutions to pro- tem’s actual functionalities [71]. Over the years, HCI scholars have
gramming problems [4, 16, 74, 99] and completing coding tasks investigated users’ mental models of complex systems including
quickly [54]. However, developers tend to avoid using AI assistants AI technologies [8, 26, 42, 66]. Researchers found that users with
due to the challenges in controlling these tools to produce desired oversimplified mental models of voice assistants have a limited
output [54]. In the creative domain, Inie et al. [45] found that cre- understanding of the privacy risks associated with those tools [42].
ative professionals are concerned about intellectual property issues When users encounter unexpected behaviors from voice assistants,
and GenAI weakening human creative sparks. Furthermore, GenAI they require explanations to refine their mental models for more
often produce inaccurate and outdated information [11, 99] and effective interaction [40]. Others captured how individuals develop
fabricated but plausible-sounding content, commonly known as mental models of the error boundary of AI systems, highlighting
hallucinations [15], which limit the reliability of these tools in im- that a good mental model can assist individuals in achieving bet-
portant use contexts. Due to these challenges, users often do not ter performance [6]. To empower individuals to actively engage
trust GenAI output and feel the need for human supervision of with AI tools rather than being passive consumers and to facilitate
AI-generated answers [4]. informed decision-making, researchers call for increased effort to
promote public AI literacy [58, 85].
Recently, researchers have started exploring how individuals
2.2 Research on AI and Accessibility conceptualize the responses from GenAI tools [3, 97]. Liao and
Set against the large and growing literature on GenAI tools and Wortman Vaughan [55] caution that interacting with LLMs with
practices, limited studies have explored the implications for GenAI flawed mental models can lead to unsafe use, over-reliance, and
among people with disabilities. Notably, Glazko et al. [30] conducted other interaction-based harms. Further, oversimplified and erro-
an authoethnographic study within a team of researchers with and neous mental models of LLMs encourage disclosures of sensitive
without disabilities to demonstrate how they used GenAI to create topics which leads to privacy risks [98]. Our study extends this
access for themselves and others and how existing GenAI tools scholarship that focused on non-disabled people by contributing
sometimes failed in this regard. Others examined GenAI use among to new insights about mental models of GenAI tools from the per-
AAC users [86] and autistic people [46], highlighting that GenAI spectives of blind individuals.
How Blind People Use and Understand Generative AI Tools ASSETS ’24, October 27–30, 2024, St. John’s, NL, Canada

3 METHOD “The picture shows a man walking in the park with

a guide dog. The man is holding a white cane in his
3.1 Participants
left hand and the dog leash in his right hand. He’s
We conducted interviews with 19 blind and visually-impaired indi- wearing sunglasses, a light jacket, a scarf, and casual
viduals who had experience using GenAI chatbots such as ChatGPT trousers with sneakers. The dog appears to be a black
3.5 or 4, Google Gemini (formerly Bard), Microsoft Copilot (formerly Rottweiler with a red collar. They are on a dirt path
Bing Chat),1 and Claude and Gen-AI powered image describer, Be surrounded by lush green trees and some grassy ar-
My AI (a feature of the Be My Eyes app). Participants were re- eas in the background. There is a small bridge over a
cruited using an online survey circulated through an organization stream with a person crossing it. The setting is peace-
that works with blind people, our research networks, and snowball ful and suggests a quiet and natural environment.”
sampling. Out of 27 respondents, we selected 19 participants, screen- Note that this Be My AI-generated description alter-
ing for their GenAI usage in the last four months. Most participants nated the items in the person’s hands. In the image,
were intermediate users of GenAI except three frequent users and the white cane is in their right hand and the dog leash
one beginner. Four participants reported using text-to-image tools is in their left hand. Also, there is no visual sign that
like DALL-E and Midjourney a few times; however, we centered our identifies the dog as a guide dog. It is not wearing any
focus on participants’ experience with text-based GenAI chatbots harness that guide dogs commonly wear.
like ChatGPT and image description apps like Be My AI. In addi-
To understand participants’ mental models of GenAI, we drew
tion to GenAI tools, all but one participant frequently used voice
on the five big ideas of AI [85] and asked participants to share
assistants (e.g., Amazon Alexa and Siri) and all but one used image
their thoughts on how these tools worked, whether and how these
description apps (e.g., Seeing AI, Google Lookout, and Envision
tools could understand their questions, how the responses were
AI). Participants primarily used one or more screen readers (e.g.,
generated, how they perceived the quality of the responses, and the
JAWS, NVDA, and VoiceOver) for information access, although
overall capabilities, limitations, and social impacts of these tools.
eight also used braille displays. All participants except two lived in
We probed participants with a particular emphasis on instances
the US. Table 1 shows details of participants’ self-reported visual
of inaccurate or unexpected responses, since expectation violation
disabilities, occupation, GenAI tools used, and frequency of using
has been found to be helpful in revealing user mental models [20].
GenAI. Table 2 shows participants’ demographic information on
To keep our study procedure accessible, we did not incorporate
an aggregate level to maintain anonymity.
any mental model drawing activity [48, 98]. Instead, we developed
our interview protocol following prior studies that used interviews
3.2 Procedure
to reveal users’ mental models of technologies [20, 22, 78]. We in-
We conducted semi-structured interviews remotely over Zoom be- tentionally avoided technical jargon like LLM, training data, or
tween January–March 2024, with approval from our university’s ‘generative AI’ unless participants mentioned these terms them-
Institutional Review Board. Interviews began with obtaining par- selves. We did not answer any questions from participants about
ticipants’ verbal consent. We first asked participants to share what how AI or GenAI worked. Interviews lasted for about 60-90 minutes.
GenAI tools they used and for what purposes. We requested them Participants were compensated with US$30 per hour (prorated) via
to walk us through their process of interacting with their pre- Amazon gift card or PayPal. All interviews were video-recorded
ferred text-based GenAI chatbot (e.g., ChatGPT, Copilot, Gemini) and transcribed for analysis.
via screen sharing, along with sharing the screen reader speech.
Participants showed examples from their previous chat histories 3.3 Data Analysis
and performed live demonstration of how they formulated prompts, We analyzed data following a reflexive thematic analysis method
read answers, and wrote follow-up prompts and any accessibility [10]. Taking an inductive approach, the first author open-coded the
issues they encountered on these tools. To demonstrate Be My AI, entire corpus while both coauthors closely read and reviewed all
participants opened the Be My Eyes app on their phone, since it codes and the data. Our initial codes captured instances, such as
was not available on desktop. Their phone screen reader read out lack of keyboard navigation support, hallucinations, techniques for
the image description generated by Be My AI. We requested partic- verifying accuracy, and more. We met weekly to discuss the codes
ipants to increase their phone volume and bring it closer to their and excerpts and compare data to data and data to codes to develop
computer (which they were using for our interview) so that we initial themes. Through this iterative process, we constructed five
could listen to and record the description through Zoom. We probed overarching themes that capture the core aspects of the ways in
participants for deeper reflections on the descriptions generated by which blind individuals use and make sense of GenAI tools.
Be My AI. For this demonstration, we sent participants two sample
images before the interviews: one showed a person with a dog and 4 FINDINGS
the other showed two persons in a shopping outlet. The images can Our analysis reveals that for blind people, using mainstream GenAI
be found at these links: image 1 and image 2. Some participants also tools in everyday practices involves leveraging its strengths for
used their own images and shared those with us after the sessions. content creation and information retrieval while navigating vari-
Below we provide the description generated by Be My AI for the ous accessibility issues, inaccuracies, and idiosyncracies of these
sample image with a dog. tools. In doing so, blind people develop interesting (and at times er-
1While describing individual participants’ experience, we use the name or version of roneous) mental models about how GenAI work and think through
the tool they reported. the harms and biases associated with these technologies.
ASSETS ’24, October 27–30, 2024, St. John’s, NL, Canada Rudaiba Adnin and Maitraye Das

Table 1: Details of interview participants. All names are pseudonyms. GenAI Usage: Frequency or an approximate count of
times GenAI tools were used in the last 4 months prior to the interviews. Gemini (formerly Bard) is a Google product. Copilot
(formerly Bing Chat) is a Microsoft product. We report the tool name/version participants mentioned. All participants regularly
used Be My AI. * denotes that the participant has programming experience, although not everyone had expertise in AI.

Name Self-reported Visual Disability Occupation GenAI Tools Used GenAI Usage
Adam* Totally blind Accessibility experience designer ChatGPT 4.0 5-6 times a day
Bella* Totally blind Adaptive tech instructor ChatGPT 4.0, Gemini, Bing Chat 2-3 times a week
Carla Legally blind with limited light perception Clinical psychologist ChatGPT 3.5 Once a week
Daisy Totally blind Personal care provider Gemini, ChatGPT >15 times
Ethan Blind Assistive tech manager ChatGPT 3.5, Copilot >15 times
Frank* Retinal degeneration Retired Bing Chat, ChatGPT, Bard >15 times
Gina Blind Instructor for the blind ChatGPT 4.0, Gemini >15 times
Henry* Totally blind. No light perception Musician ChatGPT 3.5, Bard >15 times
Ivan* Blind since birth. Leber’s Congenital Amaurosis Adaptive tech instructor ChatGPT 4.0, Bard >15 times
Julia* Visual impairment. Glaucoma, Retinopathy of Community manager ChatGPT 3.5, Bard, Bing Chat, Claude >15 times
Prematurity
Kevin Blind with limited residual vision Small business owner ChatGPT 3.5, Claude >15 times
Lily* Blind Accessibility tester ChatGPT 3.5, Bard, Copilot >15 times
Mike* Totally blind. Some light perception Student ChatGPT 3.5, Bard, Claude >15 times
Nancy Totally blind Works in a committee ChatGPT 3.5, Copilot, Gemini, Perplexity 11-15 times
Noah* Legally blind all life, totally blind 4+ years. Glau- ADA compliance testing ChatGPT 3.5 6-10 times
coma, cataracts, and Corneal Edema
Portia* Totally blind Advocate Claude, ChatGPT, Bard 6-10 times
Ruby Totally blind 4+ years Accessibility trainer intern ChatGPT 3.5, Copilot 6-10 times
Sara* No vision Worked for tech support ChatGPT 3.5, Bing Chat 6-10 times
Theo Totally blind. Retina damage and cataract in Unemployed Bard 1-5 times
right eye; prosthetic left eye

Table 2: Participants’ (n=19) demographic information on an aggregate level

Age (years) Count Race Count
18–24 1 White 10
Gender Count
25–34 8 Black 1
Male 8
35–44 3 Hispanic 2
Female 10
45–54 2 British 1
Not disclosed 1
55-75 3 Asian 1
Not disclosed 2 Not disclosed 4

4.1 Adapting to Accessibility Issues in

Generative AI Tools
Many widely adopted technologies are rife with accessibility issues
that blind people must navigate by devising various workarounds
and coping mechanisms [19, 59, 89]. GenAI tools are no exception
to this; while on the surface text-based GenAI tools like ChatGPT
may appear to be “technically” accessible, our participants reported
encountering a number of challenges due to these tools’ disregard
for established Web Content Accessibility Guidelines (WCAG) [90]
coupled with their “terrible UI” (Kevin).
During our sessions, all participants demonstrated that buttons
for copying, regenerating, and downvoting ChatGPT responses
were unlabeled (Figure 1). Hence, blind users must figure out the
Figure 1: Screenshot of the ChatGPT 3.5 interface. Unlabeled
functionalities of these buttons through trial and error or ignore
buttons are marked, e.g., buttons for copying, regenerating,
them altogether. Moreover, ChatGPT and Claude neither provided
and downvoting responses. Screen readers announced the
appropriate heading labels, regions, or landmarks for screen reader
user’s avatar and the ChatGPT icon near the prompt/response
users to swiftly traverse around the interface nor enabled any short-
as ‘graphic’, and blind users repurposed the shortcut to move
cuts to jump between previous and next prompts or responses. Our
between graphics as a workaround for quick navigation.
How Blind People Use and Understand Generative AI Tools ASSETS ’24, October 27–30, 2024, St. John’s, NL, Canada

participants had to spend considerable time in finding GenAI re- multiple GenAI tools to create the final product. For instance, Gina
sponses or navigating to the message box to type a new prompt, and Julia used Be My AI to produce descriptions of images, copied
since it required them to “brute force” (Kevin) their way through the those into ChatGPT, and prompted ChatGPT to write social me-
entire conversation repeatedly by scrolling with arrow keys. Partic- dia posts or stories based on the image descriptions. In another
ipants who used Copilot and Bard also expressed frustration with strategy, participants used GenAI as a “jumping off point” (Ivan)
circumventing extraneous sample prompts and ads, which created for brainstorming different possibilities when they felt “clueless
“a lot of clutter to sort through” (Ruby) before reaching the chat fields. about how to approach something” (Mike). Julia and Bella described
Furthermore, blind users did not get any instantaneous notification using GenAI to “make [their] own content more accessible for others”,
when the system finished response generation, requiring them to for example, by creating accessible webpages or forms [30]. Many
“dig for it” (Portia) by moving screen reader focus. participants—especially those who were English language learners
To adapt to these challenges, our participants came up with var- or were not proficient in writing—used GenAI to fix spelling, gram-
ious workarounds. Some repurposed the shortcut to move between mar, and formatting errors and translate text from one language
graphics (G + Screen Reader Modifier) to quickly jump between to another. These participants felt that GenAI made writing tasks
prompts on ChatGPT, since the user’s avatar and the ChatGPT icon “a lot less daunting” (Julia) for them. Such proofreading support is
adjacent to the prompt / response were announced as ‘graphic’ by crucial for blind people, given that blind screen reader users are
the screen reader (Figure 1). Carla, Gina, and Frank used GenAI more likely to make spelling [53, 75] and formatting errors due to
tools on phone because they found it relatively easier to locate accessibility issues in writing applications [19, 67].
information on a smaller screen using tap gestures. However, this In addition to content creation, most participants also used
approach “still involved a fair amount of hide and seek. . . because GenAI for information seeking [12, 41, 83], for instance, searching
sometimes the page scrolled” (Frank) inconsistently. Kevin, in con- about TV shows, products for shopping, or accessibility guidelines.
trast, avoided reading AI responses on the native apps altogether Some participants used GenAI for planning events or getting advice
and manually copied those to a notepad for reading and editing. on handling everyday situations. As examples, Bella gathered ideas
Participants also noted several usability issues on Be My AI, for about tactile activities to throw a party for her blind daughter’s
example, not being able to import images directly into the app from birthday, Ethan queried suggestions about raising a child as a blind
the gallery and losing conversation history once they exited the chat parent, Mike generated a roadmap on how to manage a PR (pub-
instance for a particular image. Although some of these difficulties lic relations) vertical for his college fest as a blind person, Ruby
(e.g., extraneous ads) may affect sighted users as well, our partic- curated a weight loss program, and Daisy consolidated her health
ipants highlighted the compounding impact of navigating these symptoms before talking to her doctor.
challenges using screen readers, which “slowed them down” (Julia) Related to information seeking, one unique use context of GenAI
and made their experience with GenAI tools “annoying” (Gina). for blind individuals is visual question answering [31, 35], for which
Thus, unlike sighted users, blind individuals must work through participants primarily used Be My AI but also sometimes newer
additional accessibility issues to reap the benefits of GenAI tools. GenAI models that can describe visual information e.g., ChatGPT
4. Almost every participant appreciated that compared to “one or
two sentence” (Portia) descriptions given by sighted people, Be
4.2 Leveraging Generative AI for Content My AI provided richer descriptions in a systematic way, starting
Creation and Information Retrieval with the foreground followed by the background, including details
Our blind participants incorporated GenAI tools in a wide variety of people’s attire, surroundings, objects, colors, and the overall
of content creation and information retrieval tasks, ranging from vibe. Gina shared, “I’ve actually just grown used to the Be My AI
preparing copywriting materials, emails, course outlines, resumes, descriptions because some people just don’t know how to describe
elevator pitch, cover letters, recommendation letters to program- things to blind people. They have no idea what [blind people] can and
ming and creative writing (e.g., stories, poems, songs). Although cannot see and what they want to know and don’t wanna know.”
many of these use cases and the advantages of GenAI our partici- Participants appreciated using GenAI for the above-mentioned
pants described align with the experience of sighted users [45, 84], tasks because it made their workflow “efficient” (Carla) such that
our analysis foregrounds certain contexts in which using GenAI they “didn’t have to spend hours” (Gina) to search information online
tools carry important and unique implications for blind individu- [12] and combine, revise, reformat, and proofread all that informa-
als, such as visual question answering [31, 35]. Below we broadly tion to produce the end result. Additionally, participants felt that
describe how our participants utilize GenAI tools to enhance their GenAI tools helped them develop new skills and enhance profi-
workflow, drawing out specific examples that relate to their experi- ciency in areas they had tried to learn before but did not have much
ence with blindness. success. As self-learners, some participants found it helpful to re-
Echoing findings from recent work involving sighted users [27, ceive feedback and explanations from GenAI when they needed
28, 84], blind participants shared different strategies they adopted pointers to get unstuck on problems. Julia shared, “I don’t feel silly
while using GenAI tools for creating content. In one strategy, par- asking stupid questions” to GenAI. Among our participants, Nancy
ticipants started with one or more source material(s), such as an tried to learn songwriting with ChatGPT, Mike learned simplified
outdated resume, an article, or quick scribbles jotted during a meet- explanations of academic jargon, and Bella explored mathemat-
ing, and then “ran it through the AI” (Portia) to summarize, expand, ical concepts. Reflecting on how ChatGPT helped him practice
combine, rephrase, or organize those materials into revised and programming, Henry said,
improved content. In some cases, participants chained outputs from
ASSETS ’24, October 27–30, 2024, St. John’s, NL, Canada Rudaiba Adnin and Maitraye Das

“I’m not very good at taking in heaps of information to sense whether those “sounded really weird” (Adam) or were “very
that come from all these [coding] documents. So, trying different” (Lily) from what they had anticipated based on their prior
to teach myself program has been a nightmare. Then knowledge. For instance, while exploring a sample image, Ruby,
again, I’ve never been able to find people who have time Sara, and Ethan doubted Be My AI’s description of the dog’s breed.
or patience to teach me how to do it either... And I’ve Ruby said, “Knowing what I know about guide dogs, Rottweilers aren’t
learned more from ChatGPT in the past year than I’ve generally a breed that’s used.” Refer to Section 3.2 for the full image
learned in the past 20 years... ChatGPT told me what description given by Be My AI.
functions I needed [to create an Windows app]. I was When in doubt, participants tried to find alternate ways to vali-
able to sort of look for those [functions]... in the relevant date GenAI responses. Most often, they “turned to Google” (Ivan)
documentation... and look against the snippet of code it or visited websites that were likely to contain accurate information
gave me and I was like, that should actually work.” about their queries [84]. For example, Bella checked Freedom Scien-
tific webpage for confirmation on a JAWS screen reader command.
Despite the benefits elaborated above, participants acknowl-
Besides checking external applications, participants sometimes used
edged the limitations of AI-generated content in terms of linguistic
the GenAI tools to assess accuracy. They repeated the question to
quality. Most participants observed that ChatGPT presented content
the same tool at different times (Carla) or by starting a new chat
in a distinct “bulleted” structure, which made it appear “too robotic”
(Julia) to see if they received different responses. Alternatively,
(Ethan, Ivan, Daisy), “very bookish” (Mike), and “formulaic” (Daisy).
they ran it through multiple equivalent genAI tools (e.g., ChatGPT
They also noted that ChatGPT produced redundant phrases and
and Gemini) to confirm whether they got similar responses after
overused certain “huge words” (Lily). Participants characterized this
repeated try. Additionally, participants engaged in a process of “de-
syntax as “way over the top” (Bella, Daisy), “really flowery... almost
ductive reasoning” (Ethan) with the GenAI tool by asking follow-up
too sweet” (Ruby), and “overly verbose in a way that doesn’t quite feel
questions to judge the accuracy or completeness of its response. For
human” (Ivan). Most participants felt that they can easily recognize
instance, Ethan probed Be My AI about the breed of the guide dog
“the ChatGPT style” (Noah) when they encounter unknown text.
in our sample image: “You sure it’s a Rottweiler?” Upon receiving a
Hence, while creating content, participants made sure to readjust
confident response, he followed up, “Prove to me that it’s a guide
GenAI responses to eliminate overused words and “sprinkle in some
dog.” In response, Be My AI acknowledged that it assumed the dog
of me” (Daisy). Nancy reflected on this balancing act in human-AI
to be a guide dog due to the presence of a white cane in the image.
creation: “I want AI to help me out, but I also want to put in my
Given the back-and-forth process required for verification, our
own words. . . I don’t want it to be 100% AI. So, I definitely modify
participants were judicious about when they must check accuracy
it where it sounds good, but it’s also coming from me.” Considering
of GenAI responses or when they could forego checking. They
this multi-step process for reviewing and editing GenAI responses,
agreed it was not safe to “100% rely on AI” (Nancy) for information
Ethan refrained from using GenAI for tasks like writing emails
that would be used to make financial, medical, or health-related
“because it’s actually more work.” These examples highlight that
decisions (e.g., which products to buy, which medicines to take,
incorporating GenAI in work practices often requires some extra
or food expiry date), incorporated in a professional or academic
effort from our participants. In some cases though, having to tune
context (e.g., writing a paper), or shared publicly (e.g., on someone’s
the language and structure of responses is just the start, as there is
website). However, personal use cases had more “tolerance for errors”
additional work that is necessary to navigate the inaccuracies and
(Frank) where participants felt accuracy “doesn’t matter. It’s not the
quirks of GenAI responses.
end of the world” (Ethan).
Nevertheless, we observed an overall trend among participants
4.3 Working through Inaccuracies and toward minimizing the gravity and likelihood of hallucinations.
Idiosyncrasies of Generative AI Several participants commented that they “generally trust” ChatGPT
A known limitation of mainstream GenAI tools is their tendency to responses for “high-level details”, considering those to be right “90%”
generate information that is fabricated, inaccurate, or inconsistent (Kevin, Ivan) or “99.999% of the time” (Adam) [61]. Their trust on
with input data [15]. To tackle these inaccuracies and idiosyncracies, GenAI responses bolstered when tools like Copilot or Perplexity
our participants need to navigate the challenges associated with cited links to source websites. Portia said, “It’ll already give me
verifying generated information and improving response quality. references to where that information came from, like according to the
Journal of Psychiatry... So then I don’t have to fact check.” Even those
4.3.1 Identifying Hallucinations and Verifying Accuracy. Our partic- who were more skeptic about GenAI accuracy (Ivan, Lily, Ruby,
ipants shared many examples of factually inaccurate or fabricated Julia, Daisy, Mike) tended to trust image descriptions from Be My
information (i.e., hallucination) provided by GenAI. For instance, AI because those seemed to be “detailed enough” (Ruby) and also
Be My AI described Daisy’s raincoat pattern as “hearts and stars” al- because participants appreciated the “independence” afforded by Be
though it was “clouds and raindrops” and ChatGPT replaced Ethan’s My AI descriptions over seeking sighted help for verification.
name with a fictitious name ‘Chris’ when he asked it to revise
and update his old resume. Daisy emphasized that GenAI tools 4.3.2 Enhancing Response Quality through Prompt Engineering.
projecting “confidence” in their hallucinated responses, especially Blind participants assumed suboptimal prompts or other inputs to
regarding visual information, “can be misleading to someone who GenAI to be a likely reason for erroneous or low-quality responses.
isn’t able to visually verify for themselves what something looks like.” For instance, all participants’ first reaction to inaccurate image
To assess the validity of GenAI responses, blind participants tried descriptions from Be My AI was poor image quality or visually
How Blind People Use and Understand Generative AI Tools ASSETS ’24, October 27–30, 2024, St. John’s, NL, Canada

uncertain objects in input images. They commented that lighting, under the hood. Ruby (ChatGPT 3.5 user) gave an example where
camera angle and distance from the target object, blurriness, reflec- she felt that ChatGPT response to her query about a weight loss
tion or glare, partially obscured objects—all these could “confuse” program “might have been pulled from maybe a nutritional website or
(Frank) GenAI tools and lead to inaccurate image descriptions. Some medical website, the fact that it tells you to consult with your doctor.”
participants were more willing to critique the photos they took than Informed by this mental model, most participants used GenAI tools
attribute unsatisfactory image descriptions to GenAI limitations, as a “replacement of Googling” (Ivan), given that it significantly
and often tried to capture better photos before running it through streamlined their information foraging workflow [12]. Ethan called
GenAI. In doing so, they exhibited a tendency to downplay GenAI ChatGPT “Google search on steroids” that can “deliver what you
limitations while putting the onus of capturing good quality photos need to know right on the screen without having to sift through”
on themselves. For instance, while describing a sample image, Be (Gina) links and articles returned by search engines. Certain GenAI
My AI consistently alternated the items in the person’s left and limitations bolstered this perception. Participants speculated that
right hands (refer to Section 3.2 for the full description). To specu- GenAI occasionally provided inaccurate information because it
late why this might have happened, Sara said, “People have told me gathered data from outdated and unverified internet content.
when I’ve uploaded things...it’ll stretch them out or do wacky stuff to While some participants used GenAI tools that indeed had inter-
my pictures. So, I wonder if it got turned around when it got uploaded,” net search capabilities (e.g., ChatGPT 4, Copilot), this perception
although this was not the case. Likewise, participants tended to was also evident among participants who used ChatGPT 3.5 that
blame the quality of their text prompts when interpreting possible did not have such features. Some ChatGPT 3.5 users knew that
reasons for “off base” GenAI responses. Ethan said, “If I get a bad it was unable to correctly answer questions about recent events.
response, I gave it a bad prompt. It’s my fault.” However, they were either unsure exactly how this older version
Participants reflected on their effort to learn “prompt crafting” or collected information or assumed that it also gathered information
“prompt engineering” to improve GenAI response quality [51]. For from the internet. Nancy (ChatGPT 3.5 and Copilot user) said, “If
example, they often tried to create specific and detailed prompts “to you are researching something on Copilot, I think it literally goes to
lead ChatGPT in the correct direction” (Carla). Ethan, Julia, Henry, Google and pulls up information. I think for things like ChatGPT, it
Adam, and Kevin considered follow-up questioning as a useful probably does that too, but I don’t know how it’s different because it’s
mechanism to remind GenAI tools their original asks if it started not as current as Copilot.”
making assumptions about their queries. Julia explained: “I just
played whack-a-mole and whack-a-mole until it finally did what 4.4.2 “King of Knowledge”. Unlike the previous mental model that
I wanted it to do. And you have to problem-solve these instances equated GenAI tools with a search engine, some participants spec-
to realize . . . What did it assume? Let’s figure that out and fix it.” ulated that GenAI tools conducted a keyword-based query directly
Interestingly, other participants recalled that follow-up questioning into one (or more) of its “massive database” which contained “all
degraded response quality, because the GenAI tools “lost track of the kinds of information” gathered from textbooks, archives, coding
conversation” (Mike) as the chat got longer than “5 or 10 messages” manuals, and other web content. Portia elaborated, “I feel like the in-
(Henry). This divergence in opinions indicates how participants formation has to sit somewhere, even if it’s in the cloud. . . It (ChatGPT
formed different ideas and expectations about GenAI based on their 3.5) doesn’t just spit out of nothing. . . It had to go have those answers
own experiences and perceptions about these tools. from some database that has already researched it. . . I don’t know
how that database got created or who’s adding or removing.” Due to
4.4 Developing Mental Models of Generative AI this mental model, Sara believed in the superiority of ChatGPT for
information seeking: “Honestly, I look at it as the king of knowledge. . .
Our participants actively tried to hypothesize how GenAI tools
So, if ChatGPT doesn’t know what to say, I’m just not gonna find what
worked and accordingly adapted their interactions with these tools.
I’m looking for because it ain’t there.”
They described “playing around” with these tools by progressively
Interestingly, Ethan and Bella, who were familiar with the term
asking simple to advanced questions to “test its knowledge” (Sara)
LLM, thought of it as a huge database on which GenAI tools per-
and “learn its limits” (Adam). They also learned about GenAI tools
formed a keyword-based search. Ethan said,
from discussion with friends, family members, and other blind
users, accessibility webinars and training programs, news articles, “An LLM is basically an infinite amount of data essen-
official documentations, and participating in beta testing of GenAI tially that’s been fed into a computer. That computer
tools. Below we distill salient mental models of GenAI tools we can then parse for information that you’re looking for.
observed among our participants. These mental models particularly They can gather specific data from that model to give
relate to the text generation capabilities of GenAI and do not dive you an answer to something you wanted. So, if you ask a
deeper into image recognition processes of tools like Be My AI. question, it goes back into its system and sees—Has that
While some mental models align with that of sighted users [98], question been asked before? Can I string information
there are important differences that are shaped by our participants’ together?—to give that person an answer.”
experience with blindness and usage of assistive technologies, and
we return to this point in the Discussion (Section 5.2). The above two mental models guided our participants’ belief
that GenAI tools were better at addressing “straightforward” (Noah)
4.4.1 “Google Search on Steroids”. Most participants believed that questions that had “factual” (Bella, Noah), “concrete” (Lily), and
GenAI tools “pulled up information from the internet” either by “objective” (Portia) answers, because the tools can “come back with
directly searching and/or using a search engine like Google or Bing pretty much correct answers based on information it pulls” (Bella).
ASSETS ’24, October 27–30, 2024, St. John’s, NL, Canada Rudaiba Adnin and Maitraye Das

4.4.3 “Word Generating Machine”. Unlike models described in Sec- anthropomorphization extended beyond GenAI capabilities to their
tions 4.4.1 and 4.4.2, Mike, Kevin, Julia, and Adam understood GenAI idiosyncratic behaviors. He said,
tools as “word generating machines” (Adam). Mike explained, “I “I just find it quite amusing that sometimes I’ll have
think it works on the word prediction or language prediction model... good days with it, sometimes I have bad days with
I would probably call it an advanced parrot who can understand lan- it and I’m not sure what’s causing that. That’s very
guage. It understands the patterns and the technicalities behind the human—humans have off days and good days, but I
language, how the language works.” Kevin also believed that Chat- can’t imagine why that would happen with AI.”
GPT could predict the next likely word given an input sequence of
Adam used anthromorphized metaphors to describe how he had
words. However, he held a more skeptic view of GenAI:
developed certain levels of trust and comfort with GenAI tools over
“When I throw a sentence into it. . . it spits out connected time by “treating it like a partner in getting things done... At the
words that kind of go in the direction of the prompt. beginning, you’re getting used to each other’s quirks... By the time
There is clearly no overarching intelligence in there. It you’ve been working together for a while, you know how each other
just comes up with words that string together and they work, what you can trust and what you can’t.”
kind of sound vaguely intelligent based off what you’re
trying to ask it.” 4.4.7 “Still a Computer, Not a Human”. Despite some anthropomor-
phization, upon deeper reflection, Ethan, Adam, Daisy, and Mike
Unsurprisingly, participants who shared this mental model had agreed that GenAI tools are “just a computer. It’ll never be as good
programming backgrounds themselves or partners who worked in as a human.” This mental model was informed by situations where
the technology sector. GenAI tools faltered at solving “mathematical and logical problems”
(Mike) and issues that require “nuanced reasoning” (Daisy), such
4.4.4 Stores and Reuses User Prompts. Some participants thought
as determining words containing certain letters for an anagram
that GenAI tools gathered information from users’ conversation
game or solving complicated coding problems. Adam explained, “It
histories. Sara explained,
doesn’t think like a regular developer does. . . So, bugs [in codes] that
“It will learn from what other people put in, I think. might seem obvious to a user or developer. . . through the amount of
If I ask it to write me an email... maybe it is able to experience we’ve had coding, might not seem obvious to it.” Likewise,
grab from someone else who asked the same question or participants believed GenAI tools could not perform well at ad-
someone who wrote a similar email. It probably stores dressing requests that require “creativity” (Daisy, Adam, Noah) or
stuff that we all do, anonymously in some way, in order expressing “subjective opinions” (Lily, Noah, Adam), such as writing
for it to learn what people want and what people like.” novels or presenting arguments on whether mountains or oceans
Important to note is that participants here did not refer to the are better for vacations. Adam shared that the reason GenAI tools
mechanism of user feedback in GenAI tools; in fact, many par- are not creative lies in the way they are fundamentally constructed.
ticipants did not know about upvote/downvote buttons because “This is not an idea generating machine. . . If I wanted
those were unlabeled and inaccessible (see Section 4.1). In this case, it to write about the quests that the characters [in a
participants understood users’ conversation histories to be con- story] undertook, it’s going to regurgitate very common
tributing to other sources of information used by GenAI (e.g., the themes from classic lit. . . It’s not going to come out with
large database mentioned in Section 4.4.2). a new one out of whole cloth that’s gonna turn anyone
into a bestseller author. . . It does not have a human
4.4.5 “More In-depth AI”. Participants conceptualized how GenAI spark of creativity, doesn’t think outside the box.”
worked by comparing them with other applications they had used Collectively, these mental models reveal the ways in which our
for accomplishing similar tasks. They compared text-based GenAI participants conceptualized how GenAI tools retrieve and store
like ChatGPT with voice assistants (e.g., Alexa, Siri, Google Assis- information and generate responses, drawing comparisons with
tant) and Be My AI with other image description apps (e.g., Seeing other AI systems and even humans. These mental models were
AI, Tap Tap See). In both cases, they considered GenAI to be “more rooted in their firsthand experience with GenAI as blind screen
in-depth AI” (Ivan) for their ability to provide comprehensive infor- reader users, as are their perceptions about harms and biases of
mation in a more conversational way. In fact, when contrasted with GenAI, which we elaborate on in the next section.
GenAI, several participants critiqued voice assistants and older
image description apps for being “not too intelligent” (Lily) and 4.5 Reflecting on Biases and Harms of
questioned whether those apps can be truly characterized as AI.
Generative AI
4.4.6 Partner, Friend, Mentor, Secretary... Participants used sev- Aligning with prior work [25, 29, 60], our analysis foregrounds how
eral anthropomorphized [21, 44] metaphors to describe the ways blind individuals think through biases and harms of GenAI while us-
in which they conceived of GenAI tools, such as “your notetaker ing these tools for content creation. Participants recounted several
slash secretary you never had” (Gina), “a librarian working for you” instances where GenAI tools could not “handle nuanced concepts”
(Frank), “a friend who’s your personal proofreader. . . and who can (Daisy) about disability and produced ableist and ageist content.
be honest to help clarify what you’re really trying to say” (Portia), During our sessions, Noah and Gina prompted ChatGPT to create
and “a writing assistant, editor. . . best friend, and mentor kind of a story involving a blind person traveling to a new country. In the
thing that never judges and is never in a bad mood” (Julia). To Henry, stories, ChatGPT described the blind person as ‘courageous’ and
How Blind People Use and Understand Generative AI Tools ASSETS ’24, October 27–30, 2024, St. John’s, NL, Canada

‘resilient.’ Noting this characterization of blind people as “stereotyp- biases. And the disabled community and the commu-
ical,” Noah elaborated on why certain linguistic choices by GenAI nity of minorities face bias every day. And so, these
might be subtly inappropriate: “It’s not a major deal breaker... But artificial intelligence models that are being built, when
I can hear some of my blind friends [say] we’re not all courageous.” they are searching the internet, their sources are going
Daisy recalled asking ChatGPT to generate a story about a “dis- to be impacted by bias... racism, ableism, and so on.”
abled person going on an adventure. It couldn’t do that. It was like, Reflecting on GenAI’s other negative impacts, participants were
‘Here’s a family of adventurers, but then they had a disabled person concerned about the proliferation of mis- and disinformation through
and that person stayed home.’” Similarly, while brainstorming ideas GenAI. Sara, Henry, Mike, and Ethan shared that the “biggest fear”
for messages to include in a birthday card for her mother, Carla of the GenAI boom was the rise in propaganda, deep fake videos
prompted ChatGPT “to give a compliment about being old. . . but it and images impersonating people without their consent, and the
made a couple of negative comments about being old.” use of voice cloning AI to scam others. Henry said, “Usually I’m
GenAI tools exibited ableism not only in the generated content very good at kind of detecting scams but [voice cloning] is one scam
but also in their interaction with users. Ethan, Bella, and Carla en- I don’t think I will be able to detect.” Our participants maintained
countered situations where they were looking for blindness-related extra caution before using GenAI responses in content they would
information on ChatGPT and Copilot and the tools expressed grief publicly share, because they did not want to “create more fake news
for their disability, saying “I’m sorry, you’re blind.” This frustrated in the world” (Kevin). Julia echoed this sentiment, saying: “If I’m alt
Bella: “Seriously, can we move on? I don’t really need the AI thing apol- tagging [an image using Be My AI] for purposes of others being able
ogizing to me because I’m blind.” Participants tried to correct GenAI to access it with a screen reader, I will check with someone sighted. . .
by explaining in the chat why those reactions were inappropriate, to make sure that everything is described correctly because I don’t
because they believed that chat history stored by GenAI tools would want to give misinformation.” Portia, Theo, Henry, and Ethan were
be reused for future improvements (see Section 4.4.4). Carla elabo- apprehensive about privacy issues due to a limited understanding
rated, “I’ll start with trying to give a correction, like you shouldn’t about whether GenAI tools stored their information, for what pur-
tell blind people that you feel sorry for them... in the hopes that that poses and how long, and how that information would be used [98].
would be incorporated in the future.” Thus, blind participants ex- The ability to retrieve previous chat history fueled their concerns.
pended considerable effort in providing feedback to mitigate ableist While reasoning through these promises and perils of GenAI
GenAI responses, which exemplifies the significant advocacy labor on individual, interpersonal, and social level, our participants ex-
disabled individuals must perform to voice their needs and reduce pressed willingness to embrace the growth in GenAI. They were
equity gaps reified by technologies [80]. cognizant of and concerned about potential harms of GenAI; how-
Echoing findings in prior work [25], our participants hypothe- ever, they did not want GenAI’s progress to be stifled, given its
sized biased dataset as a key reason behind ableist and inappropriate positive impacts on enhancing and scaling accessibility [30, 37, 86].
GenAI responses. Kevin explained, “The dataset is probably mostly Sara said, “The worst that could happen is it would all just go away
[nondisabled] people writing about us rather than people in the dis- and people would stop developing it, and it would be sort of like some-
abled community.” Carla thought that GenAI responses might be thing that we had for a minute, and it was great and then it just sort of
driven by prevalent “misunderstandings about particular disabilities” fizzled out.” Overall, these perspectives from blind participants have
and would further reinforce those misconceptions, such as blind implications for future efforts to reconcile the positive and negative
people desiring to “feel [someone’s] face... to help them visualize” effects of GenAI tools, which we will revisit in the Discussion.
how others look like. Given the impact of biased datasets on GenAI
output, participants felt that “feeding the model large amounts of 5 DISCUSSION
data written from the disability perspective would be good” (Kevin).
We have presented one of the first detailed accounts of the prac-
Besides issues related to disability and ableism, our participants
tices and mental models of generative AI among blind people. Be-
were cognizant of GenAI showing biased portrayals of other aspects
low we synthesize our findings to rethink access and information
of identity like race and gender. In one instance, Carla noticed that
verification in GenAI, unpack erroneous mental models of GenAI
Be My AI were “assuming short hair meant a boy as opposed to a
among blind individuals, and reflect on harms and benefits of GenAI
girl.” Similarly, Kevin proactively edited ChatGPT responses when
through an accessibility-centric lens.
it misgendered somebody or described one’s disability in a way not
preferred by them. As another example of biases against underrep-
resented populations, during our session, Adam asked ChatGPT to
5.1 Rethinking Access and Information
formulate sentences in Maori language and found two out of ten Verification in Generative AI
resultant Maori sentences to be grammatically incorrect. He specu- Building accessible technology requires critically considering not
lated the lack of representative dataset as the reason behind this: only whether disabled people can access it on a basic level but also
“Maori is a very low resource language. There aren’t a lot of people the extent to which they can leverage the full benefits of these
that put it on the web. It’s very underrepresented in the datasets. So, technologies. As our analysis demonstrates, the inherent accessi-
it’s not gonna know as much about it as it will know English.” Daisy bility of text-based interaction enables basic levels of nonvisual
critiqued issues around AI fairness and bias more broadly: access in current GenAI tools; however, these tools still leave a lot
to be desired for blind users due to the lack of accessible keyboard
“We call them artificial intelligence, but they are ulti- navigation, unlabeled buttons, and poor UI design. For example, the
mately based on humans and humans have internal suboptimal and inaccessible UI of some GenAI tools required our
ASSETS ’24, October 27–30, 2024, St. John’s, NL, Canada Rudaiba Adnin and Maitraye Das

participants to piece together a cumbersome workflow involving believability as yet another key factor that problematizes the veri-
multiple GenAI and writing applications just to read and edit AI- fication decision-making process for blind users, especially those
generated responses. For blind users, such issues greatly diminish who may not be knowledgeable of the idiosyncrasies of GenAI and
the efficiency gained from using GenAI, which has been positioned are less likely to be skeptical about generated responses. These
as one of the biggest benefits of using it for content creation [45, 84]. issues are not confined to visual information only but apply more
On top of this, inaccuracies and hallucinations in GenAI [11, 83] broadly to all GenAI tools. For instance, references to source web-
add additional layers of complexity for blind users. The possibility sites, as tools like Microsoft Copilot and Perplexity include in their
of inaccurate responses requires users to decide whether and to responses, reinforce their perceived believability among our partic-
what extent they need to verify the generated responses, and we ipants. While citing references may seem to help users understand
observed four factors that shaped this decision: context of use, stakes, provenance i.e., source of the information presented, researchers
verifiability, and believability. While all users—blind and sighted have found that GenAI-cited references are often inaccurate or do
alike—need to grapple with these factors to counteract GenAI inac- not substantiate the associated statements [57, 63].
curacies, we argue that blind individuals’ use cases require unique It is also important to highlight that the four factors stated above
considerations. Moreover, accessibility issues of GenAI tools can se- are not siloed, rather they are often competing or at tension with
verely limit which of these factors they can prioritize when deciding each other. For example, when stakes are higher, a user would be
whether to verify a response, adversely affecting their likelihood of willing to verify a response despite high believability. Conversely,
avoiding misinformation [81, 100]. even if a generated image description has slightly low believability,
First, with regards to the context of use, we observed that blind a blind user might be willing to forgo verification because of low
users are more likely to verify GenAI responses that would be verifiability, such as due to the unavailability of sighted help.
used in medical, health, education, financial, professional, or pub- Given these issues and tensions, we argue that researchers and
lic contexts compared to personal use cases. Second, even within developers must work toward reducing frictions that minimize the
the same context, blind users consider whether the stakes are high benefits of using GenAI for blind users. In addition to enforcing es-
enough to justify verification. For instance, when generating image tablished accessibility principles within individual GenAI tools, we
descriptions for food labels during cooking, the verification stakes suggest that further attention be given toward tailoring the reading
are higher for information about the presence of allergens or the and editing experiences within GenAI tools for nonvisual access
expiry date of a product than for other information with higher so that the effort needed from blind individuals to review or edit
tolerance for error, such as heating time or how much seasoning generated content (which currently requires switching back and
to add. Similarly, in the context of sharing images on social media, forth between multiple apps) does not diminish the benefits they
the AI-generated alt text of a personal photo shared for fun pur- receive from using GenAI. More importantly, we feel there is an
poses has lower stakes for accuracy than the one for an infographic acute need for seamless ways to verify information in GenAI tools
containing important health-related information. Third, users con- so that blind users do not need to expend significantly extra effort
sider how easily GenAI responses can be verified to judge whether for verification. Developers may consider integrating strategies
verification is worth the effort (i.e., verifiability) [57]. When pro- blind users already adopt to further streamline their verification
ducing image description using GenAI (e.g., By My AI), sighted workflow. For example, GenAI tools may test its response consis-
users can readily determine mismatches between visual content tency across repeated tries in the same tool or in multiple tools and
and the generated description, whereas blind users need to seek summarize these inconsistencies for blind users which may encour-
sighted help, significantly reducing verifiability for blind users in age constructive skepticism among users [18] towards otherwise
situations where sighted help is unavailable or inconvenient. Even believable GenAI responses.
for textual responses given by GenAI, efforts needed to verify (e.g.,
through another search engine) can be higher for blind individu-
als due to the inaccessibility and usability issues associated with
copying and pasting responses across different platforms [59, 89].
5.2 Unpacking (Erroneous) Mental Models of
Finally, blind users rely on perceived believability [70, 91] of GenAI Generative AI among Blind People
responses for deciding whether or not to verify the information pre- Our analysis reveals that blind users often develop flawed, incom-
sented. Our participants shared that they often forego fact-checking plete, or oversimplified mental models of GenAI tools [55, 71, 95].
if GenAI responses do not “seem fishy” or “unexpected.” As prior Several of these mental models align with that of sighted users, as
work found, blind users tend to unduly trust auto-generated im- found in recent work [98]. For instance, our participants’ mental
age descriptions [61]. GenAI tools amplify this issue since their models of GenAI chatbots, even ChatGPT 3.5, as collecting infor-
responses are relatively richer, more comprehensive, and detailed, mation from the internet through a search engine (Section 4.4.1) or
which increases the believability of these responses compared to employing keyword-based search on a massive database (Section
other image recognition applications (e.g., Be My AI vs Seeing AI). 4.4.2) match sighted users’ understanding of ChatGPT as a “super
Thus, the richness of the GenAI responses becomes a double- searcher.” Similarly, those with more technical know-how (both
edged sword, which on one hand significantly improves access to blind and sighted [98]) understood GenAI chatbots as “advanced
visual information while also making it more likely for people to parrots” or stochastic “word generating machines.” Our analysis also
trust inaccurate information. The issues of stake and verifiability reveals potentially severe implications of erroneous mental models
have also been discussed by Glazko et al. [30]. Our analysis reveals [55]. For example, most blind participants had a misconception that
GenAI tools were good at answering factual questions [57, 63, 94],
How Blind People Use and Understand Generative AI Tools ASSETS ’24, October 27–30, 2024, St. John’s, NL, Canada

informed by the belief that these tools (even the ones without ac- highlighting the need for more representative datasets [25, 30], we
tual online access) always pulled up the most relevant and accurate argue that researchers and developers need to critically examine
information available on the internet. and update GenAI tools’ default response behavior such that it is
Importantly, blind users’ mental models are also shaped by their not codified to reinforce ableist narratives about disability.
experience with blindness, the assistive technologies they use, and Furthermore, we call for a nuanced approach in addressing bi-
the level of accessibility in GenAI tools, which may differ from the ases and harms around GenAI such that measures taken to alleviate
way sighted users form mental models. For example, sighted users harms do not disregard the accessibility support disabled people
understood ChatGPT to be incorporating quality-related feedback receive from GenAI tools [37]. Our participants emphasized how
provided through the upvote/downvote buttons [98]. In contrast, GenAI helped them address critical needs, such as spellchecking
many blind participants were unaware of the upvote/downvote but- and formatting support while writing and coding as well as getting
tons because those were unlabeled and inaccessible. They instead detailed visual information—areas where existing systems are inac-
provided feedback by typing follow-up prompts in the chat (e.g., cessible and extremely challenging to navigate for screen reader
asking ChatGPT to not express grief for someone’s blindness), and users [19, 53, 67]. Thus, while the use of GenAI in academic writing
because they could go back to their conversation history later, they brings forth legitimate concerns around plagiarism [17, 69], imple-
assumed that ChatGPT would utilize these follow-up corrections menting extreme measures (such as outright bans on GenAI use
in future conversations to improve responses. at schools) to counteract plagiarism may deprive blind students of
This juxtaposition between mental models of blind and sighted the accessibility benefits of GenAI such as automated proofreading.
users from a user experience perspective shows how accessibility Another recent example of this is the public outcry on social media
oversights as simple as an unlabeled button may lead to divergent when Be My AI stopped describing images with people’s faces [23].2
and erroneous mental models among blind users. We argue that This exemplifies the tensions around the ways in which address-
any measures taken to improve transparency and trust of users for ing harms in one dimension (e.g., privacy violation due to facial
GenAI tools [55, 79] (e.g., UI redesign to minimize deceptive or dark recognition in images [82]) may reify harms in another dimension
patterns [98]) must be examined from the perspective of nonvisual (e.g., revoking the accessibility benefits for blind users, which might
access. This is just one example of how flawed mental models can be construed as a quality-of-service harm [82]). Finding an ethical
be shaped by inaccessible design even on a simple, text-based chat and productive way forward to combat GenAI-related harms with-
interface; however, as GenAI interfaces continue to evolve and out minimizing the progress in accessibility is a tough challenge
integrate more complex, multimodal features [64], supporting blind that does not have a clear-cut solution but one that researchers,
users in forming or shifting to accurate mental models will remain designers, and policymakers in AI, fairness, and accessibility must
a critical challenge in AI and accessibility. approach thoughtfully and carefully together.
Furthermore, our participants’ anthropomorphized perception of
GenAI as “best friend” or “partner in getting things done” may have 5.4 Limitations and Future Work
contributed to their heightened (and often misplaced) trust on these An important limitation of our study is that most of our participants
tools. However, the ways in which anthropomorphized descriptions were intermediate users of GenAI, although we had a few experts
of AI influence the public’s trust and reliance on these tools are and one beginner. Future studies can specifically focus on the expe-
complicated [44]. As such, further research is required to uncover riences of blind people who are beginner or expert users to uncover
the nuances of anthropomorphization, trust, and mental models of the similarities and differences in their usage patterns and mental
GenAI among blind people and how these aspects are shaped by models. Additionally, we purposefully kept our focus broad to reveal
the publicity and media representation of GenAI capabilities [55] the general GenAI usage patterns among blind users across diverse
and their impacts on accessibility. contexts including information retrieval, coding, copywriting, cre-
ative writing, and more. Future work can extend our findings by
5.3 Reconsidering Harms and Benefits of investigating nuanced practices and accessibility within specific
use cases. Finally, our analysis only focused on blind users’ experi-
Generative AI through the Lens of Access
ence with text-based and image-description GenAI tools. Further
Our analysis joins that of others who call attention to harmful dis- research is needed to explore how blind people interact with multi-
ability representations in GenAI, both in text produced by LLM chat- modal GenAI tools, e.g., image and music generation to develop a
bots [25, 29] and images generated with text-to-image models [60]. holistic understanding of GenAI accessibility.
We bring out empirically-driven insights from blind participants’
everyday experiences, reconfirming the prevalence of ableist and 6 CONCLUSION
ageist biases in GenAI. For instance, our participants encountered
Through our inquiry into Generative AI (GenAI) usage of blind
stereotypical characterization of blind people (e.g., courageous, re-
individuals, we uncover in-depth empirical understandings of the di-
silient) that bordered on ‘inspiration porn’ [34], i.e., languages that
verse ways in which blind people use GenAI to streamline their con-
objectify disabled people as being inspirational for the gratification
tent creation and information retrieval workflows, often working
of non-disabled people. Important to note is that biases exist in not
around various accessibility and usability issues in GenAI interfaces.
only the content produced by GenAI but also the ways these tools
Through this, we shed light on the complex cost-benefit analysis
interact with users. Participants found that often questions about
ideas for performing a task as a blind person are met with ChatGPT 2A later update reverted this change and as of the writing of this paper, Be My AI was
expressing pity and grief about their blindness. Thus, in addition to describing images with faces again [77].
ASSETS ’24, October 27–30, 2024, St. John’s, NL, Canada Rudaiba Adnin and Maitraye Das

blind users perform to navigate the inaccuracies and idiosyncrasies [14] Tiffany Chen, Cassandra Lee, Jessica R Mindel, Neska Elhaouij, and Rosalind
of GenAI tools while managing the effort for information verifica- Picard. 2023. Closer Worlds: Using Generative AI to Facilitate Intimate Con-
versations. In Extended Abstracts of the 2023 CHI Conference on Human Factors
tion. Additionally, we reveal blind individuals’ mental models of in Computing Systems (Hamburg, Germany) (CHI EA ’23). ACM, Article 68,
GenAI systems which both align with and differ from that of sighted 15 pages. https://doi.org/10.1145/3544549.3585651
[15] Yuyan Chen, Qiang Fu, Yichen Yuan, Zhihao Wen, Ge Fan, Dayiheng Liu,
users but are often erroneous and oversimplified nonetheless. We Dongmei Zhang, Zhixu Li, and Yanghua Xiao. 2023. Hallucination Detec-
argue that to enable equitable opportunities for blind individuals to tion: Robustly Discerning Reliable Answers in Large Language Models. In
leverage the benefits of GenAI, we must revisit the design and pol- Proceedings of the 32nd ACM International Conference on Information and Knowl-
edge Management (Birmingham, United Kingdom) (CIKM ’23). ACM, 245–255.
icy discussions around supporting users in building accurate mental https://doi.org/10.1145/3583780.3614905
models, verifying information accuracy, and combating biases and [16] Ruijia Cheng, Ruotong Wang, Thomas Zimmermann, and Denae Ford. 2024. “It
harms of GenAI through the lens of nonvisual access. would work for me too”: How Online Communities Shape Software Developers’
Trust in AI-Powered Code Generation Tools. ACM Transactions on Interactive
Intelligent Systems (TIIS) 14, 2, Article 11 (may 2024), 39 pages. https://doi.org/
7 ACKNOWLEDGMENT 10.1145/3651990
[17] Debby R. E. Cotton, Peter A. Cotton, and J. Reuben Shipway. 2024. Chatting
We thank our participants for their contributions to this study. We and cheating: Ensuring academic integrity in the era of ChatGPT. Innovations
are grateful to Abir Saha for his feedback on multiple iterations of in Education and Teaching International 61, 2 (2024), 228–239. https://doi.org/
10.1080/14703297.2023.2190148
this draft. We also thank Cynthia Bennett for a thought-provoking [18] Maitraye Das, Alexander J. Fiannaca, Meredith Ringel Morris, Shaun K. Kane,
conversation about GenAI harms. and Cynthia L. Bennett. 2024. From Provenance to Aberrations: Image Creator
and Screen Reader User Perspectives on Alt Text for AI-Generated Images.
In Proceedings of the CHI Conference on Human Factors in Computing Systems
REFERENCES (Honolulu, HI, USA) (CHI ’24). 21 pages. https://doi.org/10.1145/3613904.3642325
[1] Ali Abdolrahmani, Kevin M. Storer, Antony Rishin Mukkath Roy, Ravi Kuber, [19] Maitraye Das, Darren Gergle, and Anne Marie Piper. 2019. “It doesn’t win you
and Stacy M. Branham. 2020. Blind Leading the Sighted: Drawing Design friends”: Understanding Accessibility in Collaborative Writing for People with
Insights from Blind Users towards More Productivity-oriented Voice Interfaces. Vision Impairments. Proceedings of the ACM on Human-Computer Interaction 3,
ACM Transactions on Accessible Computing (TACCESS) 12, 4, Article 18 (jan CSCW, Article 191 (nov 2019), 26 pages. https://doi.org/10.1145/3359293
2020), 35 pages. https://doi.org/10.1145/3368426 [20] Michael Ann DeVito, Jeremy Birnholtz, Jeffery T. Hancock, Megan French, and
[2] Gus Alexiou. 2023. Envision Adds ChatGPT AI Sight Assistance Sunny Liu. 2018. How People Form Folk Theories of Social Media Feeds and
To Its Smart Glasses For The Blind. Retrieved 23 April, 2024 from What it Means for How We Study Self-Presentation. In Proceedings of the 2018
https://www.forbes.com/sites/gusalexiou/2023/04/30/envision-adds-chatgpt- CHI Conference on Human Factors in Computing Systems (Montreal QC, Canada)
ai-sight-assistance-to-its-smart-glasses-for-the-blind/?sh=d443c835446f. (CHI ’18). ACM, 1–12. https://doi.org/10.1145/3173574.3173694
[3] Safinah Ali, Prerna Ravi, Randi Williams, Daniella DiPaola, and Cynthia Breazeal. [21] Nicholas Epley, Adam Waytz, and John T Cacioppo. 2007. On seeing human: a
2024. Constructing Dreams Using Generative AI. Proceedings of the AAAI three-factor theory of anthropomorphism. Psychological review 114, 4 (2007),
Conference on Artificial Intelligence 38, 21 (Mar. 2024), 23268–23275. https: 864. https://doi.org/10.1037/0033-295X.114.4.864
//doi.org/10.1609/aaai.v38i21.30374 [22] Motahhare Eslami, Karrie Karahalios, Christian Sandvig, Kristen Vaccaro, Aimee
[4] Matin Amoozadeh, David Daniels, Daye Nam, Aayush Kumar, Stella Chen, Rickman, Kevin Hamilton, and Alex Kirlik. 2016. First I "like" it, then I hide
Michael Hilton, Sruti Srinivasa Ragavan, and Mohammad Amin Alipour. 2024. it: Folk Theories of Social Feeds. In Proceedings of the 2016 CHI Conference on
Trust in Generative AI among Students: An exploratory study. In Proceedings of Human Factors in Computing Systems (San Jose, California, USA) (CHI ’16). ACM,
the 55th ACM Technical Symposium on Computer Science Education V. 1 (Portland, 2371–2382. https://doi.org/10.1145/2858036.2858494
OR, USA) (SIGCSE 2024). ACM, 67–73. https://doi.org/10.1145/3626252.3630842 [23] Be My Eyes. 2023. Image to Text AI and Blurred Faces: What’s Going On?
[5] Anthropic. 2024. Claude: Conversational AI by Anthropic. Retrieved 23 April, Retrieved April 23, 2024 from https://www.bemyeyes.com/blog/image-to-text-
2024 from https://www.anthropic.com/claude. ai-and-blurred-faces-whats-going-on.
[6] Gagan Bansal, Besmira Nushi, Ece Kamar, Walter S Lasecki, Daniel S Weld, and [24] Stefan Feuerriegel, Jochen Hartmann, Christian Janiesch, and Patrick Zschech.
Eric Horvitz. 2019. Beyond Accuracy: The Role of Mental Models in Human-AI 2024. Generative AI. Business & Information Systems Engineering 66, 1 (2024),
Team Performance. In Proceedings of the AAAI Conference on Human Computa- 111–126. https://link.springer.com/article/10.1007/s12599-023-00834-7
tion and Crowdsourcing, Vol. 7. 2–11. https://doi.org/10.1609/hcomp.v7i1.5285 [25] Vinitha Gadiraju, Shaun Kane, Sunipa Dev, Alex Taylor, Ding Wang, Emily
[7] Edward C Bell and Natalia M Mino. 2015. Employment Outcomes for Blind Denton, and Robin Brewer. 2023. “I wouldn’t say offensive but...": Disability-
and Visually Impaired Adults. Journal of Blindness Innovation and Research 5, 2 Centered Perspectives on Large Language Models. In Proceedings of the 2023
(2015). http://dx.doi.org/10.5241/5-85 ACM Conference on Fairness, Accountability, and Transparency (Chicago, IL, USA)
[8] Nathan Bos, Kimberly Glasgow, John Gersh, Isaiah Harbison, and Celeste (FAccT ’23). ACM, 205–216. https://doi.org/10.1145/3593013.3593989
Lyn Paul. 2019. Mental models of AI-based systems: User predictions and [26] Katy Ilonka Gero, Zahra Ashktorab, Casey Dugan, Qian Pan, James Johnson,
explanations of image classification results. In Proceedings of the Human Factors Werner Geyer, Maria Ruiz, Sarah Miller, David R. Millen, Murray Campbell,
and Ergonomics Society Annual Meeting, Vol. 63. 184–188. https://doi.org/10. Sadhana Kumaravel, and Wei Zhang. 2020. Mental Models of AI Agents in a
1177/1071181319631392 Cooperative Game Setting. In Proceedings of the 2020 CHI Conference on Human
[9] Stephen Brade, Bryan Wang, Mauricio Sousa, Sageev Oore, and Tovi Grossman. Factors in Computing Systems (Honolulu, HI, USA) (CHI ’20). ACM, 1–12. https:
2023. Promptify: Text-to-Image Generation through Interactive Prompt Explo- //doi.org/10.1145/3313831.3376316
ration with Large Language Models. In Proceedings of the 36th Annual ACM [27] Katy Ilonka Gero, Tao Long, and Lydia B Chilton. 2023. Social Dynamics of
Symposium on User Interface Software and Technology (San Francisco, CA, USA) AI Support in Creative Writing. In Proceedings of the 2023 CHI Conference on
(UIST ’23). ACM, Article 96, 14 pages. https://doi.org/10.1145/3586183.3606725 Human Factors in Computing Systems (Hamburg, Germany) (CHI ’23). ACM,
[10] Virginia Braun and Victoria Clarke. 2021. Thematic analysis: a practical guide. Article 245, 15 pages. https://doi.org/10.1145/3544548.3580782
SAGE Publications Ltd. https://us.sagepub.com/en-us/nam/thematic-analysis/ [28] Sarah Gibbons, Tarun Mugunthan, and Jakob Nielsen. 2023. Accordion Editing
book248481 and Apple Picking: Early Generative-AI User Behaviors. Retrieved 23 April, 2024
[11] Ritvik Budhiraja, Ishika Joshi, Jagat Sesh Challa, Harshal D. Akolekar, and Dhruv from https://www.nngroup.com/articles/accordion-editing-apple-picking/.
Kumar. 2024. “It’s not like Jarvis, but it’s pretty close!” - Examining ChatGPT’s [29] Kate Glazko, Yusuf Mohammed, Ben Kosa, Venkatesh Potluri, and Jennifer
Usage among Undergraduate Students in Computer Science. In Proceedings of Mankoff. 2024. Identifying and Improving Disability Bias in GPT-Based Resume
the 26th Australasian Computing Education Conference (Sydney, NSW, Australia) Screening. In Proceedings of the 2024 ACM Conference on Fairness, Accountability,
(ACE ’24). ACM, 124–133. https://doi.org/10.1145/3636243.3636257 and Transparency (Rio de Janeiro, Brazil) (FAccT ’24). ACM, 687–700. https:
[12] Raluca Budiu. 2024. Generative AI Diary. Retrieved 23 April, 2024 from https: //doi.org/10.1145/3630106.3658933
//www.nngroup.com/articles/generative-ai-diary/. [30] Kate S Glazko, Momona Yamagami, Aashaka Desai, Kelly Avery Mack, Venkatesh
[13] Nicholas Carlini, Florian Tramèr, Eric Wallace, Matthew Jagielski, Ariel Herbert- Potluri, Xuhai Xu, and Jennifer Mankoff. 2023. An Autoethnographic Case Study
Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Úlfar Erlingsson, of Generative Artificial Intelligence’s Utility for Accessibility. In Proceedings of
Alina Oprea, and Colin Raffel. 2021. Extracting Training Data from Large Lan- the 25th International ACM SIGACCESS Conference on Computers and Accessibility
guage Models. In 30th USENIX Security Symposium (USENIX Security 21). USENIX (New York, NY, USA) (ASSETS ’23). ACM, Article 99, 8 pages. https://doi.org/
Association, 2633–2650. https://www.usenix.org/conference/usenixsecurity21/ 10.1145/3597638.3614548
presentation/carlini-extracting
How Blind People Use and Understand Generative AI Tools ASSETS ’24, October 27–30, 2024, St. John’s, NL, Canada

[31] Ricardo E Gonzalez Penuela, Jazmin Collins, Cynthia Bennett, and Shiri Azenkot. [48] Ruogu Kang, Laura Dabbish, Nathaniel Fruchter, and Sara Kiesler. 2015. “My
2024. Investigating Use Cases of AI-Powered Scene Description Applications data just goes everywhere": User mental models of the internet and implications
for Blind and Low Vision People. In Proceedings of the CHI Conference on Human for privacy and security. In Proceedings of the Eleventh USENIX Conference on
Factors in Computing Systems (Honolulu, HI, USA) (CHI ’24). ACM, Article 901, Usable Privacy and Security (Ottawa, Canada) (SOUPS ’15). USENIX Associa-
21 pages. https://doi.org/10.1145/3613904.3642211 tion, USA, 39–52. https://www.usenix.org/conference/soups2015/proceedings/
[32] Steven M. Goodman, Erin Buehler, Patrick Clary, Andy Coenen, Aaron Dons- presentation/kang
bach, Tiffanie N. Horne, Michal Lahav, Robert MacDonald, Rain Breaw Michaels, [49] Majeed Kazemitabaar, Justin Chow, Carl Ka To Ma, Barbara J. Ericson, David
Ajit Narayanan, Mahima Pushkarna, Joel Riley, Alex Santana, Lei Shi, Rachel Weintrop, and Tovi Grossman. 2023. Studying the effect of AI Code Generators
Sweeney, Phil Weaver, Ann Yuan, and Meredith Ringel Morris. 2022. LaMPost: on Supporting Novice Learners in Introductory Programming. In Proceedings
Design and Evaluation of an AI-assisted Email Writing Prototype for Adults of the 2023 CHI Conference on Human Factors in Computing Systems (Hamburg,
with Dyslexia. In Proceedings of the 24th International ACM SIGACCESS Confer- Germany) (CHI ’23). ACM, Article 455, 23 pages. https://doi.org/10.1145/3544548.
ence on Computers and Accessibility (Athens, Greece) (ASSETS ’22). ACM, Article 3580919
24, 18 pages. https://doi.org/10.1145/3517428.3544819 [50] David E Kieras and Susan Bovair. 1984. The role of a mental model in learning
[33] Google. 2024. Gemini. Retrieved 12 April, 2024 from https://gemini.google.com/ to operate a device. Cognitive science 8, 3 (1984), 255–273. https://doi.org/10.
app. 1207/s15516709cog0803_3
[34] Jan Grue. 2016. The problem with inspiration porn: a tentative definition [51] Tanya Kraljic and Michal Lahav. 2024. From Prompt Engineering to Collabo-
and a provisional critique. Disability & Society 31, 6 (2016), 838–849. https: rating: A Human-Centered Approach to AI Interfaces. Interactions 31, 3 (may
//doi.org/10.1080/09687599.2016.1205473 2024), 30–35. https://doi.org/10.1145/3652622
[35] Danna Gurari, Qing Li, Abigale J. Stangl, Anhong Guo, Chi Lin, Kristen Grauman, [52] Sri H. Kurniawan, Alistair G. Sutcliffe, Paul L. Blenkhorn, and Jae Eun Shin.
Jiebo Luo, and Jeffrey P. Bigham. 2018. VizWiz Grand Challenge: Answering 2003. Mental Models of Blind Users in the Windows Environment. International
Visual Questions from Blind People. In IEEE/CVF Conference on Computer Vision Journal of Rehabilitation Research 26, 2 (June 2003), 145–147. https://doi.org/10.
and Pattern Recognition (CVPR). IEEE Computer Society, Los Alamitos, CA, USA, 1097/00004356-200306000-00011
3608–3617. https://doi.org/10.1109/CVPR.2018.00380 [53] Nimatul Lailiyah, Iswahyuni Iswahyuni, and Alies Poetri Lintangsari. 2020.
[36] Orit Hazzan and Yael Erez. 2024. Generative AI in Computer Science Educa- English Spelling Errors Committed by the Blind. Indonesian Journal of Disability
tion. In Proceedings of the 55th ACM Technical Symposium on Computer Sci- Studies 7, 1 (May 2020), 66–71. https://doi.org/10.21776/ub.ijds.2019.007.01.7
ence Education V. 2 (Portland, OR, USA) (SIGCSE 2024). ACM, 1899. https: [54] Jenny T. Liang, Chenyang Yang, and Brad A. Myers. 2024. A Large-Scale
//doi.org/10.1145/3626253.3633409 Survey on the Usability of AI Programming Assistants: Successes and Chal-
[37] Amanda Heidt. 2024. ‘Without these tools, I’d be lost’: how generative AI aids lenges. In Proceedings of the IEEE/ACM 46th International Conference on Soft-
in accessibility. Retrieved 23 April, 2024 from https://www.nature.com/articles/ ware Engineering (Lisbon, Portugal) (ICSE ’24). ACM, Article 52, 13 pages.
d41586-024-01003-w. https://doi.org/10.1145/3597503.3608128
[38] Laurie Henneborn. 2023. Designing Generative AI to Work for People with [55] Q. Vera Liao and Jennifer Wortman Vaughan. 2024. AI Transparency in the Age
Disabilities. Retrieved 23 April, 2024 from https://hbr.org/2023/08/designing- of LLMs: A Human-Centered Research Roadmap. Harvard Data Science Review
generative-ai-to-work-for-people-with-disabilities. Special Issue 5 (feb 29 2024). https://hdsr.mitpress.mit.edu/pub/aelql9qy
[39] Jonggi Hong, Jaina Gandhi, Ernest Essuah Mensah, Farnaz Zamiri Zeraati, [56] Lauren Lin and Duri Long. 2023. Generative AI Futures: A Speculative Design
Ebrima Jarjue, Kyungjun Lee, and Hernisa Kacorri. 2022. Blind Users Accessing Exploration. In Proceedings of the 15th Conference on Creativity and Cognition
Their Training Images in Teachable Object Recognizers. In Proceedings of the (Virtual Event, USA) (C&C ’23). ACM, 380–383. https://doi.org/10.1145/3591196.
24th International ACM SIGACCESS Conference on Computers and Accessibility 3596616
(Athens, Greece) (ASSETS ’22). ACM, Article 14, 18 pages. https://doi.org/10. [57] Nelson F. Liu, Tianyi Zhang, and Percy Liang. 2023. Evaluating Verifiability
1145/3517428.3544824 in Generative Search Engines. In Conference on Empirical Methods in Natural
[40] Aike C. Horstmann, Clara Strathmann, Lea Lambrich, and Nicole C. Krämer. Language Processing (EMNLP ’23). https://doi.org/10.48550/arXiv.2304.09848
2023. Alexa, What’s Inside of You: A Qualitative Study to Explore Users’ Mental [58] Duri Long and Brian Magerko. 2020. What is AI Literacy? Competencies and
Models of Intelligent Voice Assistants. In Proceedings of the 23rd ACM Interna- Design Considerations. In Proceedings of the 2020 CHI Conference on Human
tional Conference on Intelligent Virtual Agents (Würzburg, Germany) (IVA ’23). Factors in Computing Systems (Honolulu, HI, USA) (CHI ’20). ACM, 1–16. https:
ACM, Article 5, 10 pages. https://doi.org/10.1145/3570945.3607335 //doi.org/10.1145/3313831.3376727
[41] Irene Hou, Sophia Mettille, Owen Man, Zhuo Li, Cynthia Zastudil, and Stephen [59] Darren Lunn, Simon Harper, and Sean Bechhofer. 2011. Identifying Behavioral
MacNeil. 2024. The Effects of Generative AI on Computing Students’ Help- Strategies of Visually Impaired Users to Improve Access to Web Content. ACM
Seeking Preferences. In Proceedings of the 26th Australasian Computing Education ACM Transactions on Accessible Computing (TACCESS) 3, 4, Article 13 (April
Conference (Sydney, NSW, Australia) (ACE ’24). ACM, 39–48. https://doi.org/ 2011), 35 pages. https://doi.org/10.1145/1952388.1952390
10.1145/3636243.3636248 [60] Avery Mack, Rida Qadri, Remi Denton, Shaun K Kane, and Cynthia L Bennett.
[42] Yue Huang, Borke Obada-Obieh, and Konstantin (Kosta) Beznosov. 2020. Ama- 2024. “They only care to show us the wheelchair”: disability representation
zon vs. My Brother: How Users of Shared Smart Speakers Perceive and Cope in text-to-image AI models. In Proceedings of the CHI Conference on Human
with Privacy Risks. In Proceedings of the 2020 CHI Conference on Human Fac- Factors in Computing Systems (Honolulu, HI, USA) (CHI ’24). 21 pages. https:
tors in Computing Systems (Honolulu, HI, USA) (CHI ’20). ACM, 1–13. https: //doi.org/10.1145/3613904.3642166
//doi.org/10.1145/3313831.3376529 [61] Haley MacLeod, Cynthia L. Bennett, Meredith Ringel Morris, and Edward Cutrell.
[43] Mina Huh, Yi-Hao Peng, and Amy Pavel. 2023. GenAssist: Making Image 2017. Understanding Blind People’s Experiences with Computer-Generated
Generation Accessible. In Proceedings of the 36th Annual ACM Symposium on Captions of Social Media Images. In Proceedings of the 2017 CHI Conference on
User Interface Software and Technology (San Francisco, CA, USA) (UIST ’23). Human Factors in Computing Systems (Denver, Colorado, USA) (CHI ’17). ACM,
ACM, Article 38, 17 pages. https://doi.org/10.1145/3586183.3606735 5988–5999. https://doi.org/10.1145/3025453.3025814
[44] Nanna Inie, Stefania Druga, Peter Zukerman, and Emily M. Bender. 2024. From [62] Kamil Malinka, Martin Peresíni, Anton Firc, Ondrej Hujnák, and Filip Janus.
"AI" to Probabilistic Automation: How Does Anthropomorphization of Technical 2023. On the Educational Impact of ChatGPT: Is Artificial Intelligence Ready to
Systems Descriptions Influence Trust?. In Proceedings of the 2024 ACM Conference Obtain a University Degree?. In Proceedings of the 2023 Conference on Innovation
on Fairness, Accountability, and Transparency (Rio de Janeiro, Brazil) (FAccT ’24). and Technology in Computer Science Education V. 1 (Turku, Finland) (ITiCSE
ACM, 2322–2347. https://doi.org/10.1145/3630106.3659040 2023). ACM, 47–53. https://doi.org/10.1145/3587102.3588827
[45] Nanna Inie, Jeanette Falk, and Steve Tanimoto. 2023. Designing Participatory [63] Shahan Ali Memon and Jevin D. West. 2024. Search engines post-ChatGPT:
AI: Creative Professionals’ Worries and Expectations about Generative AI. In How generative artificial intelligence could make search less reliable. Retrieved
Extended Abstracts of the 2023 CHI Conference on Human Factors in Computing April 27, 2024 from https://www.cip.uw.edu/2024/02/18/search-engines-chatgpt-
Systems (Hamburg, Germany) (CHI EA ’23). ACM, Article 82, 8 pages. https: generative-artificial-intelligence-less-reliable/
//doi.org/10.1145/3544549.3585657 [64] Meta AI. 2023. Multimodal Generative AI Systems. Retreived 23 April, 2024 from
[46] JiWoong Jang, Sanika Moharana, Patrick Carrington, and Andrew Begel. 2024. https://ai.meta.com/tools/system-cards/multimodal-generative-ai-systems/.
“It’s the only thing I can trust”: Envisioning Large Language Model Use by [65] Microsoft. 2024. Copilot. Retrieved 12 April, 2024 from https://copilot.microsoft.
Autistic Workers for Communication Assistance. In Proceedings of the CHI com/.
Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI [66] Omid Mohaddesi, Noah Chicoine, Min Gong, Ozlem Ergun, Jacqueline Griffin,
’24). ACM, Article 77, 18 pages. https://doi.org/10.1145/3613904.3642894 David Kaeli, Stacy Marsella, and Casper Harteveld. 2023. Thought Bubbles: A
[47] Rie Kamikubo, Kyungjun Lee, and Hernisa Kacorri. 2023. Contributing to Proxy into Players’ Mental Model Development. In Proceedings of the 2023 CHI
Accessibility Datasets: Reflections on Sharing Study Data by Blind People. In Conference on Human Factors in Computing Systems (Hamburg, Germany) (CHI
Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems ’23). ACM, Article 729, 21 pages. https://doi.org/10.1145/3544548.3581570
(Hamburg, Germany) (CHI ’23). ACM, Article 827, 18 pages. https://doi.org/10.
1145/3544548.3581337
ASSETS ’24, October 27–30, 2024, St. John’s, NL, Canada Rudaiba Adnin and Maitraye Das

[67] Lourdes M. Morales, Sonia M. Arteaga, Peter Cottrell, and Sri Kurniawan. 2014. Intelligence in Education 33 (2023), 233–266. https://doi.org/10.1007/s40593-022-
Design Guidelines of Tools for Facilitating Blind People to Independently For- 00314-1
mat Their Documents. In Computers Helping People with Special Needs, Klaus [86] Stephanie Valencia, Richard Cave, Krystal Kallarackal, Katie Seaver, Michael
Miesenberger, Deborah Fels, Dominique Archambault, Petr Peňáz, and Wolfgang Terry, and Shaun K. Kane. 2023. “The less I type, the better”: How AI-Language
Zagler (Eds.). Springer International Publishing, Cham, 634–641. Models can Enhance or Impede Communication for AAC Users. In Proceedings
[68] Cecily Morrison, Martin Grayson, Rita Faia Marques, Daniela Massiceti, Camilla of the 2023 CHI Conference on Human Factors in Computing Systems (Hamburg,
Longden, Linda Wen, and Edward Cutrell. 2023. Understanding Personalized Germany) (CHI ’23). ACM, Article 830, 14 pages. https://doi.org/10.1145/3544548.
Accessibility through Teachable AI: Designing and Evaluating Find My Things 3581560
for People who are Blind or Low Vision. In Proceedings of the 25th International [87] Sheena Vasani. 2023. Be My Eyes AI offers GPT-4-powered support for blind
ACM SIGACCESS Conference on Computers and Accessibility (New York, NY, Microsoft customers. Retrieved 23 April, 2024 from https://www.theverge.com/
USA) (ASSETS ’23). ACM, Article 31, 12 pages. https://doi.org/10.1145/3597638. 2023/11/15/23962709/microsoft-blind-users-open-ai-chatgpt-4-be-my-eyes.
3608395 [88] Mathias Peter Verheijden and Mathias Funk. 2023. Collaborative Diffusion:
[69] Fatemeh Mosaiyebzadeh, Seyedamin Pouriyeh, Reza Parizi, Nasrin Dehbozorgi, Boosting Designerly Co-Creation with Generative AI. In Extended Abstracts of
Mohsen Dorodchi, and Daniel Macêdo Batista. 2023. Exploring the Role of the 2023 CHI Conference on Human Factors in Computing Systems (Hamburg,
ChatGPT in Education: Applications and Challenges. In Proceedings of the 24th Germany) (CHI EA ’23). ACM, Article 73, 8 pages. https://doi.org/10.1145/
Annual Conference on Information Technology Education (Marietta, GA, USA) 3544549.3585680
(SIGITE ’23). ACM, 84–89. https://doi.org/10.1145/3585059.3611445 [89] Markel Vigo and Simon Harper. 2013. Coping Tactics Employed by Visually
[70] Dan Myers and Brian P Blake. 2017. An Evaluation of the Conformed Dimensions Disabled Users on the Web. International Journal of Human-Computer Studies
of Data Quality in Application to an Existing Information Quality-Privacy-Trust 71, 11 (November 2013), 1013–1025. https://doi.org/10.1016/j.ijhcs.2013.08.002
Research Framework. (2017). https://ualr.edu/informationquality/files/2023/02/ [90] World Wide Web Consortium (W3C). 2024. WCAG 2 Overview. Retrieved 21
P26-ICIQ2017-Conformed-Dimenstions-of-DQ.pdf April, 2024 from https://www.w3.org/WAI/standards-guidelines/wcag/.
[71] Donald A. Norman. 2014. Some Observations on Mental Models. In Mental [91] Jingran Wang, Yi Liu, Peigong Li, Zhenxing Lin, Stavros Sindakis, and Sakshi
models. Psychology Press, 15–22. Aggarwal. 2023. Overview of Data Quality: Examining the Dimensions, An-
[72] National Federation of the Blind. 2024. Blindness Statistics. Retrieved 23 April, tecedents, and Impacts of Data Quality. Journal of the Knowledge Economy
2024 from https://nfb.org/resources/blindness-statistics. (2023), 1–20. https://doi.org/10.1007/s13132-022-01096-6
[73] OpenAI. 2024. ChatGPT. Retrieved 12 April, 2024 from https://openai.com/ [92] Ruotong Wang, Ruijia Cheng, Denae Ford, and Thomas Zimmermann. 2024.
chatgpt. Investigating and Designing for Trust in AI-powered Code Generation Tools.
[74] Eng Lieh Ouh, Benjamin Kok Siew Gan, Kyong Jin Shim, and Swavek Wlod- In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and
kowski. 2023. ChatGPT, Can You Generate Solutions for my Coding Exercises? Transparency (Rio de Janeiro, Brazil) (FAccT ’24). ACM, 1475–1493. https:
An Evaluation on its Effectiveness in an Undergraduate Java Programming //doi.org/10.1145/3630106.3658984
Course. (2023), 54–60. https://doi.org/10.1145/3587102.3588794 [93] Yunlong Wang, Shuyuan Shen, and Brian Y Lim. 2023. RePrompt: Automatic
[75] Konstantinos S. Papadopoulos, Evmorfia K. Arvaniti, Despina I. Dimitriadi, Prompt Editing to Refine AI-Generative Art Towards Precise Expressions. In
Vasiliki G. Gkoutsioudi, and Christina I. Zantali. 2009. Spelling performance Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems
of visually impaired adults. British Journal of Visual Impairment 27, 1 (2009), (Hamburg, Germany) (CHI ’23). ACM, Article 22, 29 pages. https://doi.org/10.
49–64. https://doi.org/10.1177/0264619608097746 1145/3544548.3581402
[76] Alisha Pradhan, Kanika Mehta, and Leah Findlater. 2018. “Accessibility Came [94] Yuxia Wang, Minghan Wang, Muhammad Arslan Manzoor, Fei Liu, Georgi
by Accident": Use of Voice-Controlled Intelligent Personal Assistants by People Georgiev, Rocktim Jyoti Das, and Preslav Nakov. 2024. Factuality of Large
with Disabilities. In Proceedings of the 2018 CHI Conference on Human Factors Language Models in the Year 2024. arXiv (2024). https://doi.org/10.48550/arXiv.
in Computing Systems (Montreal QC, Canada) (CHI ’18). ACM, 1–13. https: 2402.02420 arXiv:2402.02420
//doi.org/10.1145/3173574.3174033 [95] Laura Weidinger, Jonathan Uesato, Maribeth Rauh, Conor Griffin, Po-Sen Huang,
[77] David Redmond. 2023. This Face Looks Redacted: The Be My Eyes Faces John Mellor, Amelia Glaese, Myra Cheng, Borja Balle, Atoosa Kasirzadeh, Court-
Controversy Explained. Retrieved 23 April, 2024 from https://vi.ie/this-face- ney Biles, Sasha Brown, Zac Kenton, Will Hawkins, Tom Stepleton, Abeba
looks-redacted-the-be-my-eyes-faces-controversy-explained/. Birhane, Lisa Anne Hendricks, Laura Rimell, William Isaac, Julia Haas, Sean
[78] Matthew Rueben, Jeffrey Klow, Madelyn Duer, Eric Zimmerman, Jennifer Pia- Legassick, Geoffrey Irving, and Iason Gabriel. 2022. Taxonomy of Risks posed
centini, Madison Browning, Frank J. Bernieri, Cindy M. Grimm, and William D. by Language Models. In Proceedings of the 2022 ACM Conference on Fairness,
Smart. 2021. Mental Models of a Mobile Shoe Rack: Exploratory Findings from Accountability, and Transparency (Seoul, Republic of Korea) (FAccT ’22). ACM,
a Long-term In-the-Wild Study. ACM Transactions on Human-Robot Interaction 214–229. https://doi.org/10.1145/3531146.3533088
(THRI) 10, 2, Article 16 (feb 2021), 36 pages. https://doi.org/10.1145/3442620 [96] Meredith Whittaker, Meryl Alper, Cynthia L. Bennett, Sara Hendren, Eliza-
[79] Heleen Rutjes, Martijn Willemsen, and Wijnand IJsselsteijn. 2019. Consid- beth Kaziunas, Mara Mills, Meredith Ringel Morris, Joy Lisi Rankin, Emily
erations on Explainable AI and Users’ Mental Models. In CHI 2019 Work- Rogers, Marcel Salas, and Sarah Myers West. 2019. Disability, Bias & AI Report.
shop: Where is the Human? Bridging the Gap Between AI and HCI. ACM, Inc. Technical Report. AI Now Institute. https://ainowinstitute.org/publication/
https://www.martijnwillemsen.nl/recommenderlab/RutjesChHI2019ws.pdf disabilitybiasai-2019
[80] Abir Saha and Anne Marie Piper. 2020. Understanding Audio Production Prac- [97] J.D. Zamfirescu-Pereira, Richmond Y. Wong, Bjoern Hartmann, and Qian Yang.
tices of People with Vision Impairments. In Proceedings of the 22nd International 2023. Why Johnny Can’t Prompt: How Non-AI Experts Try (and Fail) to Design
ACM SIGACCESS Conference on Computers and Accessibility (Virtual Event) (AS- LLM Prompts. In Proceedings of the 2023 CHI Conference on Human Factors in
SETS ’20). ACM, Article 36, 13 pages. https://doi.org/10.1145/3373625.3416993 Computing Systems (Hamburg, Germany) (CHI ’23). ACM, Article 437, 21 pages.
[81] Filipo Sharevski and Aziz N Zeidieh. 2023. “I Just Didn’t Notice It:” Experiences https://doi.org/10.1145/3544548.3581388
with Misinformation Warnings on Social Media amongst Users Who Are Low [98] Zhiping Zhang, Michelle Jia, Hao-Ping (Hank) Lee, Bingsheng Yao, Sauvik Das,
Vision or Blind. In Proceedings of the 2023 New Security Paradigms Workshop Ada Lerner, Dakuo Wang, and Tianshi Li. 2024. “It’s a Fair Game”, or Is It?
(Segovia, Spain) (NSPW ’23). ACM, 17–33. https://doi.org/10.1145/3633500. Examining How Users Navigate Disclosure Risks and Benefits When Using LLM-
3633502 Based Conversational Agents. In Proceedings of the CHI Conference on Human
[82] Renee Shelby, Shalaleh Rismani, Kathryn Henne, AJung Moon, Negar Ros- Factors in Computing Systems (Honolulu, HI, USA) (CHI ’24). ACM, Article 156,
tamzadeh, Paul Nicholas, N’Mah Yilla-Akbari, Jess Gallegos, Andrew Smart, 26 pages. https://doi.org/10.1145/3613904.3642385
Emilio Garcia, and Gurleen Virk. 2023. Sociotechnical Harms of Algorithmic [99] Yong Zheng. 2023. ChatGPT for Teaching and Learning: An Experience from
Systems: Scoping a Taxonomy for Harm Reduction. In Proceedings of the 2023 Data Science Education. In Proceedings of the 24th Annual Conference on Infor-
AAAI/ACM Conference on AI, Ethics, and Society (Montréal, QC, Canada) (AIES mation Technology Education (Marietta, GA, USA) (SIGITE ’23). ACM, 66–72.
’23). ACM, 723–741. https://doi.org/10.1145/3600211.3604673 https://doi.org/10.1145/3585059.3611431
[83] Marita Skjuve, Asbjørn Følstad, and Petter Bae Brandtzaeg. 2023. The User [100] Jiawei Zhou, Yixuan Zhang, Qianni Luo, Andrea G Parker, and Munmun
Experience of ChatGPT: Findings from a Questionnaire Study of Early Users. In De Choudhury. 2023. Synthetic Lies: Understanding AI-Generated Misinforma-
Proceedings of the 5th International Conference on Conversational User Interfaces tion and Evaluating Algorithmic and Human Solutions. In Proceedings of the 2023
(Eindhoven, Netherlands) (CUI ’23). ACM, Article 2, 10 pages. https://doi.org/ CHI Conference on Human Factors in Computing Systems (Hamburg, Germany)
10.1145/3571884.3597144 (CHI ’23). ACM, Article 436, 20 pages. https://doi.org/10.1145/3544548.3581318
[84] Yuan Sun, Eunchae Jang, Fenglong Ma, and Ting Wang. 2024. Generative AI
in the Wild: Prospects, Challenges, and Strategies. In Proceedings of the CHI
Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI
’24). ACM, Article 747, 16 pages. https://doi.org/10.1145/3613904.3642160
[85] David Touretzky, Christina Gardner-McCune, and Deborah Seehorn. 2023. Ma-
chine Learning and the Five Big Ideas in AI. International Journal of Artificial