Generative Artificial Intelligences (GAIs) have great potential for people with cognitive disabilities, since they can increase their autonomy, ofer them personalized learning and break down the barriers that this group traditionally has to access information. To address this gap, this paper investigates the specific dificulties experienced by users with cognitive disabilities when interacting with GAIs. We present an exploratory case study involving 16 students with cognitive disabilities enrolled in a universitybased educational program. ChatGPT was selected as the representative GAI tool due to its popularity and the similarity of its interface and interaction paradigms to those of other widely used Generative AI systems. The intervention was integrated into the students' academic curriculum and involved completing practical academic tasks using ChatGPT. A mixed methodological approach was employed, combining pre- and post-intervention questionnaires completed by students and tutors/experts, along with on-site observations during the sessions.
The advancement of Large Language Models (LLMs), powered by deep neural networks, has resulted in human-like conversational assistants, including chatbots and voice bots. Leading technology companies, such as OpenAI, Google and Meta, among others, are developing chatbots that can generate text and images in response to user-defined prompts. These are commonly referred to as Generative Artificial Intelligences (GAIs). Their outputs often appear remarkably human, and the use of chatbots has become widespread, with both the general public and professionals utilizing this technology to tackle a wide range of tasks.
However, the benefits of this technological progress may not be reaching all segments of the population equally. For individuals with cognitive impairments, accessing and efectively using GAIs might continue being a significant challenge. Despite their broad capabilities, current GAI systems are often not designed for cognitive accessibility, creating barriers that limit this user group’s ability to interact with these tools on equal terms. Yet, GAIs hold enormous potential for people with cognitive disabilities: they can promote autonomy, ofer personalized learning, and help break down traditional barriers to Interacción’25: XXV CONGRESO INTERNACIONAL DE INTERACCIÓN PERSONA-ORDENADOR, September 03–05, 2025, Val
CEUR
ceur-ws.org accessing information, communication, and knowledge. To ensure that this group can take advantage of this potential, it is necessary to research and identify the specific dificulties these users face in accessing GAI. In this context, equitable access to emerging technologies must explicitly consider and address the needs of individuals with cognitive impairments.
Generative AI interfaces may often appear as counterintuitive and potentially lack necessary adaptations for cognitive accessibility, usually featuring complex menus, confusing icons, and limited customization options. Additionally, these tools may employ complex and abstract language, potentially making it dificult for people with cognitive impairments to understand and interact efectively with them. These challenges might include dificulties in using interfaces, creating clear prompts, interpreting results, and customizing options. All of this could potentially have a negative impact on the user experience and the quality of the responses generated for this specific group. To date, there is a significant gap in research on the adoption, impact, and cognitive accessibility needs associated with the use of GAI tools by individuals with cognitive disabilities. It is essential to identify functional, linguistic, and design adaptations that enable efective, understandable, and safe use of these tools.
Addressing this gap, this paper poses the following research question: What are the challenges faced by people with cognitive disabilities in accessing and using generative AI? The exploratory case study presented in this paper aims to identify the specific dificulties and barriers that users with cognitive disabilities face when interacting with a representative Generative AI chatbot, seeking to corroborate our initial assumption of unequal access.
It was essential for the purpose of testing our hypothesis and achieving our research goals to have the
individuals with cognitive disabilites to interact with a GAI. ChatGPT was selected as the representative
tool for this study due to its status as the most widely known text-based generative AI system. Developed
by OpenAI and released for public use in 2022 with the launch of version 3.5, it quickly became the
fastest-growing application, reaching 100 million users [
The study presented in this work was conducted with students from the ACCEDE program, a university-specific certificate ofered by Universidad Complutense de Madrid (UCM) for individuals with intellectual and developmental disabilities. The activity was intentionally integrated into the students’ academic curriculum to ofer a tangible benefit to the participants, ensuring that the research provided a meaningful and relevant learning experience for them.
This study pursues the following specific objectives: • To evaluate the satisfaction and perceived usefulness of ChatGPT by students with cognitive disabilities. • To identify the main dificulties encountered during the interaction, including asking questions, understanding answers, and verifying information. • To explore improvements suggested by participants to make ChatGPT, in particular, and all GAIs in general, more accessible.
• To analyze students’ perceptions of their future ability to use this tool independently.
The expected outcomes of the study include the collection of relevant information to facilitate more equitable access and more efective use of GAIs, the identification of specific areas for improvement in cognitive accessibility in tools such as ChatGPT, and the generation of useful knowledge to guide future research on the inclusive design of GAIs for people with cognitive disabilities.
The rest of the paper is organized as follows. Section 2 presents the state of the art on generative AI technologies and their use and utility for users with cognitive disabilities. The methodology of this research is described in Section 3. Then, the obtained results are presented in Section 4, and discussed in Section 5. Finally, we acknowledge the limitations of the study in Section 6 and the conclusions and future work are addressed in Section 7.
The rapid growth of Generative AI (GAI) technologies has led to increased research on their use and
utility for end users [
Even though the usability of GAI tools is currently understudied, some research is emerging on the
design principles to consider for Generative AI User Experience and their implications in the field of
Human-Computer Interaction (HCI) [
One of these overlooked groups is people with cognitive disabilities. These individuals face significant
challenges when accessing GAIs, such as cognitively inaccessible interfaces and language not adapted
to their needs. Glazko et al. [
Recent evaluations using established accessibility standards reveal systematic barriers in current
GAI tools. Acosta-Vargas et al. [
These findings are consistent with concerns raised by Alshaigy and Grande [
Despite these challenges, there is growing research exploring the potential benefits of GAI for people
with cognitive disabilities, particularly in educational environments. Liu et al. [
Several studies have examined the real-world use of generative tools by students with disabilities.
Pierrés et al. [
Other recent eforts have explored GAI tools specifically designed for or tested with neurodivergent
users. TwIPS [
In summary, while GAI technologies ofer great potential, especially for people with cognitive disabilities, current design practices often fail to meet their needs as they do not take them into account. The literature highlights both significant opportunities, and significant accessibility gaps resulting from non-inclusive design. Most existing studies focus on specific use cases or individual tools, often without exploring how these tools perform in real contexts involving people with cognitive disabilities. Our study addresses this gap by evaluating its accessibility in authentic tasks, with the direct participation of users with cognitive disabilities. With this, we seek to provide practical design ideas that go beyond compliance with accessibility standards and move toward genuinely inclusive interaction paradigms.
This research takes the form of an exploratory case study that investigates the access, perceived usefulness, dificulties, and future potential of using ChatGPT by students with cognitive disabilities within the context of a specific educational activity at university. Our main goal is not to assess the efectiveness of ChatGPT as a learning tool, but shed light on the underlying challenges related to cognitive accessibility, HCI, and the comprehension of information generated by GAIs, from the perspective of users with cognitive disabilities. Identifying these dificulties is essential for designing more inclusive GAIs and for developing efective support strategies.
Students of the ACCEDE program were selected for this study. The ACCEDE program1 is a universityspecific certificate ofered by the Complutense University of Madrid (Universidad Complutense de Madrid, UCM) for individuals with intellectual and developmental disabilities. Its goal is to train participants to become assistant technicians in the evaluation of inclusive environments, promoting the social and labor inclusion of young people with cognitive disabilities through a training program for employment and university inclusion within the environment of the UCM.
Next sections present a thorough description of the research design, the participants involved in the study, the used data collection instruments and the tools employed for data analysis.
To explore the interaction of students with cognitive disabilities with ChatGPT, a mixed methodological approach was adopted, combining quantitative and qualitative data collection and analysis. Data collection was conducted through questionnaires given to students before and after the intervention, as well as parallel questionnaires completed by their tutor at the ACCEDE program and the support experts that participated in the intervention. Figure 1 presents the structure and stages of the study.
Stage 1 comprised the completion of pre-test questionnaires by the participants (students with cognitive disability) and their tutor in the ACCEDE program. Each student individually completed their own questionnaire, while the tutor filled out a separate questionnaire for each participant in the study. More details about the questionnaires can be found later in Section 3.3.
Stage 2, which involved the main intervention, was structured into two main phases: • Formative Phase: The first phase consisted of a training session designed to introduce the fundamental concepts of Artificial Intelligence and Generative Artificial Intelligence (GAI), with a specific focus on ChatGPT. The contents addressed included: – Conceptual introduction to AI and its various applications. – Explanation of the distinctive characteristics of AIs and their ability to generate content. – Presentation of ChatGPT as an example of conversational GAI, explaining its basic functions. – Exploration of practical examples of how ChatGPT can be used as a support tool. – Emphasis on the importance of being aware of the possible limitations and errors of ChatGPT, promoting a critical attitude towards the information generated. – Recommendations for a responsible and safe use of the tool. – Basic strategies for the formulation of efective questions and for the verification of the information provided by ChatGPT, emphasizing the importance of being specific in the input questions, asking for clarifications and contrasting the information with other sources. • Practical Phase: In the second phase, students participated in practical activities in a computer laboratory where they used ChatGPT to perform specific academic tasks. To ensure the tasks’ practical relevance and the pedagogical value of the activities, they were selected in collaboration with the students’ tutors, prioritizing those they had already worked in class. From various tutor-suggested ideas (e.g., preparing interview questions, creating a university leaflet on cultural activities, researching the university’s coat of arms, evaluating building/poster accessibility, classifying expenses,...), three were chosen. Our selection criteria focused on tasks representing distinct AI interaction purposes and varied forms of engagement to ensure a broad exploration of challenges. Feasibility, quick completion, minimal student input, and clear ChatGPT applicability were also key factors (e.g., complex accessibility evaluations were excluded). Each selected task was designed to elicit diferent cognitive demands and interaction patterns, enabling a comprehensive analysis of potential barriers and facilitators relevant to individuals with cognitive disabilities. The following three academic tasks were finally selected to investigate interaction challenges with ChatGPT: – Research about the origin of some elements on the university coat of arms. This activity focused on ChatGPT’s ability to provide factual information and required the formulation of specific search questions. Students formulated direct questions (e.g., “Why is the swan the symbol of the Complutense University of Madrid?”), read and attempted to comprehend the information provided, identifying key facts, and finally, put their findings in common in a group discussion. – Classifying diferent expenses into needs and wants. This task aimed to assess how students could use ChatGPT for conceptual reasoning and classification tasks, requiring understanding definitions and applying diverse criteria. Students first asked ChatGPT to define “need-based expense” vs. “want-based expense” (e.g., “What is the diference between a need and a want regarding expenses?”). Subsequently, they classified a provided list of expenses (e.g., Trousers, Education, Soft drinks, Holidays, Housing, etc.) individually without ChatGPT, then prompted ChatGPT to classify the same list, and finally compared their personal classifications with ChatGPT’s, reflecting on any discrepancies. – Preparation of interview questions for the Vice-Dean of Quality of their school. This task focused on the use of ChatGPT as a tool for planning and generating content for a formal academic interaction, involving the identification of relevant topics and the formulation of relevant and well-structured questions. Students informed ChatGPT of the interviewee’s role and interview purpose, asked it to generate relevant questions, and then reviewed and selected the most appropriate and professional ones. Finally, students put their questions in common in a group discussion.
During the practical phase, the students were assisted by their ACCEDE tutor and four intervention support experts. In order to focus their attention on specific participants, each of the four experts attended three specific students, and the ACCEDE tutor took care of the remaining four.
After the intervention, Stage 3 consisted on the completion of post-test questionnaires by all the actors involved. Each participant filled out a personal questionnaire, while each support expert and the tutor completed one questionnaire for each of the 3-4 students they had assisted during the intervention. More details about these materials are included in Section 3.3.
Sixteen students (n=16) with cognitive disabilities from the ACCEDE program participated in the study. Participants were between 18 and 27 years old (M = 21, SD = 2.50), with an equal gender distribution (50% women, 50% men). Only one student reported having an additional motor impairment (6.25%), and another student reported a visual impairment (6.25%).
Regarding the level of cognitive disability, 75% of the students reported having a mild level, while 18.75% had a borderline level, and 6.25% had a moderate level. All participants have dificulties with logical reasoning, 56.25% have alternating attention dificulties and 25% have cognitive flexibility dificulties. Less frequent dificulties included selective attention (18.75%), inhibition (18.75%), and problems with working memory (18.75%), short-term memory (12.5%), written expression (12.5%), sustained attention (6.25%), and long-term memory (6.25%).
In terms of technological autonomy, the majority of participants (81.25%) reported being very autonomous or needing little help when using technology. All participants reported using technology on a regular basis, with 87.5% using it daily and only 12.50% using it occasionally or rarely. Half of the students (50%) felt very comfortable using new applications, 43.75% felt somewhat comfortable, and only one participant (6.25%) reported low comfort levels.
The review of relevant studies shown in Section 2 provided an overview of various data collection instruments such as semi-structured interviews, Likert-scale surveys, qualitative data collection, and observational methods. However, there remains limited availability of standardized or directly applicable instruments specifically designed to evaluate the nuanced interaction challenges posed by Generative AI technologies like ChatGPT, particularly within our academic context and for this specific population. Existing measures often focus on general usability, diferent types of AI, or populations without cognitive disabilities, and many are designed for a broader scope than our focused inquiry. Therefore, to align with the exploratory nature and specific objectives of this case study, new questionnaires were specifically developed. These custom-designed instruments allowed us to directly assess the specific interaction aspects most relevant to our research questions and target population in a contextually appropriate manner.
To achieve this comprehensive assessment, data was collected primarily through these newly developed questionnaires, administered to students and their ACCEDE tutor before and after the intervention, and to support experts after the intervention. The questionnaires had closed-ended questions (with dichotomous categorical answers -Yes/No- and 3- and 4-category ordinal answers) as well as open-ended questions. All questionnaires were administered in Spanish, the native language of both the students with cognitive disabilities and their ACCEDE tutors/support experts.
The inclusion of the ACCEDE tutor as informant was due to her in-depth knowledge of the students’ individual abilities, dificulties and learning styles. In addition, the responses of the tutor and support experts to the post-test questionnaires were based not only on their prior knowledge of the students, but also on direct observations made during the intervention. This on-site observation of students’ interaction with ChatGPT ofered a privileged perspective and allowed cross-validation of students’ self-reported data, helping to identify possible discrepancies between self-perception and behavioral manifestation during GAI use, especially in areas where students might have dificulties to self-assess accurately their skills or understanding (e.g., dificulty admitting to not understanding something or specifying the cognitive areas they have afected).
The researchers and the ACCEDE tutor collaborated actively in the development of the data collection instruments. Based on her direct experience with students in the ACCEDE program, the tutor provided valuable input regarding the language and forms of expression most accessible to the participants. The joint review made it possible to identify and modify possible ambiguities or terms that could be confusing for the students, ensuring that the questions were interpreted in the manner intended by the researchers. The collaboration also ensured that the Likert range used for each question was appropriate for the participants’ ability to discriminate and that the response options were mutually exclusive and exhaustive.
Prior to the intervention, pre-test questionnaires, which can be consulted in Table 12, were administered to both students and their tutor: • The student questionnaire collected information about their general familiarity with technology, comfort using new applications, confidence in searching for information online, prior knowledge on Artificial Intelligence and specifically about ChatGPT, etc. • The tutor questionnaire collected student demographics (age, gender, type and level of disability, main areas of cognitive dificulty), their level of autonomy with technology, and any other considerations relevant to the student’s participation in the study. The data of participants described in Section 3.2 were obtained from the pre-test questionnaire filled out by her tutor.
PRE-TEST QUESTIONNAIRE FOR STUDENTS Do you like to ask for help or do you prefer to do things on your own? Do you usually ask questions if something isn’t clear? What do you do when you don’t understand the answer to your question? How often do you use technology (computer, phone, tablet)? Are you comfortable using new apps or programs? What technology do you use the most? Is it easy for you to search for information on the internet? Do you always trust the information you find on the internet? Who would you ask for help if you need to know if something on the internet is true? Have you heard of Artificial Intelligence? Could you describe it in your own words? Do you know any AI tools? Which ones? Have you ever used them? What have you used them for? Do you know ChatGPT? What is ChatGPT and what is it used for? What do you hope to learn from this activity? Is there anything specific that worries you or seems dificult about this activity? Do you think technology can help you complete college assignments?
PRE-TEST QUESTIONNAIRE FOR TUTOR Age Sex Type of disability Level of disability Main areas of cognitive dificulty Level of autonomy with technology
Other issues that should be taken into account 2It is important to note that the questionnaires presented are English translations of the original Spanish instruments used during the study.
During the practical phase of the intervention, the tutor and four support experts observed the students’ interaction with ChatGPT and took note of everything that happened. In order to focus their attention on specific participants, each of the four experts attended three specific students, and the ACCEDE tutor took care of the remaining four. Once the intervention was completed, post-test questionnaires were administered to both the students and the tutor and support experts. These questionnaires explored participants’ experiences with ChatGPT, including dificulties encountered in asking questions and understanding answers, their ability to verify information, their perception of the usefulness of the tool, their willingness to use it in the future, and improvements they would suggest to make it more accessible. The questionnaire for the tutor/experts also included questions about their perception of the students’ learning and autonomy during the activity, as well as general recommendations for improving the accessibility of ChatGPT for students with cognitive disabilities. Post-test questionnaires can be consulted in Table 2.
POST-TEST QUESTIONNAIRE FOR STUDENTS Did you enjoy using ChatGPT? Did you find ChatGPT useful for the activities? What did you like the most about ChatGPT? What did you like the least about ChatGPT? Did you have any dificulties when asking ChatGPT questions? What kind of dificulties? Were you able to check if the information was correct? How did you try to verify it? Did you understand the answers ChatGPT gave you? Were you able to use the information ChatGPT gave you to complete the tasks? Was there anything in the answers that you didn’t understand? How did you solve it? Do you feel you could use ChatGPT for schoolwork on your own in the future? Do you think ChatGPT can give you useful answers? Would you like to keep using ChatGPT to learn? Do you feel more confident using technology after this activity? Do you think ChatGPT needs improvements to be easier to understand? If yes, what would you suggest?
POST-TEST QUESTIONNAIRE FOR TUTOR/EXPERTS Did the student need a lot of help to formulate appropriate questions? Did the student understand ChatGPT’s answers? Was the student able to follow the steps to verify the information? Did the student show frustration or lack of motivation at any point? Did the student seem to enjoy the activity? Did the student seem to be learning during the activity? Did you observe any improvements in the student’s autonomy when using ChatGPT during the activity? What kind of dificulties did the student have? Do you think ChatGPT would be useful in future activities for the student?
What changes would you recommend to make ChatGPT more accessible?
All questionnaires were completed in physical format and later digitized and anonymized for subsequent analysis.
The quantitative data were analyzed using descriptive and inferential statistics with SAS 9.4 statistical software. Variables were summarized using absolute frequencies and percentages. To examine associations between variables, Pearson’s Chi-square test was used when the expected cell counts were suficient, and Fisher’s Exact Test was applied when Chi-square assumptions were not met. For comparisons involving ordinal variables, the Wilcoxon rank-sum test was employed. In addition, Spearman’s rank correlation coeficient was used to assess the strength and direction of monotonic relationships between ordinal variables that did not meet normality assumptions.
The qualitative data, derived from open-ended responses, were analyzed using thematic analysis, following the procedures outlined by Green and Thorogood [23] and Braun and Clarke [24]. A classical content analysis approach was used to identify, analyze, and report patterns within the data [25]. First, each sentence or meaningful segment was assigned one or more codes that summarized its core idea. These codes were then grouped into categories based on semantic similarity. From these categories, key descriptive themes emerged. Representative textual quotations were selected to illustrate each theme, and all quotes were translated into English for reporting.
This section presents the results of the study, structured around four main themes that emerged from the data analysis. These themes reflect main aspects of the participants’ experience before, during, and after their interaction with ChatGPT. First, we describe the technological profile of participants and their prior expectations and beliefs about Artificial Intelligence (Section 4.1). We then analyze the dificulties encountered during the interaction with ChatGPT, focusing on question formulation, comprehension of responses, and information verification (Section 4.2). Next, we explore students’ attitudes, emotional experiences, and perceptions of usefulness and learning after using the tool (Section 4.3). Subsequently, we summarize improvement suggestions provided by both students and tutor/experts to enhance the accessibility and efectiveness of ChatGPT (Section 4.4).
The results presented in this subsection are based on the pre-test questionnaires completed by the students. Regarding their help-seeking behaviors, a large majority of participants (87.5%) stated that they ask for help (to other people or computer tools) when something is unclear at least occasionally.
Concerning information literacy, half of the participants reported that they always find it easy to search for information on the internet, while 43.75% indicated that they sometimes do. Only a small percentage (6.25%) reported dificulty with this task. However, a considerable percentage (43.75%) admitted to trusting the information they find online without verifying it.
Regarding their familiarity with Artificial Intelligence (AI), 62.5% of the students had heard of AI, although only 37.5% were able to name a specific AI tool. Among this latter group, only three students (23.1% of those responding afirmatively) had actually used an AI tool before. Specific knowledge of ChatGPT was even more limited: 37.5% of the group knew about it, and another 18.75% had heard of it but did not know exactly what it was. Notably, no students with moderate or borderline intellectual disabilities were familiar with ChatGPT, suggesting a potential relationship between the level of disability and the degree of familiarity with this type of AI tool (Fisher’s exact test, p = 0.0346).
Finally, the majority of participants expressed a positive belief in the potential of technology to support their academic tasks (75%), with 18.75% expressing uncertainty and only one student reporting a negative view.
The analysis of students’ interaction with ChatGPT is structured around three dimensions that reflect the main stages of a typical user experience with a generative AI tool: asking questions, understanding the answers, and verifying the accuracy of the information provided. These findings are based on the analysis of the post-test questionnaires filled out by the students, as well as the tutor and support experts involved in the intervention.
Regarding the students’ experience in asking questions to ChatGPT, the results indicate a low perceived level of dificulty, with only 19% of participants reporting having encountered challenges when formulating questions. The tutor/experts’ perspective ofered a more nuanced view regarding the need for help (students could only answer Yes or No, while the tutor/experts had four response categories): 19% required a lot of help, 25% moderate help, another 25% little help, and 31.25% did not need any help.
A significant relationship was found between dificulties in writing and the need for help formulating questions (Fisher’s exact test, p = 0.0250). Specifically, 100% of the students with writing dificulties were rated as needing a lot of help, while only 7.1% of students without such dificulties received that same rating. Among students without writing dificulties, the remaining distribution was: 28.6% moderate help, 28.6% little help, and 35.7% no help. These diferences were also confirmed by a Wilcoxon rank-sum test (W = 4.00, p = 0.0401).
An analysis of the open-ended responses provided by students and the tutor/experts led to the identification of several main categories of dificulties that students experienced when asking questions to ChatGPT: • Dificulties related to the question’s conceptualization and clarity. It was observed that some participants struggled to formulate questions that were relevant to the context of the task or required a specific focus. One student exemplified this by commenting: “I phrased the expense question poorly because I asked about unnecessary expenses instead of necessary ones”. Similarly, the lack of clarity manifested itself in the formulation of broad and poorly defined questions, as evidenced by the following tutor’s observation: “She asks about the vice dean in general and not which specific vice dean she wants to interview” . • Dificulties in question generation and articulation: In addition to conceptualization, some participants faced challenges when it came to actually generating the questions needed for the task. Some participants expressed uncertainty about how to start asking the questions they needed to complete the activity, as one student noted: “I didn’t know how to start asking the questions I needed for the activity”. Furthermore, dificulty identifying necessary information and how to request it to ChatGPT was a common obstacle, as one expert noted: “She struggled with formulating questions (she didn’t know how to ask the chat what she needed, and we had to help her)”. • Linguistic formulation dificulties: The presence of spelling errors, as noted in the observation “I made many spelling mistakes when asking questions” could have afected the GAI’s ability to understand the user’s intent and provide relevant answers. • Dificulties when requesting refinement and clarification: Some dificulties related to formulating questions aimed at refining the information provided by ChatGPT or requesting clarification were identified. Comments such as “I wasn’t able to ask it to summarize or simplify” and “It’s dificult for them to ask for simplifications or to express that there are things they don’t understand” illustrate a lack of awareness or skill in using questions as a tool to deepen understanding or adapt information to their needs.
Regarding the understanding of the answers generated by ChatGPT, students’ perceptions varied: 43.75% stated they always understood the answers and 56.25% indicated that they understood them only sometimes. Going deeper into this issue, half of the students reported having understood all of the answers, while the other half reported finding some aspects incomprehensible. From the tutor/experts’ perspective, answer comprehension was evaluated more positively: 75% of the students understood the answers, 18.75% had dificulty understanding them, and only 6.25% did not understand them.
Exploring the potential influence of specific areas of cognitive dificulty on understanding ChatGPT answers, the results suggest a statistically significant relationship between cognitive flexibility dificulties and the tutor/experts’ assessment of answer comprehension (Fisher’s exact test, p = 0.0198). A Wilcoxon rank-sum test also showed that students with cognitive flexibility dificulties tended to experience more dificulties processing ChatGPT responses ( W = 50.50, p = 0.0104). Descriptively, 91.67% of students without dificulties in cognitive flexibility were rated as understanding the responses well. In contrast, among the students with such dificulties, only 25% was rated as understanding well, while 50% were rated as understanding “with dificulty”, and 25% as not understanding at all. These findings suggest that the ability to adapt to the information provided by ChatGPT, which often requires some lfexibility in processing, is compromised in students with this cognitive dificulty. Similarly, a significant relationship was identified between dificulties in written expression and tutor/experts’ assessment of answer comprehension (Fisher’s exact test, p = 0.0500). The Wilcoxon rank-sum test supported this association, indicating that students with written expression dificulties required more efort to understand the answers (W = 2.20, p = 0.0277). In fact, all students with dificulties in written expression were rated as understanding the answers with dificulty, whereas among the students without such dificulties, 85.7% were assessed as understanding well, 7.1% as understanding with dificulty, and 7.1% as not understanding.
Analyzing the responses to the open-ended questions of tutor/experts, several main categories of dificulties that students experienced when understanding answers were established to summarize the qualitative results: • Dificulties related to format and amount of information: Participants experienced dificulty when the response consisted of long, unstructured blocks of text. This was evident in comments such as: “He found it dificult to make progress when ChatGPT gave him all the text at once without bullet points or when it returned a lot of text”. • Dificulty requesting information adaptation: Another significant dificulty was students’ inability or unwillingness to ask ChatGPT to adapt the information to their needs. This was reflected in statements such as “He couldn’t ask it to summarize or simplify it”. • Dificulties understanding specific terms: Some participants encountered words or phrases within the responses that were unfamiliar to them or whose meaning they did not fully understand. This is evidenced by the following comment from this participant: “There were some words in ChatGPT’s responses that he did not understand”.
Analyzing the responses to the open-ended question about the strategies to try to solve the dificulties faced understanding the answers, several main categories of strategies were established: • Seeking external (human) help: A common strategy used by participants was to seek help from tutor/experts. This is evidenced by direct comments such as “Ask the teacher” and “To understand the answers, he needed a lot of help from me”. • Self-correction and question refinement strategies: In some cases, participants attempted to obtain more understandable answers by detailing or specifying their initial question, as described in “Further detailing and making the question more specific” . • Requesting adaptation and clarification from ChatGPT: A strategy directly related to interacting with the AI was asking it to adapt or clarify the information. Participants attempted to overcome their lack of understanding by asking for summaries, as described in “Asking for a summary”. Furthermore, when faced with unfamiliar terms, some chose to directly ask ChatGPT for their meaning or explanation, as illustrated by comments such as “Asking ChatGPT for words I didn’t understand in the answer (e.g., etymology)”.
A critical aspect of interacting with GAI tools is verifying the accuracy of the information provided. In this study, students’ perceptions of their ability to verify ChatGPT information was mostly negative, with 81.25% reporting that they were unable to perform this verification. From the tutor/experts’ perspective, a similar situation was observed, although with more variability: 53.33% of students were unable to verify the information, 26.67% did so with assistance, and only 20% were able to verify it independently. Regarding the methods used to attempt to verify information, a small percentage of students mentioned asking ChatGPT directly about the validity of their answer (6.25%), while another minority group appealed to internet searches (12.5%).
When analyzing the relationship between verification ability and specific areas of cognitive dificulty, several statistically significant relationships were identified. A significant relationship was found between the ability to follow steps to verify information and dificulties with alternating attention (Fisher’s exact test, p = 0.0174). A Wilcoxon rank-sum test further revealed that students with alternating attention dificulties exhibited a greater ability to verify information compared to those without these dificulties ( W = 69.00, p = 0.0078). Specifically, 33.33% students without alternating attention dificulties were able to verify the information on their own, 44.44% needed help, and 22.22% were unable to verify the information. In contrast, 100% students with alternating attention dificulties were able to verify the information on their own. Similarly, a significant relationship was found between verification ability and short-term memory dificulties ( Fisher’s exact test, p = 0.0286). The Wilcoxon rank-sum test revealed that none of the students with short-term memory dificulties had trouble verifying information, as all of them were able to verify it independently (W = 4.00, p = 0.0315). In contrast, the majority of students without these dificulties were unable to verify information (61.54%). Finally, a significant relationship was found between verification ability and working memory dificulties ( Fisher’s exact test, p = 0.0374): none of the students with working memory dificulties had dificulty verifying information, and they were able to do so alone or with help, while the majority of students without these dificulties (61.54%) were unable to do so.
As in the previous subsection, the findings to be shown in this subsection are based on the analysis of the post-test questionnaires completed by the students, as well as by the tutor and support experts involved in the intervention.
The majority of students expressed a positive attitude toward their future use of ChatGPT: 68.70% expressed a desire to continue using ChatGPT for learning, while 18.25% were unsure about it and 12.5% indicated they would not like to continue using it. Regarding perceived autonomy in future use, opinions were more divided: 37.5% were optimistic and believed they could use the tool without help, while a slightly larger proportion, 43.75%, anticipated the need for some assistance and 18.75% expressed a less autonomous view, believing they would require constant help to use ChatGPT. When analyzing the relationship of these results with specific areas of cognitive dificulty, a significant relationship was found between short-term memory dificulties and the tutor/experts’ assessment of whether the student had dificulties using ChatGPT ( Fisher’s exact test, p = 0.0500). Tutor/experts mostly rated students with short-term memory dificulties as not experiencing dificulties when interacting with ChatGPT, with 100% rated as not having dificulty. In contrast, 85.71% of students without short-term memory dificulties were rated by tutor/experts as having dificulties interacting with ChatGPT.
Overall, most students perceived ChatGPT as a useful tool for the activities they carried out: 56.25% considered it very useful and 43.75% considered it somewhat useful. Regarding the usefulness of the answers provided by ChatGPT for completing academic activities, the majority of students indicated that the experience was positive: 43.75% indicated that the answers were always useful, while 37.50% indicated that they were useful in almost all situations. Only 18.75% indicated that the answers were only useful in some situations.
According to the tutor/experts’ observations, most students seemed to experience learning during the activity: 66.67% of students learned during the session, 20% experienced moderate learning, and 13.33% did not seem to learn at all. It is important to note that a significant relationship was found between the students’ disability level and the tutor/experts’ perception of learning (Fisher’s exact test, p = 0.0220). Tutor/experts rated 83.33% with a mild disability level as having learned during the session, while none with moderate or borderline disabilities were perceived as having learned. In addition, a Wilcoxon rank-sum test revealed that students with mild disabilities were perceived as those who learned more during the activity compared to students with borderline or moderate disabilities (W = 38.50, p = 0.0155).
Regarding the emotional experience during the session, most of the students reported positive feelings: 75% showed no signs of frustration or demotivation at any point, while 12.5% experienced these feelings occasionally and another 12.5% consistently. Regarding enjoyment, 62.50% of students seemed to enjoy the session, 18.75% did so moderately, and another 18.75% did not seem to enjoy it at all. It is significant to note that a statistically significant association was found between enjoyment of the activity and the desire to continue using ChatGPT for learning (Fisher’s exact test, p = 0.0357; Spearman’s rank correlation, rₛ = 0.532, p = 0.0339), suggesting that a more pleasant experience with the tool is directly related to a greater willingness to use it in the future for learning purposes. 90% of the students who reported enjoying the activity expressed a desire to continue using ChatGPT for learning, while only 33.33% of the students who reported moderate enjoyment or not enjoyment shared this desire.
Regarding the question about the most appreciated aspects of ChatGPT posed by students, several main categories were established to summarize the qualitative results: • Perceived usefulness and benefits of the answers: Comments such as “The useful answers it gave me” underline the practical value they found in the AI’s responses. ChatGPT’s ability to ofer concrete examples, as mentioned in “How it gave me examples of expenses based on need and expenses based on desire” was also appreciated for facilitating understanding. Furthermore, the usefulness of the information for learning and completing academic tasks was highlighted in statements such as “Learning things with it”. The AI’s ability to provide information and help to ifnd answers was also a positive aspect, as indicated by “If you don’t know the answer to something, you can ask ChatGPT”. • Responsiveness and functionality: ChatGPT’s ability to answer a variety of questions was highly valued. This includes the ability to answer conceptual questions, as exemplified by “Asking what a need expense and a want expense mean” as well as the general function of answering questions, as indicated by “Being able to ask questions”. • Positive User Experience: The speed with which ChatGPT provided responses was a notable factor in the positive user experience, as mentioned in “How quickly it answers things”.
On the other hand, after analyzing the aspects that students liked least about ChatGPT, we obtain the following categories that summarize the qualitative responses given: • Problems with response quality and format: Several participants expressed frustration with the way the information was presented. Comments such as “the information was very long and I had to read a lot” reflected dificulty processing lengthy responses. The “lack of concise formatting” and the fact that the responses were not “outlined” were also noted as negative aspects, with a clear preference for brevity and outlines to facilitate understanding. • Questioning the reliability and accuracy of information: A significant concern among participants was the reliability of the information provided by ChatGPT. The “generation of incorrect information“ was mentioned multiple times, generating a “distrust of the information”. • Interaction and functionality issues: Some participants experienced dificulties interacting with the AI. It was mentioned that “sometimes it doesn’t understand the questions”.
Regarding the perceived impact on students’ confidence in using the technology, a vast majority of the participants (81.25%) reported feeling more confident using the technology after the session, 6.25% were undecided about the decision, and 12.5% indicated they did not feel more confident after the experience.
Participants also ofered valuable suggestions for improving ChatGPT’s functionality and user experience in post-test questionnaires. These suggestions were grouped into three main categories: • Improvements to the quality and timeliness of information: Participants expressed a desire for the tool to be “more up-to-date and to provide recent information correctly”. They also emphasized the need for “greater accuracy in responses” and to “avoid errors in information” even suggesting that incorrect responses should be eliminated altogether to increase reliability. • Improvements to the clarity and format of responses: Several participants indicated a preference for “shorter, less text-heavy responses” advocating for “more concise responses” and “less text-heavy responses”. The idea of “making responses simpler from the outset without having to be prompted” was also a recurring point to facilitate understanding. Additionally, it was suggested that “quick access to adaptation functions (simplify, summarize)” must be provided. • Improvements to user interaction and assistance: The ability to “correct questions if they are incorrect so they are answered correctly” was a key suggestion. Also mentioned was the need to “help me make questions more specific and detect general questions and ask for them to be more specific”. Furthermore, the implementation of a “guide to provide initial information” and “instructions to contextualize ChatGPT” from the outset was suggested, making it easier to provide user characteristics to obtain more personalized and relevant answers. Finally, the inclusion of a “list of predefined help commands” and “useful command suggestions” could empower users to more easily tailor answers to their needs.
Overall, the results highlight the central role that specific cognitive dificulties play in interacting with AI-based tools, afecting both question formulation and the understanding and verification of answers. These findings allow us to identify key areas for improving the accessibility of systems like ChatGPT.
The results of the presented exploratory study support our initial hypothesis that people with cognitive disabilities face significant dificulties when interacting with ChatGPT, a general-purpose GAI. While the tool demonstrated potential and generated positive feedback in terms of perceived usefulness and enjoyment (81% of participants liked using ChatGPT and 100% found it useful), the dificulties encountered in formulating questions, understanding answers, and verifying information highlights the need to redesign GAI interfaces for this group.
The dificulty in asking questions, as evidenced both by students’ self-perceptions (19% of students reported needing assistance formulating questions) and by tutor/experts’ observations (69% of students required some assistance), highlights the need for adapted interfaces. Incorporating predefined question templates for common tasks (e.g. defining a concept, explaining how to do something, etc.) or providing contextual suggestions could help users express their needs more efectively, especially for those with writing dificulties, who were the ones most frequently needing support formulating questions. Voice input, a feature suggested by students themselves, appears to be a valuable alternative for those with typing dificulties. Additionally, improved spelling error detection and correction mechanisms, beyond basic fixes to ofer meaningful rephrasings, could mitigate problems caused by spelling issues, which tutor/experts identified as a notable obstacle during participant observation.
Regarding answer understanding, the fact that a significant percentage of students (56%) reported only occasionally understanding the responses suggests a mismatch between the standard way ChatGPT presents information and the specific presentation needs of this user group. To address this, the interface should ofer customizable output formatting options such as concise responses or summary layout, which many students preferred. Incorporating language simplification features (as recommended by the ACCEDE tutor) and enabling interactive explanations of complex terms or concepts directly within the interface could significantly improve information accessibility. Additionally, tutor/experts observed that students find it dificult to admit that they do not understand something, making them less likely to ask ChatGPT for clarification, highlighting the need for more proactive and supportive interaction design.
The most critical finding of this experiment is the majority’s inability to verify information (only 19% were able to verify the information provided by ChatGPT for the activities), suggesting a limited understanding of efective verification strategies by most participants. Future GAI interfaces should integrate source reliability indicators in a visual and understandable manner. Furthermore, direct links to sources should be provided and predefined prompts for verifying information should be included, guiding users through the information verification process. The interface could also explicitly warn about the probabilistic nature of some answers, encouraging a more cautious attitude toward the information generated.
Suggestions for improvement from students and tutor/experts highlight the need for more intuitive and adaptable interfaces. Incorporating interactive tutorials and contextual assistance directly within the interface could reduce reliance on external help and encourage greater user autonomy (particularly relevant given that only 37.5% of students currently feel confident using ChatGPT independently). The observed increase in confidence in using the technology after the session indicates a positive short-term impact of the intervention, although further research is needed to determine the long-term sustainability of this efect. Interestingly, preliminary findings suggest a potential association between certain cognitive dificulties (alternating attention, short-term memory and working memory) and a greater capacity for verifying information. This unexpected result requires deeper investigation to better understand the mechanisms behind this relationship.
Finally, the observation that no students with moderate or limited disabilities were familiar with ChatGPT suggests a potential digital gap and a need to facilitate access and familiarization with these technologies for students with greater support needs.
These results are consistent with findings from previous work. Like Acosta-Vargas et al. [
This study has certain threats to the validity of our findings that should be considered when interpreting the results: • A significant threat to the external validity stems from the small sample size, which consisted of only 16 participants. Furthermore, the homogeneity of the sample (all students from a specific university-based educational program for individuals with cognitive disabilities) inherently limits the generalization of our findings. While providing rich, in-depth insights for this exploratory case study, these factors mean that the observed patterns of interaction, dificulties, and perceptions regarding ChatGPT’s utility may not be directly extrapolate to broader populations of users with diverse cognitive disabilities, or to diferent educational and social contexts. The limited sample size also constrained the statistical power of certain analyses, impacting the robustness of inferential findings. • Threats to statistical conclusion validity arose from the analysis of categorical data. Specifically, in some analyses, 75% of cells had expected frequencies less than 5. This condition can compromise the validity of Chi-square test. To address this limitation, Fisher’s exact test was used, although its sensitivity is also limited by sample size, potentially afecting the precision of the statistical inferences drawn from these specific analyses. • A potential threat to validity relates to the reliance on self-perception data obtained from studentcompleted questionnaires. Such questionnaires could introduce bias in the perception and reporting of dificulties. Although this limitation is generally attenuated by the parallel questionnaires completed by the tutor/experts, and high overall concordance was observed between their perceptions and those of the students (thereby enhancing the reliability of our findings), instances of discordance highlight a potential divergence in perspectives. This divergence suggests that ‘dificulty’ or ‘understanding’ can be perceived diferently by the user and the observer, which impacts the precise interpretation of these constructs and the full scope of faced barriers.
Addressing these threats to validity will be crucial for future research. Expanding the sample size and incorporating a broader range of objective measures to complement users’ perceptions of their own dificulties will be essential to enhance the generalization and provide more conclusive evidence regarding Generative AI accessibility for individuals with cognitive disabilities.
This exploratory study has revealed the significant dificulties faced by people with cognitive disabilities when interacting with ChatGPT, despite the perceived potential and generally positive attitude toward the tool. Although the presented study focused on ChatGPT, the dificulties identified in asking questions, understanding answers, and verifying information can be extrapolated to other GAIs that employ textbased conversational interfaces such as Gemini, DeepSeek, Claude, or Mistral. Considerations regarding the need for guided interfaces, customizable formatting options, language simplification, and verification support tools are applicable to the inclusive design of any conversational generative AI system intended for users with cognitive disabilities.
These results highlight the need for a design approach focused on cognitive accessibility for GAIs. Future research and development of these technologies should emphasize the creation of interfaces that not only enable interaction but also facilitate understanding, critical evaluation, and autonomy for users with cognitive disabilities. The exploration of multimodal interfaces, advanced customization of information output, and the integration of verification support tools are presented as promising solutions for achieving more inclusive human-computer interaction in the field of Generative AI.
Future work should focus on developing and implementing more guided interfaces, incorporating predefined question templates and contextual suggestions to facilitate query formulation. Implementing voice input and improving spelling error detection and correction with understandable rephrasing are also key areas for development. To improve response comprehension, it is recommended to explore customizable output formatting options, integrate language simplification features, and request interactive explanations. In addition, interfaces could integrate proactive comprehension support mechanisms such as: • Automatic detection of potential comprehension dificulties: Analyze user interaction (e.g., repeated questions, rephrasing of the same question, prolonged periods of inactivity after a complex answer) to infer potential comprehension dificulties. • Automatically ofer simplifications: After an initial response, the interface could discreetly ofer options like “Do you need a simpler version?” or “Would you like a summary of this?” • Automatic breakdown of complex answers: Divide long answers into smaller, more manageable sections, with the option to explore each section in detail. • Key concept highlighting: Identify and highlight key terms or ideas in the answer, ofering the ability to access definitions or explanations with only a click. • Use of analogies or examples: When AI detects a potential dificulty, it may automatically ofer analogies or simple examples to illustrate abstract concepts.
Although some recent versions of ChatGPT occasionally ofer some of these features, such as highlighting key concepts or providing explanations, they are not consistently triggered, nor can users easily control when or how they appear. Even if such features appeared to users during the intervention, the dificulties reported by participants and the observed misunderstandings still hold. Therefore, we consider that integrating these functionalities in a more systematic and user-driven way would reduce the cases in which the user has to explicitly acknowledge their lack of understanding, ofering more a intuitive and proactive support.
The widespread inability to verify information highlights the urgent need to incorporate visual and understandable indicators of source reliability, along with direct links to original sources and predefined prompts to guide the verification process. The unexpected relationships found between certain cognitive dificulties and verification ability point to the need for future research to explore the cognitive mechanisms underlying these findings. Furthermore, more research is needed to determine the long-term sustainability of the increase in technological confidence observed after participating in the study. The limited sample size of this study is acknowledged, and future research with larger samples is necessary to confirm and generalize these findings. Finally, the relationship between disability level and familiarity with AI tools should be further investigated, given the finding that no students with moderate or borderline disabilities were familiar with ChatGPT.
These lines of future work seek to advance the development of more inclusive and accessible GAIs that can truly empower people with cognitive disabilities in their learning and digital autonomy.
This publication is part of the R&D&I project HumanAI-UI, Grant PID2023-148577OB-C22 (HumanCentered AI: User-Driven Adaptative Interfaces) funded by MICIU/AEI/10.13039/501100011033 and by FEDER/UE. We want to thank the ACCEDE program tutor and students for their collaboration, and the support experts for their help during the intervention.
During the preparation of this work, the authors used ChatGPT in order to: grammar and spelling check and paraphrase and reword. Further, the authors used Google Translator for text translation. After using these tools, the authors reviewed and edited the content as needed and takes full responsibility for the publication’s content. with Disabilities Use AI Chatbots, in: Proceedings of the 26th International ACM SIGACCESS Conference on Computers and Accessibility, ASSETS ’24, Association for Computing Machinery, New York, NY, USA, 2024, pp. 133:1–133:7. doi:10.1145/3663548.3688549. [21] S. Roomkham, L. Sitbon, Restarting the conversation about conversational search: exploring new possibilities for multimodal and collaborative systems with people with intellectual disability, in: Proceedings of the 2024 Conference on Human Information Interaction and Retrieval, CHIIR ’24, Association for Computing Machinery, New York, NY, USA, 2024, p. 231–242. doi:10.1145/ 3627508.3638339. [22] J. Jang, S. Moharana, P. Carrington, A. Begel, “It’s the only thing I can trust”: Envisioning Large Language Model Use by Autistic Workers for Communication Assistance, in: Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, CHI ’24, Association for Computing Machinery, New York, NY, USA, 2024, pp. 77:1–77:18. doi:10.1145/3613904.3642894. [23] J. Green, N. Thorogood, Qualitative Methods for Health Research, Introducing Qualitative Methods series, SAGE Publications, 2009. doi:10.7748/nr.13.2.91.s14. [24] V. Braun, V. Clarke, Using thematic analysis in psychology, Qualitative Research in Psychology 3 (2006) 77–101. doi:10.1191/1478088706qp063oa. [25] U. H. Graneheim, B. Lundman, Qualitative content analysis in nursing research: concepts, procedures and measures to achieve trustworthiness, Nurse Education Today 24 (2004) 105–112. doi:10.1016/j.nedt.2003.10.001.