Generating Accessibility: Using AI to Improve Higher Education for the

Visually Impaired

Sophie Schauer

, Katharina Simbeck

and Leon Sch

ußler

School of Computing, Communication and Business, HTW University of Applied Sciences Berlin,

Wilhelminenhofstraße, Berlin, Germany

{sophie.schauer, katharina.simbeck}@htw-berlin.de, leon.schuessler@student.htw-berlin.de

Keywords:

AI for Good, Inclusive Higher Education, Accessibility.

Abstract:

Visually impaired students face many obstacles in higher education, including access to adapted course ma-

terial. While accessibility is a legal requirement in many countries, its implementation is linked to practical

challenges for educators. In this paper, we discuss the adaptation of existing course materials for blind stu-

dents in a case study and propose using generative Artiﬁcial Intelligence to facilitate the creation of accessible

educational materials, contributing to a more inclusive learning environment. We introduce and evaluate the

tool TRAIL to generate accessible textual descriptions of lecture slides.

1 INTRODUCTION

In higher education, visually impaired students en-

counter barriers that hinder their access to and intake

of information (Croft, 2020). These challenges, rang-

ing from inaccessible materials to barriers in content

delivery, highlight the need for innovative solutions

to promote inclusive education (Firat, 2021). Arti-

ﬁcial intelligence (AI) brings an opportunity to ad-

dress these barriers and redeﬁne the educational ex-

periences of visually impaired students (Mohamady

et al., 2023).

Visual impairment generally denotes a spectrum

of reduced vision, ranging from mild to severe, where

individuals may beneﬁt from visual aids or adapta-

tions (Kim et al., 2021). Blindness, on the other hand,

implies a total loss of vision, often necessitating al-

ternative means of learning and interacting with the

environment, such as braille, auditory cues or assis-

tive technology in the form of screen readers. Screen

readers read content from a computer screen to the

user for them to comprehend written text as well as

graphical visualisations (Sharif et al., 2021). Recog-

nising these distinctions is essential for tailoring sup-

port and accommodations to meet the speciﬁc needs

of individuals along the spectrum of visual challenges

(Permvattana et al., 2013).

However, inaccessible documents are only part of

https://orcid.org/0009-0006-3350-7803

https://orcid.org/0000-0001-6792-461X

https://orcid.org/0009-0003-4126-3392

the problem. Accessible higher education extends

beyond providing accessible documents but also in-

cludes accessible campus facilities and external ﬁnan-

cial support. Environmental barriers, such as access

to campus or inadequate accommodations, add to the

obstacles they encounter (Bishop and Rhind, 2011).

The paper’s focus narrows down to the challenges sur-

rounding accessible documents for visually impaired

individuals and brings attention to existing barriers.

As Kim et al. (2021) argue, there is a need for

scalable, low-cost methods for making documents

with visualisations accessible.

This paper explores the possibilities of ”generat-

ing accessibility” and ﬁnds innovative strategies and

technologies to allow visually impaired students to

navigate educational contexts with greater autonomy.

Generative AI refers to technologies that generate im-

ages, videos, sounds or, in our case, texts based on

models learned from high volumes of training data.

Generative AI has the potential to mitigate accessibil-

ity challenges by changing how educational content

is created, adapted and delivered (Leiker et al., 2023).

The inclusivity and accessibility for visually impaired

students are discussed, beginning with fundamental

principles of the right to accessibility and the current

need for accessibility in educational settings. Inclu-

sivity should not be viewed as an additional consid-

eration but as an integral part of the initial solution.

Embracing inclusivity from the start ensures that di-

verse needs are inherently addressed, regardless of

background or ability (Heyer, 2021).

Schauer, S., Simbeck, K. and Schüßler, L.

Generating Accessibility: Using AI to Improve Higher Education for the Visually Impaired.

DOI: 10.5220/0013136300003932

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Computer Supported Education (CSEDU 2025) - Volume 1, pages 15-25

ISBN: 978-989-758-746-7; ISSN: 2184-5026

To provide a practical context, the paper presents

a prototype developed for parsing inaccessible docu-

ments into a screen-readable format. Ultimately, our

study contributes to the discourse on enhancing acces-

sibility in education. It advocates for integrating AI

as a tool to empower visually impaired students with

equitable access to educational resources, as well as

giving lecturers a tool for producing accessible mate-

rial.

2 ACCESSIBILITY IN HIGHER

EDUCATION

2.1 Right to Accessibility

The right to inclusive education is recognised as a

human right, particularly for people with disabilities.

The United Nations Convention on the Rights of Per-

sons with Disabilities (CRPD) from 2006 mandates

an inclusive education system at all levels, emphasis-

ing the importance of learning together regardless of

different learning styles and rates (United Nations, n

d). In Article 24 on education in the CRPD, the right

of persons with disabilities to education is outlined,

highlighting the principles of non-discrimination and

equal opportunity. The article calls for the establish-

ment of inclusive education systems at all levels with

a focus on the full development of human potential,

respect for human rights and effective participation in

society. Individuals with disabilities should not be ex-

cluded from the general education system based on

their disability. It further addresses speciﬁc measures

for academic and social development, such as the fa-

cilitation of learning braille, alternative communica-

tion modes and sign language with qualiﬁed teachers

and professionals (United Nations, n d).

The CRPD shifts with Article 24 towards inclu-

sion in response to the history of segregation and

marginalisation of students with disabilities, to op-

pose discriminatory attitudes and to create an inclu-

sive society. However, there is a signiﬁcant gap be-

tween the aspiration of the CRPD and the global real-

ity of exclusion that many students with disabilities

are still facing (Heyer, 2021). Rethinking the pur-

pose of education and rebuilding education systems is

necessary to emphasise diversity as a positive aspect

(De Beco, 2014).

2.2 Accessibility for Visually Impaired

Students

The need for accessibility and inclusion in education

for visually impaired students originates from the fun-

damental principles of ensuring equal rights and op-

portunities for people with disabilities, as stated in the

United Nations CRPD (MacKay, 2006). The conven-

tion recognises the importance ”to promote, protect

and ensure the full and equal enjoyment of all human

rights and fundamental freedoms by all persons with

disabilities, and to promote respect for their inherent

dignity.”

Any limitations in accessibility to education for

visually impaired students imply barriers to their in-

clusion and participation in society and everyday life.

Barriers that visually impaired students face when en-

rolling at university can be grouped into four cate-

gories: attitudinal, institutional, environmental and

physical (Hutchinson et al., 1998). Attitudinal re-

lates ”to the attitudes of key individuals with whom

a visually impaired student will interact” (Bishop and

Rhind, 2011). Individuals, such as parents and insti-

tutional staff, play an important role in the academic

journey of visually impaired students, and their atti-

tudes can signiﬁcantly shape the student’s experience

and development in higher education. Institutional

barriers refer to the accessibility of educational re-

sources and materials. Further, environmental factors

can include access to buildings, classrooms and ac-

commodation, which can be overcome through lifts,

automatic doors, appropriate signs and assistive tech-

nology. Physical challenges, such as headaches and

tiredness after learning periods, must be managed by

the students themselves (Bishop and Rhind, 2011).

Educating staff members is necessary to best sup-

port students with visual impairments, including clear

communication of the student’s disability and provi-

sion of adequate materials (Bishop and Rhind, 2011).

Similar ﬁndings are backed up by (Giese et al., 2022).

They state that most barriers for students with vi-

sual impairments stem from a lack of qualiﬁed teach-

ers and inaccessible learning materials as a result

of short-term lesson planning. Thurston (2014) ad-

vocates for developing ”visually impaired literacy

skills” that would allow real-time access to educa-

tional material for inclusive learning and teaching

(Thurston, 2014).

Incorporating information and communication

technology can help students with visual impairments

access information independently and foster the de-

velopment of new skills at a similar pace as students

without disabilities. Bradea and Bl

adel (2017) stud-

ied the training level and attitude of teachers toward

CSEDU 2025 - 17th International Conference on Computer Supported Education

using ICT for educational purposes for increased ac-

cessibility. Their research focused on students in Ro-

mania and found that the majority of teachers lacked

the required skills and knowledge on how to use or

adapt educational software to the scholar curriculum

for visually impaired students. Their result showed

that only 3% of interviewed teachers believed to have

high-level skills to support school inclusion of stu-

dents with visual impairments.

Vanderheiden (2008) coined the term ”ubiquitous

accessibility” to describe the ability of people with

disabilities to access and use the same mainstream

technologies. In this context, the concept of ”plug-

gable interfaces” for computing devices is introduced.

Pluggable interfaces adapt to a user’s disability and

are adjusted to the user’s proﬁle, interaction history

and the speciﬁc kind of disability. A study found

that the goal of ubiquitous accessibility can only be

achieved through consistent screen reading experi-

ences across different devices, applications and op-

erating systems (Billah et al., 2017).

A comprehensive study review investigated fac-

tors for the academic success of students with vi-

sual impairments (Simui et al., 2018). Challenges

were identiﬁed in inﬂexible time constraints for as-

sessments, insufﬁcient availability of adaptive tech-

nologies, technical issues with e-learning platforms,

difﬁculties connecting to websites and course man-

agement systems and improper use of e-learning tools

by lecturers. Problems that the students mentioned

were incomplete online course notes, delay of mate-

rial provision on the website and pacing of in-class

presentation (Fichten et al., 2009). Consequently, stu-

dents with visual impairments are less likely to attend

courses in higher education (Athanasios et al., 2009).

Factors for success emerging from a study were a pos-

itive attitude, self-advocacy and inventiveness (Simui

et al., 2018).

2.3 Accessibility of Documents Used in

Higher Education

Students with visual impairments use assistive tech-

nology to navigate educational material. While fully

blind students might use screen or braille readers to

narrate on-screen text, those with low vision can use

magnifying software to enlarge documents. Graphi-

cal material, particularly diagrams, images and visual

cues, on the other hand, are often more difﬁcult or im-

possible to access (Armstrong and Murray, 2007). In

many disciplines, these kinds of materials, however,

make up an integral part of the studies. It has been

discovered that in response, vision-impaired students

skip graphical material and ultimately are less likely

to enrol in STEM disciplines in comparison to arts

and humanities (Butler et al., 2017).

Visualisations are used frequently in higher edu-

cational classes to communicate concepts, create va-

riety in study materials and help students keep and

regain attention during learning (Naps et al., 2002).

Visualisations can play a major role in education

to help students create mental models of scientiﬁc

concepts(Rapp, 2005; Vavra et al., 2011), course

content or learning progress (Tervakari et al., 2014;

Vieira et al., 2018). In biology, they help stu-

dents understand structures on a cellular and sub-

cellular level that cannot be observed without mi-

croscopes (Jenkinson, 2018). In chemistry, under-

standing of atomic or molecular structures can be en-

hanced using visualisations, experimental set-ups are

communicated in sketches, experimental data is plot-

ted, and simulations help understand molecular pro-

cesses (Jones et al., 2005; Burewicz and Miranow-

icz, 2002). In computer science education, visualisa-

tions to illustrate algorithms or data structures (Naps

et al., 2002; Fouh et al., 2012). Educational visu-

alisations fulﬁl several functions (Jenkinson, 2018),

such as schematic representations that point students

to the central concepts to be understood, diagrams

that display data, and ﬂow schemes that describe

dynamic processes and temporal order. However,

even learners without visual impairments sometimes

lack visual literacy to correctly interpret scientiﬁc im-

ages and understand spatial and temporal dependen-

cies (Jenkinson, 2018; Jones et al., 2005). Visual-

isations can also replace educational experiments if

those are too costly, time-consuming or dangerous to

be executed in classrooms (Burewicz and Miranow-

icz, 2002). Several researchers thus argue that educa-

tional visualisation needs to be accompanied by tex-

tual explanations to help learners interpret the images

and understand the underlying concepts (Vavra et al.,

2011; Naps et al., 2002). Beyond education, domain-

speciﬁc visualisations are commonly used, especially

in the natural sciences, to communicate data or exper-

imental results.

Three main methods for making graphical ma-

terial accessible to visually impaired students were

identiﬁed: enlargement, description and tactile graph-

ics (Butler et al., 2017). Depending on the student’s

level of vision impairment, enlarging graphics might

not be sufﬁcient. While descriptions give students

a general idea, creating an accurate mental model

of layout and content is difﬁcult. Tactile graphics

provide students with haptic drawings through raised

lines. However, this requires ﬁnancial resources, time

and tactile reading skills from the student (Butler

et al., 2017).

Generating Accessibility: Using AI to Improve Higher Education for the Visually Impaired

To provide equal opportunities to visually im-

paired students, visualisations used in class need to be

either described textually or provided as haptic expe-

riences (Jones et al., 2006; Jung et al., 2021). Alterna-

tive text for educational visualisations should enable

visually impaired learners to form visual mental mod-

els (Jung et al., 2021). Visually impaired users of alt

texts to diagrams emphasised the need for a clear de-

scription of chart type, axes including range and data

trend (Jung et al., 2021). They also preferred to learn

about the colours used and longer texts and complete

texts over shorter texts.

The need for accessible scientiﬁc PDF ﬁles was

addressed in a study (Wang et al., 2021). Scientiﬁc

papers in PDF format are often unusable on screen

readers and, therefore, inaccessible to visually im-

paired students. SciA11y was proposed as a solution

for extracting and rendering semantic content of PDFs

as an accessible HTML document (Wang et al., 2021).

There is a large number of recommendations for

accessible documents (Kim et al., 2021; Jung et al.,

2021).

Alongside graphical material, the document lay-

out can pose accessibility barriers to students with

visual impairments when using assistive technology.

New approaches have been tested with added spatial

information and extra metadata, including informa-

tion regarding the visual structure. As Moured et al.

(2023) state, AI-based image-processing technologies

can format accessible content automatically through

the extraction and analysis of pixel information. They

developed and evaluated a 2D interface for tactile dis-

plays that uses an annotated data set and real-time ob-

ject detection. Users could ”arrange a suitable reading

order based on their needs” and make the overall pro-

cess of creating an accessible document lightweight

and time-efﬁcient.

Three key criteria need to be met for making doc-

uments and presentations accessible: an appropriate

reading order, alternative text for non-text objects and

grouping of related graphical objects (Ishihara et al.,

2006). Further, decorative elements commonly found

on presentation slides must be appropriately marked

for exclusion by screen-reading software. These men-

tioned guidelines and requirements can easily be in-

tegrated using standard software platforms, such as

PowerPoint.

The German Association of the Blind and Visually

Impaired in Study and Work has released a guide for

the creation of accessible PowerPoint presentations.

They highlight the importance of using predeﬁned

slide layouts that let users of screen readers navigate

through text and graphics. They further stress the us-

age of titles for each slide since they indicate the most

important point of orientation for screen readers. Ad-

ditionally, they advise using alternative text for com-

plex tables in which the key message is mentioned

(DBSV e. V., n d).

3 METHODOLOGY

The enrolment of a student with total blindness in a

business computing bachelor’s degree program un-

veiled a critical need to assess and improve the ac-

cessibility of educational materials. This revela-

tion prompted the conduction of a case study to de-

crease the barriers to accessibility in higher education,

speciﬁcally of existing documents. It focused particu-

larly on the documents’ compatibility with commonly

used screen readers and the development of an AI tool

that parses documents into an accessible format.

Our approach to the case study involved three

main steps in a participatory design process:

1. Material Assessment. The ﬁrst phase entailed a

detailed examination of the existing educational

content. This review aimed to understand how

these materials interacted with screen readers and

identify areas of incompatibility or difﬁculty.

2. Technology Integration. In response to the chal-

lenges identiﬁed, we explored AI-based solutions,

particularly those leveraging generative AI mod-

els. This step involved assessing the potential of

these technologies to create more accessible ver-

sions of educational content tailored to the needs

of visually impaired students.

3. Evaluation. The prototype was iteratively reﬁned

with human oversight for validation, prompt en-

gineering to enhance the AI response consistency

and user feedback from visually impaired individ-

uals to assess usability and identify areas for im-

provement.

The following chapter will detail how the material

was assessed, and the ﬁndings were translated into the

development and deployment of the AI tool.

4 IMPLEMENTATION

4.1 Material Assessment

In the business computing program in question, stu-

dents are typically provided with lecture notes, usu-

ally slides, in PDF format. Lecturers also use stan-

dard software such as integrated development envi-

ronments in programming courses, the enterprise re-

source planning (ERP) software SAP and the statistics

CSEDU 2025 - 17th International Conference on Computer Supported Education

package R. A ﬁrst analysis revealed that the standard

software solutions were found to be accessible, but

not the lecture notes.

Lecturers in the bachelor’s degree program in

business computing were interviewed about the use

of domain-speciﬁc visualisations. As a result, sev-

eral diagram types were identiﬁed as relevant for

business computing studies. In software engineer-

ing and object-oriented programming courses, class

diagrams are used to visualise relationships between

classes as well as their functions and attributes using

the widely established Uniﬁed Modelling Language

(UML). Business process model notation (BPMN) di-

agrams depict the ﬂow of events and activities in busi-

ness processes. In several classes, graphical represen-

tations of functions are used, such as the linear cost

function in cost accounting or linear regression func-

tions in statistics.

As UML diagrams can be viewed as a hierarchical

representation of textual elements, they can be made

accessible to blind students more easily. BPMN di-

agrams, however, contain a variety of symbols with

different meanings, elements are ordered in so-called

”swim lanes” to represent sub-organisational areas of

responsibility, while the arrangement of elements rep-

resents the order of process ﬂow. As BPMN diagrams

are a major tool in business computing and are more

difﬁcult to describe textually, the study program used

tactile printing for an exemplary BMPN diagram.

To assess the current state of accessibility in the

presentations and documents used at our institution,

we initially tested an exemplary presentation slide

(Figure 1) in both the open-source software PAVE

and closed-source software Eye-Able for evaluation.

These applications are designed to analyse PDF struc-

tures, identifying key elements such as alternative

texts for images, logical structural trees that dictate

reading order and the marking of non-essential ele-

ments as decorative.

The analysis revealed a signiﬁcant deﬁcit in ac-

cessibility: a majority of the educational materials

utilised were almost entirely inaccessible for screen

readers and, therefore, for blind students. Many tables

were saved as graphics and code snippets or mathe-

matical formulas were inserted as screenshots instead

of markdown or LaTeX. In addition, even if diagrams

or plain text were readable with a screen reader, the

missing structure meant they were not easily com-

prehensible or otherwise made navigation difﬁcult for

screen readers. Lecturers also used unreadable graph-

ics to transfer information. For example, the course

structure, deadlines and grading criteria could have

been presented tabularly. We primarily attributed this

to the lack of consideration for the key elements dur-

Figure 1: This exemplary slide showcases a Linear Regres-

sion graph with relevant formulas on the right side. The

graphic contains decorative elements, a graph and formulas

for intelligibility. For screen reader users, neither the read-

ing order nor the difference between decorative elements

and those for comprehensibility is clear.

ing the initial creation of the documents, namely an

appropriate reading order, alternative text and group-

ing of related graphical objects.

The lecturers were surprised to learn that not only

slides with visualisations were unusable for the blind

student but also slides with mainly textual elements

that only made sense with a speciﬁc reading order.

While a typical course often provided hundreds of

slides as lecture notes, the material used only in the

six courses of the ﬁrst semester added up to dozens

of documents with hundreds of pages. The lack of at-

tention to accessibility during course creation led to

the need to quickly improve the accessibility of many

documents created iteratively over several years.

To solve these deﬁcits, initially, Microsoft Pow-

erPoint’s built-in accessibility tab was considered to

provide a list of images without alternative texts

where a structure or reading sequence was missing

or where elements were not marked as decorative.

Despite its limitations, the built-in suggestions indi-

cated that manually making the material accessible

would require signiﬁcant time. It became apparent

that recreating the material entirely while incorporat-

ing key accessibility principles from the outset would

be more time-efﬁcient. This led to the question of how

the pre-existing course materials could be adapted and

made available to blind students quickly.

The main requirements for the envisioned solution

were:

• Automatically provide textual representations of

existing course material in PDF format

• Create well-structured texts with references to

original slides (heading, slide number)

Generating Accessibility: Using AI to Improve Higher Education for the Visually Impaired

• Handle different types of visualisations

4.2 Technology Integration

Machine learning approaches have made documents

accessible to visually impaired and blind users. Choi

et al. (2019) built a browser extension using deep

learning technology to provide textual descriptions of

charts, including chart type and data tables. Lundgard

and Satyanarayan (2021) propose a four-level seman-

tic model for the usefulness of descriptions generated

using natural language processing (NLP).

To meet the requirements described in the previ-

ous section, the use of large language models (LLMs),

speciﬁcally the generative pre-trained transformers

(GPT) provided by OpenAI, was explored to pro-

cess these documents and generate screen-reader-

accessible alternative texts in a separate ﬁle. While

this approach is fast and scalable, there are several

challenges:

1. Accurate Text Generation. The LLM must ac-

curately interpret and convert the content of the

original document into alternative text. This in-

cludes not just the textual content but also the cor-

rect and complete interpretation of charts, graphs

and other non-textual elements.

2. Contextual Understanding. The LLM should

not lose the document’s context and subject matter

in the generation process. This is crucial for en-

suring that the alternative text accurately reﬂects

the intended message and information of the orig-

inal content.

3. Avoid Hallucinations: The LLM must not make

up content not part of the document.

4. Built for Use with Screen Readers. The alterna-

tive text needs to be optimised for screen reader

software. This involves structuring the text in

an easily navigable and understandable way when

read aloud by these tools.

The output format of choice is HTML to ensure

compatibility with screen readers. This format fa-

cilitates the creation of a structured reading experi-

ence through language, header, and paragraph tags.

Such structures allow visually impaired students to ef-

ﬁciently navigate the document using a screen reader,

particularly if sections are demarcated with descrip-

tive headers. For instance, a header using the <h1>-

Tag like ”Sequential Diagrams” is markedly more in-

formative than a vague designation such as ”Page 2”.

The tool was implemented as a web application using

the Flask framework, Python and the OpenAI GPT-

API.

The PDF documents were handled using the

libraries PyPDF2 and pdf2image. PyPDF2’s

PdfReader class was instrumental in reading and

manipulating the content of PDF ﬁles. Meanwhile,

pdf2image allowed the conversion of PDF pages into

images—a crucial step in processing the documents

for the intended use case.

The proof-of-concept prototype TRAIL

inte-

grates the GPT-4 Vision API to process images of the

PDF slides. The Vision API is parameterised for low

temperature to moderate the creativity level of GPT-4

during the generation process, ensuring the produced

text remains closely aligned with the content and con-

text of the original PDF document. The generated text

is then compiled into an HTML document, optimised

for compatibility with screen readers.

Speciﬁcally, the prototype disassembles the PDF,

segregating each page into an individual image. Sub-

sequently, each image is processed through the GPT-4

Vision API with a prompt to generate a detailed al-

ternative text and an appropriate title for each page.

Upon process completion, these responses are me-

thodically assembled into an HTML document. This

document is structured with a distinct header, repre-

senting each page of the original PDF, followed by the

corresponding alternative text for the page as a para-

graph, enhancing the navigability and accessibility of

the content for visually impaired users. Mathemati-

cal formulas are written as LaTeX in the HTML ﬁle

for correct parsing by screen readers. The processed

HTML text for the exemplary slide from Fig. 1 is

shown in Fig. 2, followed by the ﬁnal LaTeX result.

Resulting LaTeX

The slide is titled ”Simple

Case: Linear Regression With One Input” and

is divided into two sections. On the left, there

is a scatter plot with the x-axis labelled as ’x’

and the y-axis labelled as ’y’. The plot shows

numerous data points marked with blue ’x’ sym-

bols. A red dashed line passes through the data

points, indicating the best-ﬁtting line for the lin-

ear regression. Two annotations on the graph in-

clude ”β

= y-intercept” at the point where the

line crosses the y-axis and ”β

= slope” along

the line, indicating the slope of the best-ﬁtting

line. The linear regression model is given by

y = β

+ β

On the right side of the slide, there are sev-

eral mathematical equations and deﬁnitions re-

lated to linear regression. The deﬁnition of

residual e

is provided as e

= y

− (β

+ β

The sum of squared residuals S is deﬁned as S =

https://github.com/AltarisV/TRAIL-PDF

CSEDU 2025 - 17th International Conference on Computer Supported Education

Figure 2: HTML text for the slide shown in Fig. 1 generated with TRAIL.

∑

i=1

∑

i=1

− β

)

. Partial deriva-

tives for S with respect to β

and β

are given

∂S

∂β

= −2

∑

i=1

− β

) = 0 and

∂S

∂β

−2

∑

i=1

− β

) = 0, respectively.

The equations for solving for β

and β

are

also provided. For β

, the equation is β

∑

i=1

− β

∑

i=1

which simpliﬁes to β

= ¯y −

¯x. For β

, the equation is β

∑

i=1

− ¯x)(y

− ¯y)

∑

i=1

− ¯x)

where ¯x and ¯y represent the mean values of x and

y, respectively.

Figure 3: LaTeX for the slide shown in Fig. 1 gener-

ated with TRAIL

5 EVALUATION

The evaluation process focused on the effectiveness of

GPT-4 for rendering educational materials in accessi-

ble formats. The approach was guided by iterative re-

ﬁnement and critical examination of the AI-generated

outputs and consisted of a few main focus points:

1. Human Oversight. Human oversight was a vi-

tal part of the evaluation process to validate AI-

generated outputs and correct signiﬁcant errors.

This step was essential in ensuring the quality and

accuracy of the ﬁnal accessible documents.

2. Prompt Engineering. The evaluation also fo-

cused on reﬁning prompt engineering techniques

to enhance the consistency and relevance of AI re-

sponses. We experimented with different prompt

structures and formulations to optimise the AI’s

performance.

3. User Feedback. We gathered feedback from end-

users, including lecturers and a blind student, to

assess the usability and understandability of the

produced accessible documents. This input was

crucial in identifying areas for improvement and

validating AI technologies in educational settings.

The tool underwent the testing and evaluation pro-

cess using a variety of slides sourced from business

computing classes. The interpretation was possi-

ble for conventional text slides and worked also for

mathematical formulas and LaTeX expression. How-

ever, the technology had errors distinguishing relevant

from irrelevant content, making human oversight nec-

essary. The requirement for manual intervention to

correct signiﬁcant errors, such as incorrect semantic

interpretations or inaccuracies in graph descriptions,

remains paramount. Notably, the frequency and na-

ture of these errors depend on the input material’s

quality, with common challenges arising from low-

resolution images or misinterpreting ﬁne details in di-

agrams used in the slides.

An illustrative instance of this was observed dur-

ing our experiments with an image of quadratic equa-

tions employing the quadratic formula, a concept

prevalent in German-speaking regions, titled ”PQ-

Formel”. The AI’s response to the GPT-API accu-

rately recognised the formula’s application in solving

quadratic equations. However, the generated LaTeX

code for the formulas was erroneous. Once we re-

moved the title ”PQ-Formel” from the image, we got

ﬂawless explanations and LaTeX generation, albeit

without any reference to the quadratic formula.

A second step was iteratively testing the tool with

the blind student, which played a pivotal role in re-

ﬁning its effectiveness and accessibility. Immediate

implementation of the feedback from the blind stu-

dent was crucial to enhance the tool’s performance

and address usability concerns speciﬁc to visually im-

paired users. This process identiﬁed areas for im-

provement and made necessary adjustments to ensure

that the tool could accurately interpret and present ed-

ucational content in an accessible format.

Furthermore, an integral aspect of the evaluation

methodology was the process of prompt engineering,

which played a crucial role in reducing randomness

and minimising errors in the responses generated by

the GPT-API. Extensive experimentation with differ-

ent structures and phrasings of prompts revealed sig-

niﬁcant variations in the results. These differences

were often attributed to subtle changes or omitting

certain prompt words.

For instance, the prompt snippet ”short and pre-

cise alternative text” was interpreted literally by the

GPT model, typically yielding brief descriptions, of-

ten just one to two sentences long, for complex dia-

grams. However, modifying the snippet to ”precise

Generating Accessibility: Using AI to Improve Higher Education for the Visually Impaired

alternative text” resulted in more comprehensive and

clearer responses. This alteration also helped elimi-

nate the tendency for extraneous elaboration, encoun-

tered in earlier tests with different formulations.

To further streamline the consistency of the AI re-

sponses, the GPT-API’s temperature setting was ad-

justed to 0.2. This modiﬁcation signiﬁcantly reduced

the range of variability in the responses, aligning them

more closely with our speciﬁc requirements for alter-

native text generation (Open AI, n d).

To summarise, prioritising initial content genera-

tion, followed by evaluative revision, has proven to

be a time-efﬁcient approach, aligning with our pri-

mary objective of enhancing document accessibility

for educational purposes. This was reafﬁrmed by the

feedback from the student, who emphasised the im-

portance of well-structured and complete content and

was happy to learn with a document after such evalu-

ative revision.

6 DISCUSSION

AI holds considerable promise in enhancing the ac-

cessibility of educational materials for visually im-

paired students. As highlighted earlier, a crucial el-

ement of accessible documents is providing alterna-

tive text for graphical content. Utilising common

language models like Chat-GPT offers a viable solu-

tion for generating comprehensive image descriptions

(Johnson et al., 2023).

Moreover, AI can generate detailed descriptions

for complex diagrams and structured textual content

(He and Deng, 2017; Bernardi et al., 2016). This

functionality aids in parsing information presented vi-

sually and gives visually impaired students a more

comprehensive understanding of the material.

Engaging with content in code snippets or tables

poses a challenge for those with visual impairments.

Optical character recognition (OCR) within AI sys-

tems can be a valuable technology (Sheela et al.,

2023). Converting code and tabular data into acces-

sible formats can be an assistive tool.

Other approaches focused on accessible slide-

creation solutions where visually impaired users

could complete slide reading and authoring tasks in-

dependently through a multimodal interaction board

(Zhang et al., 2023). The developed A11yBoard was

an extension for Google Slides and gave visually im-

paired users an inclusive interface to navigate and

modify slides themselves.

Similar to our ﬁndings, Glazko et al. (2023) val-

ued generative AI technology as a useful tool but

found it required ”signiﬁcant human involvement and

iteration, rather than an out-of-the-box end-to-end so-

lution for [their] access needs” .

6.1 Learnings

AI technologies have demonstrated signiﬁcant efﬁ-

cacy in rendering previously inaccessible documents

usable for educational purposes. Advanced iterations

of GPT, such as GPT-4, exhibit remarkable proﬁ-

ciency in extracting and elucidating information from

conventional university lecture slides. This capabil-

ity extends to interpreting and generating mathemati-

cal formulas and LaTeX expressions facilitated by the

extensive training data used for these models.

During the material assessment, it also became

apparent that each course brings different require-

ments to accessibility. Many mathematical formulas

were printed as an image instead of markdown or La-

TeX and were therefore impossible for screen read-

ers to interpret. Diagrams to depict business cases

in business process modelling notation (BPMN) were

saved as screenshots. The same applied to accounting

classes where t-accounts or booking rates were pic-

tured. Most of these cases were unique to business

computing classes but showcased the need for further

material evaluation in other study programs.

Nonetheless, the journey towards full automation

in document accessibility remains incomplete. Cur-

rent large language models, including GPT-4, are not

infallible and exhibit limitations, particularly in im-

age comprehension. The ability of these AI models

to distinguish between relevant and irrelevant content

is still being reﬁned, necessitating human oversight

during the conversion process.

One key learning from this case study is the reali-

sation that, in many cases, it is more efﬁcient to recre-

ate slides entirely with accessibility in mind rather

than attempting to retroﬁt existing slides to meet ac-

cessibility standards. This highlights the importance

of incorporating accessibility principles from the out-

set of slide creation, ensuring that educational materi-

als are inherently accessible to all users.

6.2 Limitations

Our study was bound in its focus on converting a lim-

ited number of existing slides into an accessible for-

mat. The tool’s evaluation was restricted to business

computing classes and limited in broader applicability

across different educational settings and disciplines.

Further research should be done to evaluate TRAIL’s

effectiveness in other study ﬁelds.

Additionally, the development and evaluation pro-

cess involved only one blind student, and while their

CSEDU 2025 - 17th International Conference on Computer Supported Education

insights have been instrumental, they may not fully

capture the diverse needs and perspectives of the

broader blind community. We acknowledge the ne-

cessity for further testing and validation across vari-

ous contexts and with a more representative sample

of blind users.

6.3 Future Work

Moving forward, further development should address

the previously mentioned limitations by expanding

the tool’s applicability to diverse academic contexts

and a broader sample group of visually impaired stu-

dents. Engaging with a larger blind community from

different academic disciplines will ensure the tool’s

functionalities align with varied preferences and re-

quirements.

Furthermore, integrating options for direct user

feedback on the generated slides as part of the

tool’s functionality would help iterative improve-

ments based on user interactions and experiences.

By addressing these areas, researchers can con-

tribute to advancing accessible technologies for visu-

ally impaired individuals in educational settings, en-

hancing effectiveness, usability, and inclusivity for a

broader audience.

7 CONCLUSION

This paper discussed the possibilities of AI for gen-

erating accessible learning material. Despite le-

gal requirements for accessibility, visually impaired

students still encounter barriers in many aspects of

higher education, particularly regarding access to

adapted course material. These deﬁcits were high-

lighted with the enrolment of a fully blind student in

one of our study programs. This formed the basis of a

comprehensive case study in which the complexities

of adapting existing course materials for blind stu-

dents were investigated. The exploration introduced

a promising solution: leveraging generative AI tech-

nology to streamline the creation of accessible edu-

cational materials. The introduction and evaluation of

the TRAIL tool exempliﬁed the potential of AI in gen-

erating textual descriptions of lecture slides, thereby

enhancing accessibility and creating a more inclusive

learning environment. Moving forward, continued re-

search and innovation in AI-driven accessibility so-

lutions hold the key to overcoming barriers and en-

suring equitable educational opportunities for all stu-

dents, regardless of disabilities.

ACKNOWLEDGEMENTS

This paper portrays the work carried out in the con-

text of the KIWI project (16DHBKI071) that is gen-

erously funded by the Federal Ministry of Education

and Research (BMBF).

REFERENCES

Armstrong, H. and Murray, I. (2007). Remote and local de-

livery of cisco education for the vision-impaired. In

Proceedings of the 12th annual SIGCSE conference

on Innovation and technology in computer science ed-

ucation, pages 78–81.

Athanasios, K., Konstantinos, P., Doxa, P., and Eleni, K.

(2009). Students with visual impairments in higher

education institutes. In 7th European Conference of

ICEVI.

Bernardi, R., Cakici, R., Elliott, D., Erdem, A., Erdem, E.,

Ikizler-Cinbis, N., Keller, F., Muscat, A., and Plank,

B. (2016). Automatic description generation from im-

ages: A survey of models, datasets, and evaluation

measures. Journal of Artiﬁcial Intelligence Research,

55:409–442.

Billah, S. M., Ashok, V., Porter, D. E., and Ramakrishnan, I.

(2017). Ubiquitous accessibility for people with visual

impairments: Are we there yet? In Proceedings of the

2017 chi conference on human factors in computing

systems, pages 5862–5868.

Bishop, D. and Rhind, D. J. (2011). Barriers and enablers

for visually impaired students at a UK Higher Educa-

tion Institution. British Journal of Visual Impairment,

29(3):177–195.

Bradea, A. and Bl

andul, V. C. (2017). New Modalities to

Increase the Accessibility of Students with Visually

Impaired to Education Using ICT. New Trends and Is-

sues Proceedings on Humanities and Social Sciences.

Burewicz, A. and Miranowicz, N. (2002). Categorization of

visualization tools in aspects of chemical research and

education. International Journal of Quantum Chem-

istry, 88(5):549–563.

Butler, M., Holloway, L., Marriott, K., and Goncu, C.

(2017). Understanding the graphical challenges faced

by vision-impaired students in Australian universi-

ties. Higher Education Research & Development,

36(1):59–72.

Choi, J., Jung, S., Park, D. G., Choo, J., and Elmqvist,

N. (2019). Visualizing for the non-visual: Enabling

the visually impaired to use visualization. Computer

Graphics Forum, 38(3):249–260.

Croft, E. (2020). Experiences of visually impaired and blind

students in UK higher education: an exploration of ac-

cess and participation. Scandinavian Journal of Dis-

ability Research, 22(1):382–392.

DBSV e. V. (n. d.). Quick Guide Barrierefreie PowerPoint-

Folien. Online; accessed 08.01.2025.

De Beco, G. (2014). The right to inclusive education ac-

cording to Article 24 of the UN Convention on the

Generating Accessibility: Using AI to Improve Higher Education for the Visually Impaired

rights of persons with disabilities: background, re-

quirements and (remaining) questions. Netherlands

Quarterly of Human Rights, 32(3):263–287.

Fichten, C. S., Ferraro, V., Asuncion, J. V., Chwojka, C.,

Barile, M., Nguyen, M. N., Klomp, R., and Wolforth,

J. (2009). Disabilities and e-learning problems and so-

lutions: An exploratory study. Journal of Educational

Technology & Society, 12(4):241–256.

Firat, T. (2021). Experiences of students with visual impair-

ments in higher education: barriers and facilitators.

British Journal of Special Education, 48(3):301–322.

Fouh, E., Akbar, M., and Shaffer, C. A. (2012). The role of

visualization in computer science education. Comput-

ers in the Schools, 29(1-2):95–117.

Giese, M., Greisbach, M., Meier, M., Neusser, T., and

Wetekam, N. (2022). ‘I usually never got involved’:

understanding reasons for secondary students with

visual impairments leaving mainstream schooling in

Germany. European Journal of Special Needs Educa-

tion, 37(2):264–277.

Glazko, K. S., Yamagami, M., Desai, A., Mack, K. A.,

Potluri, V., Xu, X., and Mankoff, J. (2023). An Au-

toethnographic Case Study of Generative Artiﬁcial In-

telligence’s Utility for Accessibility. In Proceedings of

the 25th International ACM SIGACCESS Conference

on Computers and Accessibility, pages 1–8.

He, X. and Deng, L. (2017). Deep learning for image-to-

text generation: A technical overview. IEEE Signal

Processing Magazine, 34(6):109–116.

Heyer, K. (2021). What is a human right to inclu-

sive education? The promises and limitations of the

crpd’s inclusion mandate. Handbuch Inklusion inter-

national/International Handbook of Inclusive Educa-

tion, pages 45–58.

Hutchinson, J. S. O., Atkinson, K., and Orpwood, J. (1998).

Breaking down barriers: Access to further and higher

education for visually impaired students. Nelson

Thornes.

Ishihara, T., Takagi, H., Itoh, T., and Asakawa, C. (2006).

Analyzing visual layout for a non-visual presentation-

document interface. In Proceedings of the 8th inter-

national ACM SIGACCESS conference on Computers

and accessibility, pages 165–172.

Jenkinson, J. (2018). Molecular biology meets the learning

sciences: Visualizations in education and outreach.

Journal of Molecular Biology, 430(21):4013–4027.

Johnson, O., Alyasiri, O., Akhtom, D., and Johnson, O.

(2023). Image analysis through the lens of chatgpt-

4. Journal of Applied Artiﬁcial Intelligence, 4(2).

Jones, L. L., Jordan, K. D., and Stillings, N. A. (2005).

Molecular visualization in chemistry education: the

role of multidisciplinary collaboration. Chemistry Ed-

ucation Research and Practice, 6(3):136–149.

Jones, M. G., Minogue, J., Oppewal, T., Cook, M. P., and

Broadwell, B. (2006). Visualizing without vision at

the microscale: Students with visual impairments ex-

plore cells with touch. Journal of science education

and technology, 15(5):345–351.

Jung, C., Mehta, S., Kulkarni, A., Zhao, Y., and Kim, Y.-S.

(2021). Communicating visualizations without visu-

als: Investigation of visualization alternative text for

people with visual impairments. IEEE transactions

on visualization and computer graphics, 28(1):1095–

1105.

Kim, N. W., Joyner, S. C., Riegelhuth, A., and Kim, Y.

(2021). Accessible Visualization: Design Space, Op-

portunities, and Challenges. Computer Graphics Fo-

rum, 40(3):173–188.

Leiker, D., Gyllen, A. R., Eldesouky, I., and Cukurova, M.

(2023). Generative AI for Learning: Investigating the

Potential of Learning Videos with Synthetic Virtual

Instructors. In International conference on artiﬁcial

intelligence in education, pages 523–529. Springer.

Lundgard, A. and Satyanarayan, A. (2021). Accessible vi-

sualization via natural language descriptions: A four-

level model of semantic content. IEEE transactions

on visualization and computer graphics, 28(1):1073–

1083.

MacKay, D. (2006). The United Nations Convention on the

rights of persons with disabilities. Syracuse J. Int’l L.

& Com., 34:323.

Mohamady, A. A., Fathy, A. G., Ibrahim, A. E., Darwish,

M. S., and Salama, M. K. (2023). Enhancing Acces-

sibility and Independence of Visually Impaired Indi-

viduals through AI, ML and IoT: The Development

of a Smart Robot Assistant. Enhancing Accessibil-

ity and Independence of Visually Impaired Individuals

through AI, ML and IoT: The Development of a Smart

Robot Assistant.— IUSRJ, 4.

Moured, O., Alzalabny, S., Schwarz, T., Rapp, B., and

Stiefelhagen, R. (2023). Accessible Document Lay-

out: An Interface for 2D Tactile Displays. In Proceed-

ings of the 16th International Conference on PErva-

sive Technologies Related to Assistive Environments,

pages 265–271.

Naps, T. L., R

oßling, G., Almstrum, V., Dann, W., Fleis-

cher, R., Hundhausen, C., Korhonen, A., Malmi, L.,

McNally, M., Rodger, S., et al. (2002). Exploring the

role of visualization and engagement in computer sci-

ence education. In Working group reports from ITiCSE

on Innovation and technology in computer science ed-

ucation, pages 131–152. Association for Computing

Machinery.

Open AI (n. d.). Text generation models. Online; accessed

08.01.2025.

Permvattana, R., Armstrong, H., and Murray, I. (2013). E-

learning for the vision impaired: A holistic perspec-

tive. International Journal of Cyber Society and Edu-

cation, 6(1):15–30.

Rapp, D. N. (2005). Mental models: Theoretical issues for

visualizations in science education. Visualization in

science education, pages 43–60.

Sharif, A., Chintalapati, S. S., Wobbrock, J. O., and Rei-

necke, K. (2021). Understanding screen-reader users’

experiences with online data visualizations. In Pro-

ceedings of the 23rd International ACM SIGACCESS

Conference on Computers and Accessibility, pages 1–

16.

Sheela, S., Krishnan, R., Srivathsan, R., Sudharsan, M., and

Sriram, S. (2023). Enhancing accessibility: object

CSEDU 2025 - 17th International Conference on Computer Supported Education

detection and optical character recognition for em-

powering visually impaired individuals. In Interna-

tional Conference on Computer Vision and Internet of

Things 2023 (ICCVIoT’23), volume 2023, pages 305–

309. IET.

Simui, F., Kasonde-Ngandu, S., Cheyeka, A. M.,

Simwinga, J., and Ndhlovu, D. (2018). Enablers

and disablers to academic success of students with

visual impairment: A 10-year literature disclosure,

2007–2017. British Journal of Visual Impairment,

36(2):163–174.

Tervakari, A.-M., Silius, K., Koro, J., Paukkeri, J., and Pirt-

til

a, O. (2014). Usefulness of information visualiza-

tions based on educational data. In 2014 IEEE global

engineering education conference (EDUCON), pages

142–151. IEEE.

Thurston, A. (2014). Disability, power and equality in the

school. International Journal of Disability, Develop-

ment and Education, 61(2):105–107.

United Nations (n. d.). Convention on the rights of persons

with disabilities (crpd). Online; accessed 08.01.2025.

Vanderheiden, G. C. (2008). Ubiquitous accessibility, com-

mon technology core, and micro assistive technology:

Commentary on “computers and people with disabil-

ities”. ACM Transactions on Accessible Computing

(TACCESS), 1(2):1–7.

Vavra, K. L., Janjic-Watrich, V., Loerke, K., Phillips, L. M.,

Norris, S. P., and Macnab, J. (2011). Visualization in

science education. Alberta Science Education Jour-

nal, 41(1):22–30.

Vieira, C., Parsons, P., and Byrd, V. (2018). Visual learning

analytics of educational data: A systematic literature

review and research agenda. Computers & Education,

122:119–135.

Wang, L. L., Cachola, I., Bragg, J., Cheng, E. Y.-Y., Haupt,

C., Latzke, M., Kuehl, B., van Zuylen, M. N., Wagner,

L., and Weld, D. (2021). Scia11y: Converting scien-

tiﬁc papers to accessible html. In Proceedings of the

23rd International ACM SIGACCESS Conference on

Computers and Accessibility, pages 1–4.

Zhang, Z., Kim, G. S., and Wobbrock, J. O. (2023). De-

veloping and Deploying a Real-World Solution for

Accessible Slide Reading and Authoring for Blind

Users. In Proceedings of the 25th International ACM

SIGACCESS Conference on Computers and Accessi-

bility, pages 1–15.

Generating Accessibility: Using AI to Improve Higher Education for the Visually Impaired