Paper deep dive

Q&A or Document-Based? The Effects of Interface Type on How Screen Reader Users Access Interconnected Documents

Colleen F. Cipriano, Yichun Zhao, Miguel A. Nacenta, Kotaro Hara, Jaylee Soh

Year: 2026Venue: arXiv preprintArea: cs.HCType: PreprintEmbeddings: 125

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%

Last extracted: 8/27/2026, 5:12:48 AM

Summary

This study compares a Question-Answer Interface (QAI) powered by Large Language Models (LLMs) with a traditional Document Interface (DI) for Blind and Low Vision (BLV) screen reader users. Using a within-subjects design with 16 participants exploring fictional worlds, the research found that while the DI led to better knowledge construction (larger, more correct mental models) and broader document exploration, many participants subjectively preferred the QAI and overestimated their learning outcomes. The findings highlight risks of LLM-mediated access creating biased information patterns despite user preference.

Entities (10)

Blind and Low Vision Users · user-group · 99%Document Interface · interface · 98%Question-Answer Interface · interface · 98%Screen Readers · technology · 95%Large Language Models · technology · 92%Mental Models · concept · 90%Concept Mapping · method · 88%Tactile Overlays · technology · 87%Dominion · fictional-world · 85%Solana · fictional-world · 85%

Relation Signals (7)

Question-Answer Interface → comparedwith → Document Interface

confidence 95% · we compared a Question-Answer Interface (QAI)... with a Document Interface (DI)

Blind and Low Vision Users → uses → Screen Readers

confidence 95% · recruited 16 BLV screen reader users

Question-Answer Interface → uses → Large Language Models

confidence 95% · QAI that supports open-ended conversational inquiry... powered by an LLM

Document Interface → producesbetter → Mental Models

confidence 92% · formed larger and more correct mental models with the DI than with the QAI

Document Interface → enablesbroader → Document Exploration

confidence 90% · participants visited more distinct documents with the DI

Document Interface → utilizes → Tactile Overlays

confidence 88% · used a tactile overlay on a touchscreen device for graphical access

Question-Answer Interface → ispreferredby → Blind and Low Vision Users

confidence 85% · many still preferred the QAI

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Blind and low-vision (BLV) users are increasingly engaging with large language model (LLM) interfaces to access documents, but it is unclear how such systems support or hinder their ability to build interconnected knowledge. To examine this gap, we compared a Question-Answer Interface (QAI) that supports open-ended conversational inquiry, with a Document Interface (DI) based mostly on traditional structured text document navigation. We recruited 16 BLV screen reader users where they used both interfaces to explore two fictional worlds. Data from interaction logs, concept maps, decision-based tasks, and semi-structured interviews provide comparative insights into how interface design supports knowledge construction. Findings show that participants visited more distinct documents with the DI and formed larger and more correct mental models with the DI than with the QAI. They were also more able to apply knowledge they had gained. Simultaneously, many still preferred the QAI and often estimated that they had explored more, formed better mental models and applied their models better when acquiring the information with the QAI, despite this not being the case. Our analysis suggests possible interface design reasons for these differences and highlights some of the risks introduced by using question-answer interfaces to access information spaces.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

122,109 characters extracted from source content.

Expand or collapse full text

Q&A or Document-Based? The Effects of Interface Type on How Screen Reader Users Access Interconnected Documents Conference: The 28th International ACM SIGACCESS Conference on Computers and Accessibility; October 25–28, 2026; Vila Nova de Gaia, PortugalThe 28th International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS ’26), October 25–28, 2026, Vila Nova de Gaia, PortugalDOI: 10.1145/3797867.3829036ISBN: 979-8-4007-2521-0/2026/10CCS: Human-centered computing AccessibilityCCS: Human-centered computing Empirical studies in accessibilityCCS: Human-centered computing Accessibility technologies Colleen F. Cipriano Affiliation: Department of Computer Science, University of Victoria, Victoria, British Columbia, Canada email: colleencipriano@uvic.ca , Yichun Zhao Affiliation: Department of Computer Science, University of Victoria, Victoria, British Columbia, Canada email: yichunzhao@uvic.ca , Miguel A. Nacenta Affiliation: Department of Computer Science, University of Victoria, Victoria, British Columbia, Canada email: nacenta@uvic.ca , Kotaro Hara Affiliation: School of Computing and Information Systems, Singapore Management University, Singapore, Singapore email: kotarohara@smu.edu.sg and Jaylee Soh Affiliation: Lee Kong Chian School of Business, Singapore Management University, Singapore, Singapore email: jaylee.soh.2022@business.smu.edu.sg © c Abstract. Blind and low-vision (BLV) users are increasingly engaging with large language model (LLM) interfaces to access documents, but it is unclear how such systems support or hinder their ability to build interconnected knowledge. To examine this gap, we compared a Question–Answer Interface (QAI) that supports open-ended conversational inquiry, with a Document Interface (DI) based mostly on traditional structured text document navigation. We recruited 16 BLV screen reader users where they used both interfaces to explore two fictional worlds. Data from interaction logs, concept maps, decision-based tasks, and semi-structured interviews provide comparative insights into how interface design supports knowledge construction. Findings show that participants visited more distinct documents with the DI and formed larger and more correct mental models with the DI than with the QAI. They were also more able to apply the knowledge they had gained. Simultaneously, many still preferred the QAI and often estimated that they had explored more, formed better mental models and applied their models better when acquiring the information with the QAI, despite this not being the case. Our analysis suggests possible interface design reasons for these differences and highlights some of the risks introduced by using question-answer interfaces to access information spaces. Keywords: Document Accessibility, Mental Models, Conversational User Interface, Exploratory Search, Information Seeking, Blind People, Screen-Readers †c-license: by 1. Introduction The conversational approach to browsing information enabled by generative AI (GenAI) has transformed how people can explore document collections. This emergent tool could benefit those who use screen readers (SRs), most often those coming from the Blind and Low Vision (BLV) community. Traditionally, blind users have relied on mainstream browsing technology and assistive technologies like SRs to access information (Asakawa and Itoh, 1998; Lunn et al., 2011; Perera et al., 2025). Despite extensive efforts to establish accessibility guidelines and develop automated accessibility checkers, many documents remain inaccessible (Jordan et al., 2024). Even technically accessible documents can pose challenges when visual information lacks proper context or description, leaving SR users unaware of content on the page (Bigham et al., 2017). Some expect that GenAI tools could eliminate these barriers (Ernst & Young, 2024); poor alt text becomes less problematic when AI can describe images directly, and one could find information quickly just by prompting, without having to search through lengthy reports. However, current AI-powered tools commonly used by BLV users (e.g., ChatGPT (Atcheson et al., 2025; Adnin and Das, 2024)) often provide targeted answers to pointed questions. This approach risks creating biased access patterns, systematically directing users toward certain types of content while obscuring others (Adnin and Das, 2024; Sharma et al., 2025). This, in turn, could limit users to partial understanding of complex information in a given collection of documents (Tang et al., 2025). That is, we risk creating new accessibility barriers, much like how missing accessibility tags or image descriptions currently hide information for BLV users (Gubbi Mohanbabu and Pavel, 2024; Doore et al., 2024), the design of information access technology could require additional effort into achieving comprehensive document understanding (Adnin and Das, 2024). Better understanding of how GenAI tools influence BLV people’s information exploration and consumption behavior is crucial for the future design of accessible technologies (Tang et al., 2025; Gonzalez Penuela et al., 2025). The overarching goal of this paper is to compare traditional document browsing tools with LLM-driven conversational access tools and understand how SR users explore documents, comprehend the information they contain, and assess which aspects of each tool influence preference. We call the former a document-based interface (DI) and the latter a question and answering interface (QAI). We pose three research questions around these interface types. (RQ1) How does the difference in interface type affect the patterns by which SR users browse documents? (RQ2) How does the difference in interface type influence how SR users integrate and apply knowledge from documents? (RQ3) What aspects of the interface influence SR users’ preference over the other? To answer these questions, we conducted a study with 16 SR users to examine how they interact with a DI and a QAI when exploring multimodal documents about complex, information spaces. We created two fictional worlds, Solana and Dominion, with corresponding document sets. Our mixed-methods study comprised of four phases. In Phase 1, we logged participants’ document browsing patterns with both interfaces. In Phase 2, we asked participants to engage in concept mapping activities to examine how they construct mental models using different interfaces. In Phase 3, participants applied their mental models by answering a decision-based scenario within the worlds. Lastly, in Phase 4, we conducted interviews to record their overall experience and preferences. Our results showed that participants covered a wider range of documents with the DI, whereas the QAI supported more broad connections transitions within a smaller set of documents. Participants produced larger and broader concept maps with fewer errors with the DI, although the QAI concept maps were more densely linked. Interestingly, most participants did not detect these differences, as some preferred DI’s explicit structure and QAI for its conversational style. The subjective perceptions of the two interfaces were divided, suggesting divergence in exploration behaviors, strategies, and experiences. This study primarily contributes: (i) an empirical examination of how SR users browse and synthesize multimodal information using a DI and a QAI; and (i) insights on how these information access interface types shape knowledge construction and application. Given how long SR users have depended on mainstream browsing tools, we seek to advance understanding of how GenAI can influence and change the future of accessible information access. 2. Background and Related Work We first introduce the concept of mental model, which we extensively leverage in this paper to conceptualize how people comprehend documents and to operationalize our measurement of this comprehension. Then we review document-based interfaces and conversational user interfaces, and the efforts in making these interfaces accessible for SR users. Finally, we introduce how GenAI technologies have enabled the conversational user interfaces. 2.1. Mental Models and Mental Model Elicitation We define mental model (based on (Norman, 1983; Gentner and Stevens, 2014; Simon, 1978)) as an internal representation that allows a person to organize knowledge gained about the world and act in consequence to this knowledge. Mental models are commonly used to examine cognition in applied and theoretical areas (e.g., (Jones et al., 2011; Saucier and Dobmeier, 2025)). A mental model offers a lens through which we can examine cognitive activities that are otherwise hard to observe. For example, learners construct new mental models “when confronted with new learning tasks” (Bucciarelli and Cutica, 2012; Nadkarni, 2003). Methods for extracting these models include concept mapping and card sorting, each appropriate for different scenarios (Harper and Dorton, 2019). Prior work on blind users’ mental models has established that non-visual interaction is shaped by the structure of the internal representation of a system, but has mostly examined desktop environments and screen readers through verbal protocols, observations, and performance data, rather than externalized representations (Kurniawan and Sutcliffe, 2002; Abidin et al., 2012). Moreover, work in this area emphasizes the difficulty of eliciting mental models directly (Saei et al., 2010). In this paper, we use concept mapping as a way to assess the cognitive process of SR users as they engage with a corpus of documents and develop comprehension. Concept mapping is a way to externalize how people organize, connect, and build comprehension by representing concepts as nodes and their relationships as links (Bias et al., 2015; Gerken et al., 2011; Sanchez and Flores, 2010). Adopting on earlier critiques that verbal protocol alone is incomplete (Abidin et al., 2012), concept mapping allowed us to examine construction of internal representations, beyond just measuring retrieval from memory. 2.2. Accessing Documents and Information Access Technologies In traditional information retrieval literature, users’ tasks in accessing a corpus of documents are broadly divided into search and browsing (Baeza-Yates and Ribeiro-Neto, ; Croft et al., 2010; Agosti and Croft, 2008; Saracevic, 1997). In search, the goal is to find relevant information or a list of documents based on a search query (Preininger et al., 2021; Kim et al., 2025). In its most basic form, search prompts a user to enter a query, and the information access technology (e.g., a search engine) returns the results (Lewis and Jones, 1996). While search has evolved to accommodate more natural language queries to handle greater ambiguity (Moore, 2018), and support iterative refinement conversationally with the use of conversational user interfaces (CUIs) (Hall, 2018; Kelly, 2009; Saracevic, 1997; Dhingra et al., 2017; Kim et al., 2025), the fundamental interaction type of “query-and-get-results” paradigm remained essentially unchanged until recently. Browsing represents a more exploratory approach to information access, which is associated with learning and investigative sense-making (Marchionini, 2006)—activities that are more aligned with the focus of our work. Users typically lack a specific query in browsing; instead, they may seek to understand general concepts within documents or gain an overview of available information before conducting more targeted searches (Athukorala et al., 2016; Marchionini, 2006). Browsing occurs through document-based interaction. When navigating the Web, for instance, sighted people use browsers through direct manipulation, while SR users additionally rely on assistive technologies (Asakawa and Itoh, 1998; Lunn et al., 2011). The emergence of GenAI-based tools follows a broader shift in information retrieval toward treating search as a text-to-text transformation problem (Mo et al., 2025; Suri et al., 2024; Yang et al., 2025; Holtz et al., 2024; Preininger et al., 2021; Sowa and Przegalinska, 2025). Compared with traditional CUIs, these systems can interpret nuanced input, generate human-like responses, and—crucially for document access—–synthesize text across multiple documents and maintain conversational context (Ai et al., 2025; M.P.Geetha et al., 2024). Furthermore, tools like NotebookLM11 1 https://notebooklm.google.com and Elicit22 2 https://elicit.com enable information exploration within document corpora through iterative question-answering, and prior work shows that people do ask open-ended questions to navigate large collections without explicit queries (Atcheson et al., 2025; Adnin and Das, 2024). However, we still know little about what happens when an LLM becomes the intermediate layer between users and documents, especially in accessibility-centered exploration of interlinked documents. This gap matters because the same properties that make GenAI tools appealing may also reshape access in limiting ways. Beyond general concerns like GenAI inducing overreliance (Passi and Vorvoreanu, 2022; Bainbridge, 1983), and difficulty and cognitive burden associated with prompt-based interaction (Tankelevitch et al., 2024; Dang et al., 2023; Lehmann and Buschek, 2024; Subramonyam et al., 2024), the dominant “prompt-and-get-summary” interaction encourages targeted information extraction and may reduce exposure to a wider range of documents (Pucci et al., 2023; Furini et al., 2020; Baez et al., 2022; Kodandaram et al., 2024). Prior work has linked this to weaker exploration, comparison, and continuity of access (Torres et al., 2019; Baez et al., 2022; Fast et al., 2018). Using the stratified model of information retrieval as an analytical lens, we expect that effort is made toward prompt formulation and summary interpretation at the affective layer (Saracevic, 1997), while reducing direct interaction with the document structure at the interface layer. If we unknowingly design GenAI-powered exploration tools that could bias information access, we risk creating new barriers for SR users; just as lack of appropriate document tags limits access to information, the design of information access technology may obscure parts of a collection and require additional effort into achieving comprehensive understanding. Therefore, in this work, we conduct a study comparing two interfaces—document-based interface and CUI-based question and answering interface powered by GenAI—that manifest different types of interaction. 2.3. Document Accessibility Our goal is to understand the effects of interaction types on how SR users explore and comprehend information in documents. This requires us to isolate the accessibility of documents and interfaces by making them as accessible as possible. The development of accessible document-based interfaces and study of information access using these tools has been a significant focus for the HCI and accessibility research communities. For example, past work has explored ways to present and extend alt-text or image descriptions to enrich screen reader-friendly interaction with static content (Jung et al., 2022; Zong et al., 2022a; Zhao et al., 2024; Thompson et al., 2023; Chheda-Kothary et al., 2025; Zong et al., 2022b). Tactile graphics offer another means of access: raised-line representations of visual materials that can be explored through touch (Mukhiddinov and Kim, 2021; Gupta et al., 2017). For tactile graphics there are guidelines that can aid in their creation (Wabiński et al., 2022; Erp, 2002; Butler et al., 2021). In our study, we created the document-based interface by learning from past interface designs and following existing guidelines. To make our document-based interface accessible, we provided sufficient annotations for screen-reading access and used a tactile overlay on a touchscreen device for graphical access. Early conversational assistants like Alexa and Siri opened the door to voice-based, hands-free information access for SR users. These assistants were appreciated for quick and simple tasks, which lowered barriers to accessing digital content (Oumard et al., 2022; Pradhan et al., 2018; Abdolrahmani et al., 2018). Today’s LLM-based interfaces can further facilitate access to documents (Perera et al., 2025). Rather than navigating documents line-by-line, SR users can query LLMs for descriptions, summaries, or clarifications, enabling them to extract and synthesize information more efficiently. This shift in information access allows SR users to direct their navigation path instead of adhering to rigid document structures (Adnin and Das, 2024). At the same time, LLMs may also introduce problems by limiting access to a subset of content, making it difficult to ensure complete rather than fragmented understanding (Pucci et al., 2023; Furini et al., 2020; Baez et al., 2022; Kodandaram et al., 2024). 3. Study Methodology To address the RQs from the Introduction, we designed an empirical study to observe SR users navigating unfamiliar information spaces, building mental models of those spaces, and applying the gained knowledge. The study took place over two laboratory sessions in which SR users worked with two interfaces, one based on traditional hyperlinked documents (the Document Interface—DI), and one based on a question-answer paradigm powered by an LLM—QAI). The laboratory setting allowed us to gather detailed data and control our observations. Participants experienced both interfaces (a within-subjects design), allowing explicit comparison. Horizontal process map of the study. The top row shows four phases: Exploration, Verbal Elicitation and Concept Mapping, Decision-Based Task, and Interface Reflection and Comparison. The middle row lists the activity and duration for each phase. The bottom row summarizes the main outcomes, with curved connectors linking concept listing and knowledge application to the shared insight on organization, integration, and application of knowledge. Figure 1. Overview of the study procedure and measurements. Participants completed all phases for each of the DI and QAI conditions in a counterbalanced order. At the end of the second session in Phase 4, participants completed an additional set of interview questions comparing the interfaces.Horizontal process map of the study. The top row shows four phases: Exploration, Verbal Elicitation and Concept Mapping, Decision-Based Task, and Interface Reflection and Comparison. The middle row lists the activity and duration for each phase. The bottom row summarizes the main outcomes, with curved connectors linking concept listing and knowledge application to the shared insight on organization, integration, and application of knowledge. Figure 2. Experimental setup for the DI condition showing the tactile overlay on top of the tablet (left) on a non-slip mat, with the laptop showing the Dominion interface (right).Setup for the DI condition. On the left, a tablet with a tactile overlay is placed on a non-slip mat next to a laptop, on the right, displaying the Dominion interface. 3.1. Study Materials: Fictional Worlds An important concern in this kind of investigation is the large potential effect of familiarity on people’s behavior and understanding. For this reason, we created two information spaces from scratch, guaranteeing no familiarity. These take the form of interconnected documents, similar to Wikipedia’s hyperlinked pages but describing fictional worlds. The document sets followed the following criteria: (i) minimize the influence of participants’ prior knowledge, (i) encompass sufficient diversity of information and of document types to reproduce real-life complexity, and (i) provide sufficient depth for extended exploration. We iteratively composed the Solana and Dominion worlds in a research approach related to Kieras et al.’s (Kieras and Bovair, 1984), informed by existing methodologies in fictional world building (Fischer and Mehnert, 2021). Each document set included 25 documents, including an index. Out of the 25 documents of each world, 8 (Solana) and 6 (Dominion) were diagrams representing spatial or conceptual aspects of the fictional worlds (e.g., a map of the territory and a diagram of family group configurations). The textual representations resemble wikipedia articles, with headings and meaningful inter-document links. Each document describes an aspect of the world, such as history, governance, and culture. We included spatial representations because they are common in real documents and are recognized as important to the BLV community (Elmqvist, 2023; Zhao et al., 2024). To make spatial documents accessible to our participants, we designed 3D-printed tactile overlays following best practices (Butler et al., 2021; Gupta et al., 2019), considering BANA guidelines (Braille Authority of North America, 2022) and current research on tactile representations (Butler et al., 2021; Gupta et al., 2019). We iterated over the designs three times to make them representative of the best accessible spatial representations currently available to the community (He et al., 2025). Although interactive printable tactile overlays are not currently widespread, they are not very different from printed materials that many in the BLV community are familiar with (e.g., embossed tactile prints) and may soon become practical and affordable (Barros et al., 2023). Excluding spatial information from the DI would have artificially constrained the DI from presenting spatial structure in general. The implications of this choice are discussed in the Limitations Section below. The documents that form the information spaces of both worlds, including the visual diagrams and their 3D-printed versions are open sourced for use in subsequent research and shared in the Supplementary Materials. 3.2. Interfaces (Main Condition) The two interfaces (DI and QAI, shown in Figures 3 and 5, provided access to the fictional worlds with the interaction paradigms that we want to compare: direct corpus exploration and LLM-mediated question answering. We designed the conditions to isolate these access paradigms so that the observed differences could be attributed to the interaction approach. We strove to implement state-of-the-art interfaces that are the best reasonable alternatives available or soon to be available. 3.2.1. Document Interface (DI) To minimize the need to train participants and maximize ecological validity, participants used their own device and screen reader (SR) to access textual documents. These were web pages meeting the Web Content Accessibility Guidelines—WCAG 2.1 (WCAG, 2025) (see Figure 3). The DI supported index and in-page search with the participants’ own SRs. The spatial documents were accessed as the 3D-printed paper diagrams discussed above, overlaid over a tablet (Samsung Galaxy Tab A9+— see Figures 4 and 2). Moving the finger over the overlay produced verbal descriptions of manually labeled areas, in addition to the static relief of the 3D-printed overlay. When a participant followed a hyperlink in a textual document connected to a spatial document, they received a verbal notification to switch to the tablet. If the spatial document had changed, the experimenter manually placed the corresponding document’s overlay. The spatial documents also linked to text documents: when tapped, the browser on the main computer jumped to the corresponding document. The tactile representations preserve spatial relationships that are difficult to communicate through SR navigation alone. Their inclusion reflects our goal of implementing DI as a realistic form of document access. We considered several alternatives for spatial document display such as dynamic pin displays, but decided against because of their low resolution and reliance on Braille, which is not universal among BLV users. Although printable tactile overlays are not currently widespread, they are not very different from printed materials that many in the BLV community are already familiar with (e.g., embossed tactile prints) and may soon become practical and affordable (Barros et al., 2023). Figure 3. An example of one of the documents in the DI interface for the world of Solana.A sample document in the DI trial. The website page consists of headers, paragraphs, an overview, and a large main content area with a main heading called 'The Doctrines of Solana.' There re also hyperlinks, providing additional references within the document. Figure 4. Examples of tactile overlays used in the study: a map (left), a diagram of relationships (center), and a set of icons (right).Tactile overlays. It shows three types of tactile overlays arranged horizontally. The left one represents a map, the center shows a node-link diagram, and the right presents a pyramid of symbols. 3.2.2. Question-Answer Interface (QAI) The QAI was a WCAG 2.1-compliant web-based (WCAG, 2025) chatbot accessible through participants’ devices and screen readers. Participants could enter questions by typing or using speech input. The system returned information in multiple formats: text, structured bullet points, tables, detailed image descriptions when relevant, and direct quotations from the source documents (Figure 5). The interface was designed to retrieve and summarize information only from the existing knowledge base. Figure 5. The QAI interface for the world of Dominion showing a user query and a structured tabular response.The QAI interface with the Dominion guide. A user asks, "Who are the leaders in Dominion?" with the system responding with a table listing names, seats, and tiers of ten members. The interface contains a heading and a short description of the Dominion. The underlying conversational system was built using Python with Flask, integrated with Google’s Gemini API (Gemini 2.5 Pro). To constrain conversations to information in the documents we used a custom prompt with cache-augmented generation (CAG) (Chan et al., 2025), loading the corpus of 25 documents to the context window (including the spatial documents). CAG is well-suited to closed knowledge bases because it reduces retrieval latency and errors. When direct answers were unavailable, the QAI suggested related documents to stay within content boundaries. For out-of-scope questions (e.g., “Do you have Oreo cookies in Solana?”), it explicitly stated that the information was unavailable, offering relevant alternatives such as “The archives do not mention that topic. If you’re interested, there is detailed information on Solana’s economy or its annual festivals.” We did not observe any instances of hallucination or erroneous responses. 3.3. Participants We recruited 16 participants (age ≥ 19) who reported regular use of at least one screen reader (e.g., JAWS, NVDA, VoiceOver, TalkBack). Recruitment took place through community organizations, social media, and snowball sampling. Participants reported varied GenAI uses (Table 1), including accessibility features (n=11n=11), work or productivity (n=8n=8), technical assistance (n=7n=7), writing or content generation (n=5n=5), and information seeking (n=4n=4). Participants received $40 per session. The study was approved by the local ethics board. Table 1. Self-reported demographics of the study participants. ID Age Visual ability Occupation GenAI tools used GenAI usage 1 40–44 Totally blind Program coordinator ChatGPT, Copilot, Be My AI Daily 2 19–24 Totally blind Voiceover artist ChatGPT, Copilot, Gemini, Be My AI Daily 3 50–54 Low vision Unemployed ChatGPT, Be My AI Daily 4 55–59 Totally blind Administrative work ChatGPT, Gemini, Be My AI 2–3 times a week 5 55–59 Totally blind Retired Be My AI Daily 6 25–29 Totally blind Unemployed ChatGPT, Be My AI Daily 7 70–74 Totally blind Instructor Be My AI Daily 8 65–69 Totally blind Bookkeeper None N/A 9 55–59 Low vision Retired None N/A 10 19–24 Totally blind Accessibility tester ChatGPT, Copilot, Gemini, Be My AI Daily 11 30–34 Low vision Marketer None N/A 12 50–54 Totally blind Trades ChatGPT, Gemini, Be My AI Daily 13 65–69 Totally blind Administrative work None N/A 14 55–59 Totally blind International relations Copilot Daily 15 55–59 Totally blind Client scout Gemini, Copilot Daily 16 45–49 Totally blind Administrative work ChatGPT, Gemini, Be My AI Daily 3.4. Procedure Each participant completed two 90-minute sessions on separate days at a location of their choice (e.g., public library, participant home, community center). We counterbalanced interface order (DI vs. QAI) and world order (Solana vs. Dominion) across participants. Given the study duration (Figure 1), sessions were separated to reduce cognitive load, fatigue, and carryover effects, consistent with prior cognitive work (Keppel, 1991; Hornbæk, 2013). To preserve comparative judgments, we capped the interval between sessions at five days (Cepeda et al., 2006). In practice, participants completed the second session one to three days later, depending on scheduling and location constraints. This also reflects best practices for working with BLV individuals by supporting flexible scheduling, agency in choosing comfortable locations, and use of their preferred device and screen readers. Each session proceeded with four phases in Figure 1. Phase 1. Participants had to follow their interests and openly explore the world using the assigned interface, while developing an overall understanding of the world. We explicitly told them that recall was not the goal and asked them to prioritize acquiring domain knowledge over constrained task performance (Sarrafzadeh et al., 2014; Soufan, 2023; Marchionini, 2006). Before starting their QAI session, participants were informed that the system has an underlying content of 25 documents. Participants could take notes, but none chose to. Phase 2. Participants had to describe the world as if to a friend, highlighting what they found most important. From their description, they listed important concepts and explained how concepts relate to each other. Using these responses, the researcher constructed a concept map and asked probing questions to address gaps or overlooked relationships. The map was then validated with the participant and refined as needed to ensure that it accurately reflected their understanding. The researcher followed a structured script and added or connected only concepts and relationships introduced by the participant in order to prevent bias in the analysis. Participants did not have access to the assigned interface because doing so would have shifted the evaluation from their mental models to assisted performance. This restricted interface access means that the experimental design aligns with our formulation of RQ2. This phase implements a mental model elicitation technique similar to interview-based concept mapping method (M. et al., 2018). The assisted concept mapping exercise, in which the experimenter built a graph of concepts and connections is designed to capture mental model organization beyond recall measures (Gerken et al., 2011; Bias et al., 2015). Phase 3. Participants applied their understanding by answering a question from a world-specific scenario (e.g., “As Solana expands, kinship traditions are becoming difficult to maintain. The Luminaries are debating whether to modernize certain laws or preserve Kinship laws strictly. You’ve been asked to propose a policy on this”) which tested how effectively they could use their gained knowledge. Decision-based scenarios complement the model elicitation procedure of Phase 2 by examining applied reasoning (Paolo and Warglien, 1999; Byström and Järvelin, 1995). Phase 4. Participants completed a semi-structured interview about their interactions. At the end of each session, we asked about ease of use, challenges, and overall experience. After the second session, participants additionally answered comparison questions, indicating overall preference, perceived support for exploration and learning, and confidence in answering, accompanied by free-form explanations. 3.5. Measurements and Analysis The two sessions yielded quantitative and qualitative data to address our research questions. Figure 1 indicates which measurements correspond to which phases of the experiment. 3.5.1. Document Traversal Data We logged the visited documents and transitions across the information space in Phase 1 of each session. We analyzed and plotted how participants moved within each interface. The quantitative measures extracted from this phase are: • Unique Document Visits: A count of distinct documents counted as visited at least once. • Document Visits: The total number of visits, including revisits. • Unique Document Transitions: A count of distinct document-to-documenttransitions traversed at least once. • Document Transitions: The sum of all transitions made between documents. • Transitions per Document: The transitions normalized by the number of unique document visited. Because the interfaces support different actions, visits and transitions, we necessarily had to operationalize their counts differently across interfaces maintaining equivalence as much possible. In DI, a visit occurs upon opening a document. In QAI, we manually mapped each response to the documents where its information appeared. We counted a source document as visited when the response included information corresponding to five or more contiguous sentences from that document; shorter references were treated as summaries and were not counted as visits. We chose this threshold to distinguish substantial exposure to a document’s content from a brief mention or summary. It also reflects the DI structure, where each document is preceded by a short descriptive paragraph before the main content. The index document also contains a summary of the document, which therefore allows extracting information without actually visiting it. A transition represented movement between source documents. In DI, transitions are link traversals. In QAI, transitions occurred when the source-document mapping changed within a response or across follow-up responses. Although QAI transitions were mediated by the system’s retrieval, they remained user-driven as they resulted from follow-up and compound questions. Self-loops were excluded for comparability, where rereading the currently selected document in DI does not generate a new transition. With the visits data, we constructed graphs that describe the navigation pattern of each session. Topics are nodes and transitions are edges. From the graphs, which we plotted for visual qualitative analysis, we additionally extracted the following quantitative measurements (all common graph measurement described in (Weth and Hauswirth, 2013; Broder et al., 2000; Yoon and Jetter, 2016; Nousiainen and Koponen, 2010): • Density: Measures of overall interconnectedness. Higher values correspond to more tightly connected graphs. • Directed Diameter: Measures the breadth of the graph. Lower values correspond to more tightly connected graphs. • Degree of Variance: Measures the evenness of connectivity. Higher values correspond to more unevenly connected graphs. • Average shortest path length: Measures the average number of steps needed to move between concepts. Lower values correspond to more tightly connected graphs. 3.5.2. Concept Maps In Phase 2, we documented each topic and subtopic mentioned by the participant, and the explicit connections they made to form the concept maps. We collected this data as graphs, which we visualized and analyzed to examine the structure of gained knowledge and mental models following Paul (Paul, 2014) and Haque et al. (Haque et al., 2023). Topics and Subtopics: A topic corresponds to overarching categories in the worlds represented in a document (e.g., governance, geography (map), or the community structure). Subtopics refer to subsections within documents or facts that fit under a topic such as names, roles, dates, or landmarks. We counted a topic/subtopic when the participant made an unambiguous mention that could be mapped to the predefined documents. We did not count repeated mentions of the same topic/subtopic as new instances, only newly stated connections. References to information beyond the content were marked as inaccurate by the researcher, ensuring completeness since they could still link to new topics or subtopics. Number of Connections: We counted the explicit connections from the topics/subtopics mentioned by participants and any additional connections were verified by the researcher through probing. We also calculated the number of connections per topic. Error Metrics: We counted errors in the elicitation tasks to assess the alignment of participant responses with the actual content of the documents. Errors included statements that contradicted the source documents, unsupported inferences participants made without evidence, and incorrect recall of specific details. Repetitions of the same error were not counted as new, unless they added a new incorrect connection or distinct claim. Graph Metrics: We calculated the same structural metrics as with the exploration graphs (Subsection 3.5.1) to assess the properties of participants’ mental models. 3.5.3. Decision-based Tasks We analyzed the decision-based task using the same metrics as the concept mapping task: topics and subtopics count, number of connections, and error metrics. We did not apply graph measures here because the graphs were substantially sparser and some participants were not able to produce responses. We discuss why in the Results section. 3.5.4. Qualitative Responses The qualitative analysis drew on data sources across research questions. For RQ1, we used participants’ interaction logs, queries, QAI responses, and graph visualizations, with relevant interview responses to understand exploration patterns, prompting strategies, and differences between interfaces. For RQ2 and RQ3, we conducted thematic analysis (Braun and Clarke, 2023; Braun and Clarke, 2021) of the interview reflections and comparison (Phase 4), adopting an inductive-deductive coding approach (Adams et al., 2008b; Adams et al., 2008a; Dix, 2020). Coding was guided by the RQs while remaining open to emerging concepts. To develop the initial codebook, the first author began by open-coding four transcripts from two participants selected to capture variation in interface use and preference. Initial low-level codes were grouped into preliminary categories to code 12 transcripts more. We discovered new codes, and revisited transcripts to apply them. After this stage, the sample included a balanced set of participants who favored each interface. The research team met regularly to refine and develop themes across the categories of codes. The second round involved collectively collapsing redundancies and adding codes that aligned with the RQs. Specifically, we examined whether responses provided explanations for the quantitative results, such as insights on knowledge construction (RQ2). The third round involved revision and grouping codes into broader themes that reflected patterns in the data. We agreed on a final codebook of 93 codes that was applied deductively to the remaining 16 transcripts, with earlier transcripts revisited as needed to ensure consistency across the data. Additionally, participants made direct comparisons by selecting which interface they preferred overall, which they believed supported greater exploration, more learning, and enabled more confident answers. These perceptions were followed by open-ended prompts asking to explain their choices. 3.6. Positionality Statement None of the authors identify as BLV. Nevertheless, we intentionally ground our study design, analysis, and interpretation in participants’ own accounts, practices, and perspectives. Our work centers around the lived realities of SR users, ensuring that RQs, metrics, and conclusions reflect what participants identify as meaningful and relevant. The second author has more than six years of sustained engagement with BLV communities. With the fourth author’s experience in assistive technology design, these informed the study design, accessibility of the study materials and attention to the screen-reader practices of participants. The second author’s qualitative research background guided the coding process. The third author’s expertise in knowledge elicitation and interaction techniques informed our RQs, design and interpretation of elicitation tasks, and the analysis of how participants organized and applied information. The quantitative expertise of the third and fourth authors informed the selection and interpretation of our statistical analyses. We view technology as an active mediator that shapes how people access, explore, and make sense of information. Drawing on the relational model of disability (Thomas, 2004; Reindal, 2008), while technology can reduce barriers, we believe restrictive design choices can also actively create disability and impose dysaffordances (Costanza-Chock, 2020). 4. Results This section presents the study evidence organized by the RQs. Within each subsection, we consider first analysis of quantitative measurements and then evidence from a qualitative analysis. 4.1. How do People Explore Information Spaces with the Different Interfaces? (RQ1) 4.1.1. Quantitative Results. The quantitative evidence for RQ1 is based on statistical comparisons of key measurements represented in Figure 6 and Figure 7. When counting the unique visited documents, we see that participants tended to visit more documents with the DI (μdocuments,DI=11.44 _documents,DI=11.44, σdocuments,DI=2.58 _documents,DI=2.58) than with the QAI (μdocuments,QAI=9.31 _documents,QAI=9.31, σdocuments,QAI=2.68 _documents,QAI=2.68), a 22.9% increase, (t⁡(15)=2.34t(15)=2.34, p=0.03<0.5p=0.03<0.5,η2=0.27η^2=0.27). Because documents correspond to topics, this generally supports that exploration was broader with the DI than with the QAI. This is despite the fact that our measurements of total document visits were similar (participants did roughly the same number of visits on average). We expected that the QAI’s nature would lead to more transitions between documents, or at least, more different instances of transitions between documents/topics. This is a property inherent to the QAI: when answering a query, the LLM can output material from multiple documents within the same answer (which we counted as transitions), whereas in the DI transitions can only be done in pre-established ways, such as navigating using hyperlinks between the documents or through the index document. Despite this, the counts of transitions between documents were not conclusively different (all p>0.05p>0.05 in Figure 6 b, d and e). However, a second set of measurements focuses on whether differences in the interface resulted in a tighter interconnection of more distant topics (Figure 7). This time the answer is positive. When measuring the density, directed diameter and average shortest path length of the navigation graphs (nodes are topics/documents and links are transitions between those topics) we see that the QAI enabled tighter connections between different topics as measured by a higher density of the navigation graph for QAI (μdensity,QAI=0.21 _density,QAI=0.21, μdensity,DI=0.13 _density,DI=0.13), a lower directed diameter (μdiameter,QAI=4.94 _diameter,QAI=4.94, μdiameter,DI=7.75 _diameter,DI=7.75), and a lower average shortest path length between two topics (μaspl,QAI=2.44 _aspl,QAI=2.44, μaspl,DI=3.77 _aspl,DI=3.77). 4.1.2. Qualitative Results. We used answers from the interviews, visualizations of the exploration patterns, and analysis of the participants’ questions and the QAI’s answers to make sense of how the two interfaces differ and to contextualize the quantitative results from the previous section. We report counts (e.g., n=5n=5, where n denotes the number of distinct participants that fit the finding) to make the distribution of coded observations transparent, facilitate comparison across findings, and strengthen analytical credibility (Nicmanis, 2024; Noble and Smith, 2025). We assessed the importance of each theme based on its relevance to our research questions and its relationship to the observed differences. Our qualitative analysis of participants’ explorations revealed three recurring subpatterns: overview (learning the overall structure of the information space), depth (learning in increasing detail about a particular topic), and fact finding (finding answers to specific questions). These align with existing frameworks in exploratory search, information foraging, and sensemaking literature (Shneiderman, 1996; Marchionini, 2006; Marchionini and Shneiderman, 1988) where similar knowledge-oriented intents have been documented as fundamental for complex information spaces. While these behaviors represent general patterns of engagement, we observed that for both overview and depth it did matter whether participants were content with the predetermined structure provided in the documents (i.e., the hierarchical organization and sequential order provided by the document author) or wanted to access the information in a self-defined organization. For fact finding this does not seem to matter that much. For example, in our worlds, the documents are structured according to a traditional textbook-like set of topics such as history, politics, or culture. However, there are other ways to slice and index the same information space, such as organizing the descriptions as a storyline, or focusing on the influence of specific historical figures. Although both interfaces can support any of the subpatterns (overview, depth, fact finding), the data suggests that the effectiveness of the interface depends largely on whether the participant is ready to leverage the pre-defined structure of the documents or seeks a different one, as reflected in Table 2. The following paragraphs discuss evidence for each of the Table’s cells. Overview Predetermined Overview Self-Defined Depth Predetermined Depth Self-Defined Fact Finding DI Easy Hard Easy Hard Hard QAI Hard Easy Hard Easy Easy Table 2. Summary of the difficulty of using the DI and QAI across the exploration subpatterns, based on qualitative data.Qualitative summary of which interface participants found easier for each exploration intent. The table compares DI and QAI across overview and depth tasks under two conditions: predetermined and self-defined. The table also includes fact finding. Green cells mark the interface reated easier for that subpattern; red cells mark the interface rated harder. Overview-Predetermined. Participants (n=9n=9) identified advantages of the DI when they were willing to accept the existing structure of the documents. P2 observed that they “could look at a bunch of categories and decide what to read,” reflecting the traditional way to familiarize oneself with the information space, akin to skimming the main subsections of a Wikipedia page. P12 echoes this: “The information was there. It told me the links, told me 25 in a list, gave me an idea of what I was getting into.” In contrast, the QAI presents some difficulty, since a predetermined organization is not immediately apparent, and participants (n=9n=9) did not necessarily know how to obtain an overview. P4 observes: “I didn’t know what questions to ask it because I didn’t know what information would be available to me.” Some participants did not even realize that an overview could be useful (n=6n=6), as P7 reflects, “I have no idea what the parameters are. I guess I could have asked it a question like, give me the main topics.” Overview-Self-Defined. When participants (n=5n=5) are, instead, interested in their particular overview of the information, DI can be frustrating because it requires them to define their own categories: “It felt like reading a book and trying to navigate through the different pages.” (P11) This is much easier with the QAI, because participants (n=7n=7) can ask the interface to organize information based on their interest. P1 explains that they got “to determine what I was going to look into and get those big concepts. The other interface [DI] was a bit overwhelming,”. With QAI, they were able to request cross-document summaries based on their overview preferences without having to check every document. Depth-Predetermined. When participants were looking for depth, opinions are parallel. DI is great if they do not have a particular view on how to dig deeper, and are comfortable following the document’s structure (n=7n=7). Referring to what a document has to offer, P10 notes: “I know there’s a container full of information in each that I can just explore.” P12 echoes this, “at least there was a container that I knew all the information was in, and I knew how big it was.” In contrast, this is harder to do with QAI (n=8n=8): “What happens if I keep asking for more detail, it’s eventually gonna run out of information about this one thing, so now what?” (P11) Even when an answer felt detailed, the nature of the question-answer cycle makes it difficult to know when a topic has been exhausted. Depth-Self-Defined. Nevertheless, if participants are not interested in the grouping (or “narrative”), then DI required readers to manually trace their path across documents and pursue a topic in detail based on what they deemed relevant (n=9n=9). P4 observes, “…it led me to a whole different path of where I wanted to go to understand the world.” This made self-defined depth possible, but more effortful to keep track of. Conversely, the QAI strongly supports presenting information in whichever way the participant asks (n=7n=7). For example, participants asked questions such as “How is life governed according to the full cycle?” (P4) or “Can you explain to me how is it from 12 to 50 years?” (P9). P16 explicitly highlights this advantage, “you can just keep asking questions and isolate on specific topics,” without having to look for relevant pieces from each document. Fact Finding. There is no distinction between whether participants are interested or not in the predetermined organization of the document. The QAI (n=9n=9) easily provides contextualized facts as answers (i.e., and advanced search feature) that the DI would require the reader to hunt throughout multiple documents (n=7n=7). This advantage with QAI over DI is acknowledged by P4, “you’re able to get your answers as opposed to having to read 30 pages to find out.” 4.1.3. RQ1 Synthesis The quantitative data show that DI enables a wider coverage of the information space, whereas QAI supports a more tightly connected set of transitions between the different topics. The qualitative analysis helps us connect those results to the fundamental differences between the interfaces: there is value in having access to a pre-established structure that readers can just consume, which is much easier to do with DI. Participants struggle to access structure with QAI because they have to come up with appropriate queries about a structure and information space that they do not know much about in advance. However, when the user wants to access the information space in more sophisticated or custom ways, such as based on personal interest, then the QAI can significantly facilitate this process by crafting coherent answers that comprise information scattered around different areas of the document. Violin plots comparing DI and QAI on five measures. Unique document visits averaged 11.44 for DI and 9.31 for QAI, p = 0.03. Document transitions averaged 16.31 for DI and 16.81 for QAI, p = 0.85. Document visits averaged 17.31 for DI and 17.81 for QAI, p = 0.85. Unique document transitions averaged 9.56 for DI and 12.63 for QAI, p = 0.06. Transitions per document averaged 1.49 for DI and 1.77 for QAI, p = 0.33. Figure 6. Exploration metrics (RQ1—Phase 1). DI is red, and QAI blue. Numbers above represent the averages. Error bars are standard error. Each black dot is a session.Violin plots comparing DI and QAI on five measures. Unique document visits averaged 11.44 for DI and 9.31 for QAI, p = 0.03. Document transitions averaged 16.31 for DI and 16.81 for QAI, p = 0.85. Document visits averaged 17.31 for DI and 17.81 for QAI, p = 0.85. Unique document transitions averaged 9.56 for DI and 12.63 for QAI, p = 0.06. Transitions per document averaged 1.49 for DI and 1.77 for QAI, p = 0.33. Violin plots comparing DI and QAI on four graph measures. Density averaged 0.13 for DI and 0.21 for QAI, p = 0.00. Directed Diameter averaged 7.75 for DI and 4.94 for QAI, p = 0.00. Average Shortest Path Length averaged 3.37 for DI and 2.44 for QAI, p = 0.00. Degree Variance averaged 2.40 for DI and 2.77 for QAI, p = 0.69. Figure 7. Exploration graph metrics (RQ1—Phase 1). See caption of Figure 6 for details.Violin plots comparing DI and QAI on four graph measures. Density averaged 0.13 for DI and 0.21 for QAI, p = 0.00. Directed Diameter averaged 7.75 for DI and 4.94 for QAI, p = 0.00. Average Shortest Path Length averaged 3.37 for DI and 2.44 for QAI, p = 0.00. Degree Variance averaged 2.40 for DI and 2.77 for QAI, p = 0.69. 4.2. How do the Different Interfaces Influence Integration of Knowledge? (RQ2) If RQ1’s results address how people behaved with the interfaces, RQ2 focuses on the consequences. Were participants able to build, connect, and apply what they explored, integrating it into their mental models? (Phase 2) or when using it in scenario-based responses (Phase 3)? We look at each of the two phases quantitatively first, then discuss the qualitative evidence together. 4.2.1. Concept Maps (Phase 2) We asked participants to recreate their mental structures of the worlds they had explored (the elicitation process is described in Section 3.4). The result of this process is a mental model graph for each session and participant that reflects their understanding of the information space (Figure 8 displays two examples). The graphs are analyzed in three ways: straight counts of topics, subtopics, and their links (Figure 9), metrics of their connectedness (Figure 10), and the correctness of the remembered topics and links (Figure 12). Figure 8. Participant 1’s mental model graphs from the concept mapping. DI (left) and QAI (right). Nodes represent elicited topics (red) and subtopics (yellow). Edges represent explicitly described relationships.On the left, the DI-Solana graph shows interconnected nodes arranged around central topics. There are 11 topics and 9 subtopics. On the right, the QAI-Dominion graph shows 6 topics and 16 subtopics with a more centralized structure with hub nodes connecting to many others. Nodes represent topics or subtopics, and arrows represent directed relationships between them. Violin plots showing five metrics for DI and QAI. Topics averaged 9.44 for DI and 5.63 for QAI, p = 0.00. Subtopics averaged 9.63 for DI and 8.69 for QAI, p = 0.55. Combined topics and subtopics averaged 19.06 for DI and 14.31 for QAI, p = 0.02. Connections averaged 20.69 for DI and 14.94 for QAI, p = 0.01. Connections per topic averaged 1.09 for DI and 1.03 for QAI, p = 0.27 Figure 9. Concept map metrics (RQ2—Phase 2). See caption of Figure 6 for details.Violin plots showing five metrics for DI and QAI. Topics averaged 9.44 for DI and 5.63 for QAI, p = 0.00. Subtopics averaged 9.63 for DI and 8.69 for QAI, p = 0.55. Combined topics and subtopics averaged 19.06 for DI and 14.31 for QAI, p = 0.02. Connections averaged 20.69 for DI and 14.94 for QAI, p = 0.01. Connections per topic averaged 1.09 for DI and 1.03 for QAI, p = 0.27 Violin plots showing four metrics for DI and QAI. Density averaged 0.07 for DI and 0.09 for QAI, p = 0.01. Directed diameter averaged 4.50 for DI and 3.44 for QAI, p = 0.02. Average shortest path length averaged 1.95 for DI and 1.68 for QAI, p = 0.03. Degree variance averaged 2.47 for DI and 1.70 for QAI, p = 0.01. Figure 10. Concept map graph metrics (RQ2—Phase 2). See caption of Figure 6 for details.Violin plots showing four metrics for DI and QAI. Density averaged 0.07 for DI and 0.09 for QAI, p = 0.01. Directed diameter averaged 4.50 for DI and 3.44 for QAI, p = 0.02. Average shortest path length averaged 1.95 for DI and 1.68 for QAI, p = 0.03. Degree variance averaged 2.47 for DI and 1.70 for QAI, p = 0.01. Overall, participants included 67% more topics on average in their elicited mental model graphs in the DI condition (μtopics,DI=9.44 _topics,DI=9.44) compared to the QAI condition (μtopics,QAI=5.63 _topics,QAI=5.63, t⁡(7.25)t(7.25), p=0.02p=0.02, η2=0.78η^2=0.78) and 11% more subtopics (μsubtopics,DI=9.63 _subtopics,DI=9.63, μsubtopics,QAI=8.69 _subtopics,QAI=8.69, t⁡(0.61)t(0.61), p=0.55p=0.55, η2=0.02η^2=0.02). Considering topics and subtopics together, this represents evidence that the models elicited in the DI condition were more comprehensive on average. With DI, participants made 38% more connections (μconnections,DI=20.69 _connections,DI=20.69, μconnections,QAI=14.94 _connections,QAI=14.94, t⁡(2.92)t(2.92), p=0.01p=0.01, η2=0.36η^2=0.36), but this seems to be attributable to having more nodes to connect (the connections per topic measures are fairly similar and not statistically distinguishable: μconnections/topic,DI=1.09 _connections/topic,DI=1.09, μconnections/topic,QAI=1.03 _connections/topic,QAI=1.03, t⁡(1.15)t(1.15), p=0.27p=0.27, η2=0.08η^2=0.08). We also considered each node or link and classified it as correct or incorrect with respect to the knowledge contained in the documents. In terms of raw error counts, participants made 45% fewer errors with DI than with QAI (μerrors,DI=3.31 _errors,DI=3.31, μerrors,QAI=6.00 _errors,QAI=6.00, t⁡(5)t(5), p<0.01p<0.01, η2=0.62η^2=0.62). Because the elicited models differed in size, we used error rate to show that the difference is slightly amplified (54%) (μerrorrate,DI=0.12 _errorrate,DI=0.12, μerrorrate,QAI=0.26 _errorrate,QAI=0.26, t⁡(6.10)t(6.10), p<0.01p<0.01, η2=0.71η^2=0.71). We then calculated graph metrics for the concept maps, after removing errors. The results show that the DI elicited mental model graphs were less interconnected than QAI’s, with less average density, higher diameter, higher average shortest path length and higher degree variance (all p<0.04p<0.04, see Figure 10). 4.2.2. Decision-based Scenario (Phase 3) We performed a parallel analysis of the participant answers to the scenario phase, where we asked participants to apply their knowledge of the worlds to suggest solutions to conundrums. The results follow very similar patterns as the previous subsection, which instead of describing in text we leave summarized in Figures 11 and 12. However, there are a few differences between the two analyses. First, the scenario task was much harder than the elicitation task, which left 6 participants unable to provide coherent responses in one (n=5n=5, 4 of which were with the QAI) or both (n=1n=1) of the two sessions. Second, the pattern of higher connections in DI was not strong enough for the conditions to be distinguished statistically. We did not run the graph metrics for the resulting graphs because these graphs are much sparser than in Phase 2, hence noisier and less meaningful. Violin plots showing five metrics for DI and QAI. Topics averaged 2.44 for DI and 1.38 for QAI, p = 0.02. Subtopics averaged 1.25 for DI and 0.31 for QAI, p = 0.03. Combined topics and subtopics averaged 3.69 for DI and 1.69 for QAI, p = 0.00. Connections averaged 4.19 for DI and 2.50 for QAI, p = 0.09. Connections per topic averaged 1.07 for DI and 0.96 for QAI, p = 0.72 Figure 11. Decision-based metrics (RQ2—Phase 3). See caption of Figure 6 for details.Violin plots showing five metrics for DI and QAI. Topics averaged 2.44 for DI and 1.38 for QAI, p = 0.02. Subtopics averaged 1.25 for DI and 0.31 for QAI, p = 0.03. Combined topics and subtopics averaged 3.69 for DI and 1.69 for QAI, p = 0.00. Connections averaged 4.19 for DI and 2.50 for QAI, p = 0.09. Connections per topic averaged 1.07 for DI and 0.96 for QAI, p = 0.72 Violin plots comparing DI and QAI across four error metrics. Concept map errors averaged 3.31 for DI and 6.00 for QAI (p = 0.00), while concept map error rates averaged 0.12 for DI and 0.26 for QAI (p = 0.00). Decision-based errors averaged 0.25 for DI and 0.81 for QAI (p = 0.01), and decision-based error rates averaged 0.03 for DI and 0.29 for QAI (p = 0.01). Figure 12. Concept map and Decision-based error metrics (RQ2—Phase 3). See caption of Figure 6 for details.Violin plots comparing DI and QAI across four error metrics. Concept map errors averaged 3.31 for DI and 6.00 for QAI (p = 0.00), while concept map error rates averaged 0.12 for DI and 0.26 for QAI (p = 0.00). Decision-based errors averaged 0.25 for DI and 0.81 for QAI (p = 0.01), and decision-based error rates averaged 0.03 for DI and 0.29 for QAI (p = 0.01). 4.2.3. Qualitative Results We used interview responses to analyze how participants made sense of the worlds using the two interfaces. Our qualitative analysis involved three processes: building understanding of the information space, connecting information across documents, and applying their gained knowledge to a scenario-based question. Build. Participants described clear differences in how the two interfaces supported building an understanding of the created worlds. Participants appreciated that DI did not require them to construct their own structure since it is already visible (n=5n=5). P3 explains that DI “gives you an understanding of what each chapter is and it shows you what you’re looking, what you’re headed into.” However, DI can be frustrating if readers want to build their own mental structure (n=4n=4). P10 notes, “I basically have to figure out the structure based on the description of each thing,” suggesting that they still had to determine where a given piece of information fit in their mental models. Conversely, the QAI supported participants in “being able to curate the index in my own thought” (P3) by building structure through their questions (n=7n=7). At the same time, this flexibility depended on participants having skills on what and how to ask (n=10n=10), “it was a bit more tricky because I had to have a bit more imagination.”(P6) Connect. When it comes to integrating disparate pieces of information across the information space, some participants perceived the DI as providing more information volume which, in turn, allowed them to establish more connections between the different parts of the information space (n=6n=6). For example P3 explains that, with DI, “I felt like I had more to work with and it was more that I could kind of jump off of, so it was helpful.” Yet for some, most connections with DI are implicit (i.e., not links), which forces them to make the connections mentally (n=3n=3). Even when the connection is explicit (i.e., a link), it might still be difficult to know why they are connected. P7 explains this: “if you’re going through it as link to here, link to here, you’re not necessarily getting the connections between those.”. In contrast, several participants recognized that the QAI directly supports integration of related topics, unconstrained by the structure (n=7n=7). In other words, participants appreciate that the QAI composes a coherent text using different bits of information: “AI does a better job of linking stuff together than you having to pick a piece and go to another link. It pulls it together.” (P7) However, this is not for free, since participants recognized that asking the right questions to get the right integration is not trivial (n=8n=8): “my questions were probably pretty simplistic and it sort of led you around in circles a couple of times. It’s just like any AI” (P8). Apply. The decision-based scenario (Phase 3) interrogated whether participants could use the knowledge they had gained and connected through each interface. As we mentioned before, this is likely a much harder task that goes beyond accessing and remembering and requires integration, retrieval, and even creativity; in other words, it is a task higher in Bloom’s taxonomy (Krathwohl, 2002). Some participants thought that the stable and well-organized structure of documents (DI) helped them answer the scenario (n=8n=8). P4 noted about the information to answer a question that “it gave more, helped me get a visual representation first of what was happening with the various structures.” However, some participants also recognized that, to be successfully applied, they needed to actively process the information: “It was much more of a lecture because you were just absorbing; I was filling the gaps,” (P11) (n=3n=3). With QAI the challenge is of a different nature. Although participants thought that they could ask for information that meaningfully supported their progressive understanding and ability to apply the knowledge (n=6n=6—“I was asking questions that were important to me.” (P15)), they were never sure that what they had gathered through questions and answers was sufficient (n=8n=8): “knowing what to ask and then questioning the output. Am I getting the full answer? Am I getting the correct information?”(P14). 4.2.4. RQ2 Synthesis The quantitative data show that participant’s elicited mental models contained, on average, more topics and subtopics when using the DI, although less densely connected. Importantly, the elicited graphs also had many more errors. When asked to apply what they had learn, we also observed an advantage when using the DI, which resulted in substantially more topics and subtopics brought up, and also fewer errors. The qualitative data suggests that building mental models with the DI requires effort to connect what they read with their current understanding as they go, discovering implicit connections between the topics. In contrast, the QAI offers more flexibility to build the model based on the participant’s own curiosity, and interest. The QAI answers provide custom-formulated answers that directly answer the participant’s questions and connect the topics smoothly in the answer’s text. In exchange, formulating the right questions might be challenging, and the participants were often not sure of whether they had gathered enough information or reached exhaustive cover of the available information. 4.3. What Aspects of the Different Interfaces Affect Preferences? (RQ3) At the end of Phase 4, we asked participants for comparative reflections on their experience with the interfaces. Table 3 summarizes these interview responses together with counts of their actual performance, and the counts of alignment. Overall, participants’ perceptions did not fully align with the measured patterns. This discrepancy was evident for breadth of exploration and appeared even more strongly for perceived support for learning and confidence in answering the questions. When participants chose their preferred interface, 8 participants favored each. There is no obvious mapping between perceived advantages on either exploration, mental model building or scenario answer confidence and the final stated preference. Perception Measurement Match DI QAI DI QAI Equal DI QAI Explore more 10 6 11 4 1 6 2 Learn more 6 10 11 5 0 5 4 Answer Confidence 7 9 12 1 3 5 0 Overall Preference 7 8 N/A N/A N/A N/A N/A Table 3. Comparison of subjective judgments and measured outcomes for DI and QAI. Perception shows how many participants said each interface helped them explore more (RQ1), learn more and answer with greater confidence (RQ2). Measurement shows how many participants actually performed better with each interface. Overall preference is subjective only.Comparison of participant judgments and measured outcomes for DI and QAI. Perception shows reported advantages, Measurement shows metric-based advantages, and Match shows agreement between the two. Overall preference is subjective only. Nevertheless, two main themes came up in their interviews and feedback that can help us understand participant preferences: Agency/Control and Effort/Cognitive Load. Interestingly, arguments regarding each of those topics were used to support preferences for either of the two interfaces. Agency and Control. Participants often evaluated the interfaces in terms of how independent they felt in steering the interaction and shaping control around their own interests. With DI, participants valued having the option to decide for themselves what documents to open and which links to pursue (n=9n=9) “I felt really empowered. There’s all these concepts and I’m going to go with what speaks to me,” (P1). P10 similarly describes “I could just jump into it right away. I could be independent with it.” Still, for some participants, having the freedom to choose has drawbacks (n=3n=3), “I lost control because of opening documents. I had to go back somewhere totally unrelated in my opinion.” (P4) In QAI, control meant having the ability to ask anything, which felt open-ended and self-directed (n=9n=9). P9 describes this, “I liked that I could ask the questions and gave me answers. It acknowledged me.” Similarly, P14 mentions: “I have more control over what I want to focus on,” emphasizing the value of the question-answer cycle. However, control also depended on knowing how to steer the conversation productively (n=6n=6), as P12 reflects: “I don’t think I’m asking the right questions to get the information that I wanted.” Effort and Cognitive Load. Participants often mentioned the differences in effort involved in obtaining, tracking and working with the information with each interface. Some participants appreciated DI because, as P6 explains, “there wasn’t a lot of guesswork,” and they could “jump into it right away,” (n=5n=5). P13 echoes this by describing DI as “a straightforward situation,” and P10 notes that the structure is “pretty easy to understand.” However, the effort required with DI was often placed in repeated browsing and tracking of the documents (n=4n=4). P7 particularly dislikes “I hate things that go on and on and send you off to other links. It’s too much,” with P11 also reflecting on the amount of material, “I find this very exhausting and like a lot of information.” Some participants appreciated the QAI’s ability to address several topics at once (n=9n=9), as P4 observes: “I was able to ask three questions at once,” and with P14 describing their experience as “better, quicker, and clearer.” Yet the cost of this interaction relies heavily on the reader to probe the information effectively (n=5n=5), with P8 describing: “it could be just insecurity. Trying to figure out the questions and do it in an intelligent way.” P15 also reflected that the interaction with QAI depends on the user, “I think the limitation is your mind, right?”. 5. Discussion In this section we interpret the results above in light of the three research questions stated in the introduction, then connect the results to the emerging knowledge of GenAI-based interfaces for knowledge access, including accessibility. We finish by synthesizing recommendations for practitioners and users and qualifying the results based on the limitations of our study design choices. 5.1. Answers to RQ’s and their connections Our behavioral measures of how people chose to explore the information space exposed a trade-off between the two interface styles. Participants used the more traditional DI in a way that resulted in more unique documents visited than the QAI. However, the QAI enabled more flexible transitions from different parts of the information space (RQ1). These differences are not difficult to trace to the characteristics of the interface: the QAI—backed by a current LLM—is able to synthesize answers that connect disparate pieces of information into a smooth narrative, a hitherto difficult to achieve way to navigate information that many people liked and, importantly, allows the reader to personalize their knowledge acquisition. Unfortunately, the QAI way of interacting with the information space did not seem to translate into more information remembered, better mental models, or an improved ability to apply the gained knowledge or models. In fact, the evidence that we collected supports the opposite: elicited mental models were smaller and contained more errors, although perhaps were a bit more tightly connected (RQ2). This pattern propagated to the application of their understanding, which was also poorer for QAI. We do not rely only on errors as the main or only measure of our analysis, in particular because task instructions de-emphasized recall. Nevertheless, the error measurements very clearly confirm results from the other measurements that models were overall poorer when using the QAI. Task instructions that explicitly emphasized memorization might have produced different outcomes, and this warrants further study. We anticipate two plausible causal explanations for the connection between the behavior in access (RQ1) and our appraisals of the gained information and understanding (RQ2). First, the DI requires more effort (some participants brought this up), hence more elaboration which, in turn, results in deeper processing, better memory and more functional mental models (Lockhart and Craik, 1990). Second, the QAI relinquishes the predetermined structure in the documents structure (a kind of ready-to-learn model curated by an expert in advance), forcing the reader to consider which mental structure to build and how to build it, then formulate the appropriate questions to fulfill this purpose. This might simply be too difficult and detract from the mental resources required to do the actual learning. This is, essentially, a Cognitive Load Theory argument about germane cognitive load adding to intrinsic cognitive load to overwhelm cognitive capacity (Sweller, 1988; Paas and Van Merriënboer, 1994). Interestingly, the challenges introduced by the QAI did not appear obvious to participants, who often rated it as best for the outcomes where we know they did worse (i.e., they were often wrong—see Table 3). In turn, their preferences were evenly split between the two interfaces (RQ3). This suggests that the nature of the interfaces somehow obscures their value to their users (a failure of metacognition, see the next subsection) or that preference was informed by other factors other than their ability to form more complete mental models and apply them. In any case, it is noteworthy and somewhat alarming that so many participants thought they were better with the interface that actually helped them least. 5.2. What does this mean for GenAI interfaces and non-visual information access? Our findings extend work on traditional and GenAI search. Recent studies suggest that GenAI-based tools can support exploratory search, better user experience, and lower perceived effort (Liu et al., 2021; Kaushik and Jones, 2025; Selker and Wu, 2024; Zerhoudi and Granitzer, 2025; Kim et al., 2025; Yang et al., 2025; Kim and Ji, 2026). Comparative work between chatbots and search tools afford different kinds of access, and that the strengths depend on the task, domain, and design (Pasquarelli et al., 2025; Melumad and Yun, 2025). This is consistent with our study where the QAI was particularly effective for contextualized fact-finding across multiple documents, whereas the DI better supported broad exploration of the corpus. However, most prior evaluations focus primarily on search tasks and do not measure how people build mental models. Our results therefore contribute empirical evidence suggesting that the GenAI-enabled conversational interfaces that seem poised to replace traditional search as well as much direct document access, could be detrimental for tasks that are more open, exploratory or contributing to people’s expertise and general knowledge. This is particularly important for the BLV community for two reasons. First, GenAI interfaces are being quickly adopted by the community for many tasks (Adnin and Das, 2024; Tang et al., 2025), and traditional document access is often particularly tedious and time-consuming with non-visual input-output (Lee et al., 2020). Both make it more likely that they will be used to replace document access. We also suspect that there might be important differences between BLV users and sighted users when using DI vs. QAI interfaces, but this will require further research. Our results are also consistent with concerns about the metacognitive consequences of LLM use (Tankelevitch et al., 2024). Experimental work has shown that despite how LLM-based search feels easier and faster, it can also yield superficial learning since users do less of synthesis (Melumad and Yun, 2025; Tankelevitch et al., 2024; Stadler et al., 2024; Yong et al., 2026) This is consistent with our findings. First, we did observe that participants experienced the QAI as efficient, flexible, and responsive, yet these benefits did not translate into larger concept maps, or stronger performance in applying what they had learned. Second, participants were often poor at judging which interface actually supported their exploration or learning better. Moreover, we observed that the interfaces function not only as a tool for consuming information but also as a mechanism to help users determine what they have covered, what remains unseen, and how complete or reliable their current understanding is. In our study, the DI relies on users to build an internal representation of document architecture through system interaction at the interface/surface level of the stratified model of information access (Saracevic, 1997). Users must grasp what documents exist and infer what information they contain, encouraging users to explore documents across various categories to construct this mental model. In contrast, QAI encourages users to focus on formulating their exploratory intent as prompts at the affective layer. Thus, the user’s prompt and early document choices determine the subsequent course of exploration, making coverage and completeness harder to assess. 5.3. Implications for users and future directions for practitioners Our findings have implications for the adoption of QAI technologies, for the design of future interactive systems and for future research. First, we believe that users need to be aware that, despite their own initial appraisals of the value of LLM-based QA technologies for accessing documents and other information spaces, they incur the risk of processing information more superficially and being less able to apply it. This is particularly important for BLV individuals, who often seek technological solutions to avoid being perceived as less effective than their sighted peers (Zhao et al., 2026a; Arslantas and Gul, 2022; Navas-Bonilla et al., 2025; Zhao et al., 2026b).Further confirmation should encourage wider dissemination of this issue, perhaps in educational contexts. These warnings should be qualified because readers are not always looking to form better models or remember more. Second, some of the negative effects that we observed might be ameliorated or solved by better design. For example, when people are at a loss of where to start, or how to start structuring their own learning, the system might be able to, prompted or unprompted, offer help on how to carry out the process more efficiently. We did not observe this kind of behavior (asking for help about how to explore, learn or interpret), but it is not rare in current LLMs to offer different levels of hints to the user before or after prompting. Future document interfaces should also be able to incorporate notions of desired agency and effort that adapt to user preferences. Third, designing new hybrid modes of access for document and information spaces might offer the advantages of DIs and QAIs without the disadvantages, depending on the task or the purpose of the interaction. Our study intentionally separated these modes of access to make differences easier to identify and attribute. Future systems, however, could combine document navigation, search, and conversational interaction to support different tasks and preferences. Interestingly, it might be of research interest to also consider whether documents and other current ways to structure information for human consumption can benefit from different forms and structures, especially if we accept that question-answer interfaces might become the dominant way to access them by the BLV community. 5.4. Limitations We strove to design the QAI and DI interfaces in a way representative of current interfaces and as close to the state of the art as possible. Nevertheless, many different choices to the ones we made are justifiable, and we cannot discard that they could have influenced the results in different ways. One such choice was enabling touch interaction for spatial documents. This means that we cannot separate the effects of multimodality from other aspects of the DI interface. In other words, having different types of documents might have improved people’s ability to remember and integrate the information. Similarly, we decided to include a document in each corpora that provides access to all other documents. This somewhat reduced the incentive for people to navigate between topics from other in-document links, but we also think it is more realistic. Although we still stand by our experimental design choices, we need further experiments to know whether either of these caused the improved performance in the DI condition. Specifically, future systems might include features of both DI and QAI, including both tactile spatial access, search, and presentation of literal information from the document corpus. Ecologically faithful evaluations of such system will require new experiments. Additionally, we had to decide how to count visits and transitions between visits in Phase 1 in a way that would allow comparisons between the two interfaces. We chose before analysis and with our best judgment. However, we acknowledge that other choices were possible that might result in different balances. Nevertheless, this only affects some measurements in Phase 1; other measurements in other phases are less open to variation because the model elicitation and scenario processes are identical for the two conditions. Our findings were not independently reviewed by the BLV community. We also focused on SR users to maintain a consistent access modality, as SR use imposes navigation demands central to how these interfaces are experienced, limiting generalizability to BLV users who use other access methods. Finally, we only tested participants in two sessions of ninety minutes. Prolonged use of interfaces for information access is likely to evolve, both in use patterns and in its outcomes. 6. Conclusion Our study compared two interfaces, DI and QAI, to understand how BLV users explore and build knowledge of unfamiliar domains. From this comparison, we draw attention to a meaningful distinction between interfaces. The nature of interaction with QAI through conversation can feel deceptively expansive, offering a sense of coverage even though content is synthesized. This, in turn, did not translate into better learning. In contrast, DI may demand more effort or might be overwhelming, but they result in a more comprehensive mental model due to the provided structure. This distinction is important because participants do not always recognize it. Many perceived the QAI as better even when performance suggested otherwise. For BLV users, who may be motivated to adopt these tools to reduce the burden of traditional document access, this introduces a risk of interfaces tha feel easier may also support weaker understanding. As LLM-based interfaces evolve closer to becoming intermediaries, accessible technologies should be evaluated beyond speed or convenience, but also accurate understanding and meaningful exploration. Acknowledgements. We would like to thank the Canadian Council of the Blind for their support in connecting us with our participants. We are grateful to the participants who gave their time, insights, and trust to this work. We also acknowledge the broader network of practitioners, advocates, and community members who provided us support and guidance. This research is supported by the University of Victoria, NSERC DG 2020-04401 and the Singapore Ministry of Education (MOE) Academic Research Fund (AcRF) Tier 1 grant (Project ID: 24-SIS-SMU-039). References Abdolrahmani et al. (2018) A. Abdolrahmani, R. Kuber, and S. M. Branham "Siri Talks at You": An Empirical Investigation of Voice-Activated Personal Assistant (VAPA) Usage by Individuals Who Are Blind. In Proceedings of the 20th International ACM SIGACCESS Conference on Computers and Accessibility, ASSETS ’18, New York, NY, USA, p. 249–258. External Links: ISBN 978-1-4503-5650-3, Link, Document Cited by: §2.3. Abidin et al. (2012) A. H. Z. Abidin, H. Xie, and K. W. Wong Blind users’ mental model of web page using touch screen augmented with audio feedback. In 2012 International Conference on Computer & Information Science (ICCIS), Vol. 2, p. 1046–1051. External Links: Link, Document Cited by: §2.1, §2.1. Adams et al. (2008a) A. Adams, P. Lunt, and P. Cairns A qualitative approach to hci research. In Research Methods for Human-Computer Interaction, P. Cairns and A. L. Cox (Eds.), p. 138–157 (en). External Links: ISBN 978-0-521-87012-2, Link, Document Cited by: §3.5.4. Adams et al. (2008b) A. Adams, P. Lunt, and P. Cairns A qualitative approach to HCI research. In Research Methods for Human-Computer Interaction, A. L. Cox and P. Cairns (Eds.), p. 138–157. External Links: ISBN 978-0-521-69031-7, Link, Document Cited by: §3.5.4. Adnin and Das (2024) R. Adnin and M. Das "I look at it as the king of knowledge": How Blind People Use and Understand Generative AI Tools. In Proceedings of the 26th International ACM SIGACCESS Conference on Computers and Accessibility, ASSETS ’24, New York, NY, USA, p. 1–14. External Links: ISBN 979-8-4007-0677-6, Link, Document Cited by: §1, §2.2, §2.3, §5.2. Agosti and Croft (2008) M. Agosti and W. B. Croft Information Access through Search Engines and Digital Libraries. 1 edition, The Information Retrieval Series, Vol. 22, Springer Berlin Heidelberg, Berlin, Heidelberg (eng). Note: ISSN: 1387-5264 External Links: ISBN 978-3-642-09441-5, Document Cited by: §2.2. Ai et al. (2025) Q. Ai, J. Zhan, and Y. Liu Foundations of Generative Information Retrieval. In Information Access in the Era of Generative AI, R. W. White and C. Shah (Eds.), p. 15–45 (en). External Links: ISBN 978-3-031-73147-1, Link, Document Cited by: §2.2. Arslantas and Gul (2022) T. K. Arslantas and A. Gul Digital literacy skills of university students with visual impairment: A mixed-methods analysis. Education and Information Technologies 27 (4), p. 5605–5625. External Links: ISSN 1360-2357, Link, Document Cited by: §5.3. Asakawa and Itoh (1998) C. Asakawa and T. Itoh User interface of a home page reader. In Proceedings of the Third International ACM Conference on Assistive Technologies, Assets ’98, New York, NY, USA, p. 149–156. External Links: ISBN 1581130201, Link, Document Cited by: §1, §2.2. Atcheson et al. (2025) A. Atcheson, O. Khan, B. Siemann, A. Jain, and K. Karahalios "I’d never actually realized how big an impact it had until now": perspectives of university students with disabilities on generative artificial intelligence. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI ’25, New York, NY, USA. External Links: ISBN 9798400713941, Link, Document Cited by: §1, §2.2. Athukorala et al. (2016) K. Athukorala, D. Głowacka, G. Jacucci, A. Oulasvirta, and J. Vreeken Is exploratory search different? a comparison of information search behavior for exploratory and lookup tasks. J. Assoc. Inf. Sci. Technol. 67 (11), p. 2635–2651. External Links: ISSN 2330-1635, Link, Document Cited by: §2.2. Baez et al. (2022) M. Baez, C. M. Cutrupi, M. Matera, I. Possaghi, E. Pucci, G. Spadone, C. Cappiello, and A. Pasquale Exploring challenges for Conversational Web Browsing with Blind and Visually Impaired Users. In Extended Abstracts of the 2022 CHI Conference on Human Factors in Computing Systems, CHI EA ’22, New York, NY, USA, p. 1–7. External Links: ISBN 978-1-4503-9156-6, Link, Document Cited by: §2.2, §2.3. [13] R. Baeza-Yates and B. Ribeiro-Neto Modern Information Retrieval. Pearson, Harlow (English). External Links: ISBN 978-0-201-39829-8 Cited by: §2.2. Bainbridge (1983) L. Bainbridge IRONIES OF AUTOMATION. In Analysis, Design and Evaluation of Man–Machine Systems, G. Johannsen and J. E. Rijnsdorp (Eds.), p. 129–135. External Links: ISBN 978-0-08-029348-6, Link, Document Cited by: §2.2. Barros et al. (2023) G. Barros, W. Correia, and J. M. Teixeira Towards the Effectiveness of 3D Printing on Tactile Content Creation for Visually Impaired Users. Polymers 15 (9), p. 2180 (en). Note: Publisher: Multidisciplinary Digital Publishing Institute External Links: ISSN 2073-4360, Link, Document Cited by: §3.1, §3.2.1. Bias et al. (2015) R. G. Bias, B. M. Moon, and R. R. Hoffman Concept Mapping Usability Evaluation: An Exploratory Study of a New Usability Inspection Method. International Journal of Human–Computer Interaction 31 (9), p. 571–583. External Links: ISSN 1044-7318, Link, Document Cited by: §2.1, §3.4. Bigham et al. (2017) J. P. Bigham, I. Lin, and S. Savage The effects of "not knowing what you don’t know" on web accessibility for blind web users. In Proceedings of the 19th International ACM SIGACCESS Conference on Computers and Accessibility, ASSETS ’17, New York, NY, USA, p. 101–109. External Links: ISBN 9781450349260, Link, Document Cited by: §1. Braille Authority of North America (2022) Braille Authority of North America Guidelines and Standards for Tactile Graphics. External Links: Link Cited by: §3.1. Braun and Clarke (2021) V. Braun and V. Clarke Thematic analysis: a practical guide. Cited by: §3.5.4. Braun and Clarke (2023) V. Braun and V. Clarke Thematic analysis. In APA handbook of research methods in psychology: Research designs: Quantitative, qualitative, neuropsychological, and biological, Vol. 2, 2nd ed, APA Handbooks in Psychology®, p. 65–81. External Links: ISBN 978-1-4338-4133-0 978-1-4338-4134-7, Document Cited by: §3.5.4. Broder et al. (2000) A. Broder, R. Kumar, F. Maghoul, P. Raghavan, S. Rajagopalan, R. Stata, A. Tomkins, and J. Wiener Graph structure in the Web. Computer Networks 33 (1), p. 309–320. External Links: ISSN 1389-1286, Link, Document Cited by: §3.5.1. Bucciarelli and Cutica (2012) M. Bucciarelli and I. Cutica Mental Models in Improving Learning. In Encyclopedia of the Sciences of Learning, p. 2213–2215 (en). Note: 2025-08-29 n other words, when confronted with new learning tasks, learners have to construct a mental model integrating their preexisting knowledge and the new information from the learning environment, along with proper inferences that can be drawn from these. This model has to be reconstructed several times to become a schema, i.e., to be learnt. External Links: ISBN 978-1-4419-1428-6, Link, Document Cited by: §2.1. Butler et al. (2021) M. Butler, L. M. Holloway, S. Reinders, C. Goncu, and K. Marriott Technology Developments in Touch-Based Accessible Graphics: A Systematic Review of Research 2010-2020. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, CHI ’21, New York, NY, USA, p. 1–15. External Links: ISBN 978-1-4503-8096-6, Link, Document Cited by: §2.3, §3.1. Byström and Järvelin (1995) K. Byström and K. Järvelin Task complexity affects information seeking and use. Information Processing & Management 31 (2), p. 191–213. External Links: ISSN 0306-4573, Link, Document Cited by: §3.4. Cepeda et al. (2006) N. J. Cepeda, H. Pashler, E. Vul, J. T. Wixted, and D. Rohrer Distributed practice in verbal recall tasks: A review and quantitative synthesis. Psychological Bulletin 132 (3), p. 354–380 (eng). External Links: ISSN 0033-2909, Document Cited by: §3.4. Chan et al. (2025) B. J. Chan, C. Chen, J. Cheng, and H. Huang Don’t Do RAG: When Cache-Augmented Generation is All You Need for Knowledge Tasks. In Companion Proceedings of the ACM on Web Conference 2025, W ’25, New York, NY, USA, p. 893–897. External Links: ISBN 979-8-4007-1331-6, Link, Document Cited by: §3.2.2. Chheda-Kothary et al. (2025) A. Chheda-Kothary, A. Sharif, D. A. Rios, and B. A. Smith "It Brought Me Joy": Opportunities for Spatial Browsing in Desktop Screen Readers. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI ’25, New York, NY, USA, p. 1–18. External Links: ISBN 979-8-4007-1394-1, Link, Document Cited by: §2.3. Costanza-Chock (2020) S. Costanza-Chock Design justice: community-led practices to build the worlds we need. The MIT Press. External Links: ISBN 9780262356862, Document Cited by: §3.6. Croft et al. (2010) B. Croft, D. Metzler, and T. Strohman Search Engines: Information Retrieval in Practice. Pearson, Boston (English). External Links: ISBN 978-0-13-607224-9 Cited by: §2.2. Dang et al. (2023) H. Dang, S. Goller, F. Lehmann, and D. Buschek Choice Over Control: How Users Write with Large Language Models using Diegetic and Non-Diegetic Prompting. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, CHI ’23, New York, NY, USA, p. 1–17. External Links: ISBN 978-1-4503-9421-5, Link, Document Cited by: §2.2. Dhingra et al. (2017) B. Dhingra, L. Li, X. Li, J. Gao, Y. Chen, F. Ahmed, and L. Deng Towards End-to-End Reinforcement Learning of Dialogue Agents for Information Access. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), R. Barzilay and M. Kan (Eds.), Vancouver, Canada, p. 484–495. External Links: Link, Document Cited by: §2.2. Dix (2020) A. Dix Statistics for hci: making sense of quantitative data. Morgan & Claypool Publishers (en). External Links: ISBN 978-1-68173-744-7 Cited by: §3.5.4. Doore et al. (2024) S. A. Doore, D. Istrati, C. Xu, Y. Qiu, A. Sarrazin, and N. A. Giudice Images, Words, and Imagination: Accessible Descriptions to Support Blind and Low Vision Art Exploration and Engagement. Journal of Imaging 10 (1), p. 26 (en). External Links: ISSN 2313-433X, Link, Document Cited by: §1. Elmqvist (2023) N. Elmqvist Visualization for the Blind | IX Magazine Issue X.1 January - February 2023. Interactions 30 (1) (en). External Links: Link Cited by: §3.1. Ernst & Young (2024) Ernst & Young GenAI for Accessibility: More Human, Not Less – How Does Microsoft 365 Copilot Impact the Working Experience of People Living with Disability and/or Neurodivergence?. Research Report Ernst & Young Global Limited. External Links: Link Cited by: §1. Erp (2002) J. V. Erp Guidelines for the use of vibro-tactile displays in human computer interaction. External Links: Link Cited by: §2.3. Fast et al. (2018) E. Fast, B. Chen, J. Mendelsohn, J. Bassen, and M. S. Bernstein Iris: A Conversational Agent for Complex Tasks. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, CHI ’18, New York, NY, USA, p. 1–12. External Links: ISBN 978-1-4503-5620-6, Link, Document Cited by: §2.2. Fischer and Mehnert (2021) N. Fischer and W. Mehnert Building Possible Worlds: A Speculation Based Framework to Reflect on Images of the Future. External Links: Link Cited by: §3.1. Furini et al. (2020) M. Furini, S. Mirri, M. Montangero, and C. Prandi Do Conversational Interfaces Kill Web Accessibility?. 2020 IEEE 17th Annual Consumer Communications & Networking Conference (CCNC), p. 1–6. External Links: Link, Document Cited by: §2.2, §2.3. Gentner and Stevens (2014) D. Gentner and A. L. Stevens Mental Models. Psychology Press (en). External Links: ISBN 978-1-317-76940-8 Cited by: §2.1. Gerken et al. (2011) J. Gerken, H. Jetter, M. Zöllner, M. Mader, and H. Reiterer The concept maps method as a tool to evaluate the usability of APIs. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, CHI ’11, New York, NY, USA, p. 3373–3382. External Links: ISBN 978-1-4503-0228-9, Link, Document Cited by: §2.1, §3.4. Gonzalez Penuela et al. (2025) R. E. Gonzalez Penuela, R. Hu, S. Lin, T. Shende, and S. Azenkot Towards Understanding the Use of MLLM-Enabled Applications for Visual Interpretation by Blind and Low Vision People. In Proceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems, ACM Conferences, p. 1–8. External Links: ISBN 979-8-4007-1395-8, Link, Document Cited by: §1. Gubbi Mohanbabu and Pavel (2024) A. Gubbi Mohanbabu and A. Pavel Context-Aware Image Descriptions for Web Accessibility. In Proceedings of the 26th International ACM SIGACCESS Conference on Computers and Accessibility, ASSETS ’24, New York, NY, USA, p. 1–17. External Links: ISBN 979-8-4007-0677-6, Link, Document Cited by: §1. Gupta et al. (2017) R. Gupta, M. Balakrishnan, and P.V.M. Rao Tactile Diagrams for the Visually Impaired. IEEE Potentials 36 (1), p. 14–18. External Links: ISSN 0278-6648, Link, Document Cited by: §2.3. Gupta et al. (2019) R. Gupta, P. V. M. Rao, M. Balakrishnan, and S. Mannheimer Evaluating the Use of Variable Height in Tactile Graphics. In 2019 IEEE World Haptics Conference (WHC), p. 121–126. External Links: Link, Document Cited by: §3.1. Hall (2018) E. Hall Conversational Design. Mule Design, S.l. (English). External Links: ISBN 978-1-952616-30-3 Cited by: §2.2. Haque et al. (2023) S. Haque, H. Mahmoudi, N. Ghaffarzadegan, and K. Triantis Mental models, cognitive maps, and the challenge of quantitative analysis of their network representations. System Dynamics Review 39 (2), p. 152–170 (en). External Links: ISSN 1099-1727, Link, Document Cited by: §3.5.2. Harper and Dorton (2019) S. Harper and S. Dorton A Context-Driven Framework for Selecting Mental Model Elicitation Methods. Proceedings of the Human Factors and Ergonomics Society Annual Meeting 63 (1), p. 367–371 (EN). External Links: ISSN 1071-1813, Link, Document Cited by: §2.1. He et al. (2025) T. He, M. McCracken, D. Hajas, S. Creem-Regehr, and A. Lex Using Tactile Charts to Support Comprehension and Learning of Complex Visualizations for Blind and Low-Vision Individuals. arXiv. Note: arXiv:2507.21462 [cs] External Links: Link, Document Cited by: §3.1. Holtz et al. (2024) N. Holtz, S. Wittfoth, and J. M. Gómez The New Era of Knowledge Retrieval: Multi-Agent Systems Meet Generative AI. In 2024 Portland International Conference on Management of Engineering and Technology (PICMET), p. 1–10. Note: ISSN: 2159-5100 External Links: Link, Document Cited by: §2.2. Hornbæk (2013) K. Hornbæk Some Whys and Hows of Experiments in Human–Computer Interaction. Found. Trends Hum.-Comput. Interact. 5 (4), p. 299–373. External Links: ISSN 1551-3955, Link, Document Cited by: §3.4. Jones et al. (2011) N. Jones, H. Ross, T. Lynam, P. Perez, and A. Leitch Mental models: an interdisciplinary synthesis of theory and methods. SMART Infrastructure Facility - Papers. Note: Read 2020-09-30 for HAC project Good summary of the theory of mental models, (up to the date) of methods to elicit these models, and of their application (centered on Natural Resource Management). This is a key paper. It uncovers a lot of the ideas that I came up with originally. External Links: Link Cited by: §2.1. Jordan et al. (2024) J. B. Jordan, V. Van Hyning, M. A. Jones, R. Bradley Montgomery, E. Bottner, and E. Tansil Information wayfinding of screen reader users: five personas to expand conceptualizations of user experiences. In Proceedings of the 26th International ACM SIGACCESS Conference on Computers and Accessibility, ASSETS ’24, New York, NY, USA. External Links: ISBN 9798400706776, Link, Document Cited by: §1. Jung et al. (2022) C. Jung, S. Mehta, A. Kulkarni, Y. Zhao, and Y. Kim Communicating visualizations without visuals: investigation of visualization alternative text for people with visual impairments. IEEE Transactions on Visualization and Computer Graphics 28 (1), p. 1095–1105. External Links: ISSN 1941-0506, Document Cited by: §2.3. Kaushik and Jones (2025) A. Kaushik and G. J. F. Jones Comparing Conventional and Conversational Search Interaction Using Implicit Evaluation Methods. p. 292–304. External Links: ISBN 978-989-758-634-7, Link Cited by: §5.2. Kelly (2009) D. Kelly Methods for Evaluating Interactive Information Retrieval Systems with Users. Foundations and Trends® in Information Retrieval 3 (1–2), p. 1–224 (English). Note: Publisher: Now Publishers, Inc. External Links: ISSN 1554-0669, 1554-0677, Link, Document Cited by: §2.2. Keppel (1991) G. Keppel Design and analysis: A researcher’s handbook, 3rd ed. Design and analysis: A researcher’s handbook, 3rd ed, Prentice-Hall, Inc, Englewood Cliffs, NJ, US. External Links: ISBN 978-0-13-200775-7 Cited by: §3.4. Kieras and Bovair (1984) D. E. Kieras and S. Bovair The role of a mental model in learning to operate a device. Cognitive Science 8 (3), p. 255–273. External Links: ISSN 0364-0213, Link, Document Cited by: §3.1. Kim et al. (2025) H. Kim, S. L. K. Pond, and S. MacNeil Conversations over Clicks: Impact of Chatbots on Information Search in Interdisciplinary Learning. arXiv. Note: arXiv:2507.21490 [cs] External Links: Link, Document Cited by: §2.2, §5.2. Kim and Ji (2026) N. Kim and Y. G. Ji Exploratory search with generative AI: An empirical study on the impact of interaction design strategies on information exploration and cognitive load. International Journal of Human-Computer Studies 210, p. 103771. External Links: ISSN 1071-5819, Link, Document Cited by: §5.2. Kodandaram et al. (2024) S. R. Kodandaram, U. Uckun, X. Bi, I. Ramakrishnan, and V. Ashok Enabling Uniform Computer Interaction Experience for Blind Users through Large Language Models. In The 26th International ACM SIGACCESS Conference on Computers and Accessibility, St. John’s NL Canada, p. 1–14 (en). External Links: ISBN 979-8-4007-0677-6, Link, Document Cited by: §2.2, §2.3. Krathwohl (2002) D. R. Krathwohl A Revision of Bloom’s Taxonomy: An Overview. Theory Into Practice 41 (4), p. 212–218. Note: _eprint: https://doi.org/10.1207/s15430421tip4104_2 External Links: ISSN 0040-5841, Link, Document Cited by: §4.2.3. Kurniawan and Sutcliffe (2002) S. H. Kurniawan and A. Sutcliffe Mental Models of Blind Users in the Windows Environment. In Computers Helping People with Special Needs, K. Miesenberger, J. Klaus, and W. Zagler (Eds.), Berlin, Heidelberg, p. 568–574 (en). External Links: ISBN 978-3-540-45491-5, Document Cited by: §2.1. Lee et al. (2020) H. Lee, V. Ashok, and I.V. Ramakrishnan Repurposing Visual Input Modalities for Blind Users: A Case Study of Word Processors. Conference proceedings. IEEE International Conference on Systems, Man, and Cybernetics 2020, p. 2714–2721. External Links: ISSN 1062-922X, Link, Document Cited by: §5.2. Lehmann and Buschek (2024) F. Lehmann and D. Buschek Functional Flexibility in Generative AI Interfaces: Text Editing with LLMs through Conversations, Toolbars, and Prompts. arXiv. Note: arXiv:2410.10644 [cs] External Links: Link, Document Cited by: §2.2. Lewis and Jones (1996) D. D. Lewis and K. S. Jones Natural language processing for information retrieval. Commun. ACM 39 (1), p. 92–101. External Links: ISSN 0001-0782, Link, Document Cited by: §2.2. Liu et al. (2021) B. Liu, Y. Wu, Y. Liu, F. Zhang, Y. Shao, C. Li, M. Zhang, and S. Ma Conversational vs Traditional: Comparing Search Behavior and Outcome in Legal Case Retrieval. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’21, New York, NY, USA, p. 1622–1626. External Links: ISBN 978-1-4503-8037-9, Link, Document Cited by: §5.2. Lockhart and Craik (1990) R. Lockhart and F. Craik Levels of Processing: A Retrospective Commentary on a Framework for Memory Research. Canadian Journal of Psychology 44 (1), p. 87–112 (ENGLISH). External Links: ISSN 0008-4255, Link Cited by: §5.1. Lunn et al. (2011) D. Lunn, S. Harper, and S. Bechhofer Identifying behavioral strategies of visually impaired users to improve access to web content. ACM Trans. Access. Comput. 3 (4). External Links: ISSN 1936-7228, Link, Document Cited by: §1, §2.2. M. et al. (2018) H. M., K. I.M., and M. E. Interview talk and the co-construction of concept maps. Educational Research 60 (4), p. 373–389. Note: _eprint: https://doi.org/10.1080/00131881.2018.1522963 External Links: ISSN 0013-1881, Link, Document Cited by: §3.4. M.P.Geetha et al. (2024) M.P.Geetha, G. Thirukumaran, C.Pradakshana, B.Sudharsana, and T. Ashwin Conversational AI Meets Documents Revolutionizing PDF Interaction with GenAI. In 2024 International Conference on Emerging Research in Computational Science (ICERCS), p. 1–6. External Links: Link, Document Cited by: §2.2. Marchionini and Shneiderman (1988) G. Marchionini and B. Shneiderman Finding Facts vs. Browsing Knowledge in Hypertext Systems. Computer 21 (1), p. 70–80. External Links: ISSN 0018-9162, Link, Document Cited by: §4.1.2. Marchionini (2006) G. Marchionini Exploratory search: from finding to understanding. Commun. ACM 49 (4), p. 41–46. External Links: ISSN 0001-0782, Link, Document Cited by: §2.2, §3.4, §4.1.2. Melumad and Yun (2025) S. Melumad and J. H. Yun Experimental evidence of the effects of large language models versus web search on depth of learning. PNAS Nexus 4 (10), p. pgaf316. External Links: ISSN 2752-6542, Link, Document Cited by: §5.2, §5.2. Mo et al. (2025) F. Mo, K. Mao, Z. Zhao, H. Qian, H. Chen, Y. Cheng, X. Li, Y. Zhu, Z. Dou, and J. Nie A Survey of Conversational Search. ACM Trans. Inf. Syst.. Note: Just Accepted External Links: ISSN 1046-8188, Link, Document Cited by: §2.2. Moore (2018) R. J. Moore A Natural Conversation Framework for Conversational UX Design. In Studies in Conversational UX Design, R. J. Moore, M. H. Szymanski, R. Arar, and G. Ren (Eds.), p. 181–204 (en). External Links: ISBN 978-3-319-95579-7, Link, Document Cited by: §2.2. Mukhiddinov and Kim (2021) M. Mukhiddinov and S. Kim A systematic literature review on the automatic creation of tactile graphics for the blind and visually impaired. Processes 9, p. 1726. External Links: Document Cited by: §2.3. Nadkarni (2003) S. Nadkarni Instructional Methods and Mental Models of Students: An Empirical Investigation. Academy of Management Learning & Education 2 (4), p. 335–351. External Links: ISSN 1537-260X, Link, Document Cited by: §2.1. Navas-Bonilla et al. (2025) C. d. R. Navas-Bonilla, J. A. Guerra-Arango, D. A. Oviedo-Guado, and D. E. Murillo-Noriega Inclusive education through technology: a systematic review of types, tools and characteristics. Frontiers in Education 10 (English). External Links: ISSN 2504-284X, Link, Document Cited by: §5.3. Nicmanis (2024) M. Nicmanis Reflexive content analysis: an approach to qualitative data analysis, reduction, and description. International Journal of Qualitative Methods 23, p. 16094069241236603. Cited by: §4.1.2. Noble and Smith (2025) H. Noble and J. Smith Ensuring validity and reliability in qualitative research. Evidence-Based Nursing 28 (4), p. 206–208. Cited by: §4.1.2. Norman (1983) D. A. Norman Some observations on mental models. In Mental models, p. 7–14. Cited by: §2.1. Nousiainen and Koponen (2010) M. Nousiainen and I. T. Koponen Concept maps representing knowledge of physics: Connecting structure and content in the context of electricity and magnetism. Nordic Studies in Science Education 6 (2), p. 155–172 (en). External Links: ISSN 1894-1257, Link, Document Cited by: §3.5.1. Oumard et al. (2022) C. Oumard, J. Kreimeier, and T. Götzelmann Pardon? An Overview of the Current State and Requirements of Voice User Interfaces for Blind and Visually Impaired Users. In Computers Helping People with Special Needs: 18th International Conference, ICCHP-AAATE 2022, Lecco, Italy, July 11–15, 2022, Proceedings, Part I, Berlin, Heidelberg, p. 388–398. External Links: ISBN 978-3-031-08647-2, Link, Document Cited by: §2.3. Paas and Van Merriënboer (1994) F. G. W. C. Paas and J. J. G. Van Merriënboer Instructional control of cognitive load in the training of complex cognitive tasks. Educational Psychology Review 6 (4), p. 351–371 (en). External Links: ISSN 1573-336X, Link, Document Cited by: §5.1. Paolo and Warglien (1999) B. Paolo and M. Warglien Mental models and local semantics: the problem of information integration. External Links: Link Cited by: §3.4. Pasquarelli et al. (2025) L. Pasquarelli, C. Koutcheme, and A. Hellas AI Chatbots vs. Traditional Search: A Comparative Study on Student Information Retrieval. In 2025 IEEE Frontiers in Education Conference (FIE), p. 1–8. Note: ISSN: 2377-634X External Links: ISSN 2377-634X, Link, Document Cited by: §5.2. Passi and Vorvoreanu (2022) S. Passi and M. Vorvoreanu Overreliance on AI literature review. Microsoft Research 339, p. 340. External Links: Link Cited by: §2.2. Paul (2014) C. L. Paul Analyzing card-sorting data using graph visualization. J. Usability Studies 9 (3), p. 87–104. Cited by: §3.5.2. Perera et al. (2025) M. Perera, S. Ananthanarayan, C. Goncu, and K. Marriott The Sky is the Limit: Understanding How Generative AI can Enhance Screen Reader Users’ Experience with Productivity Applications. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI ’25, New York, NY, USA, p. 1–17. External Links: ISBN 979-8-4007-1394-1, Link, Document Cited by: §1, §2.3. Pradhan et al. (2018) A. Pradhan, K. Mehta, and L. Findlater "Accessibility Came by Accident": Use of Voice-Controlled Intelligent Personal Assistants by People with Disabilities. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, CHI ’18, New York, NY, USA, p. 1–13. External Links: ISBN 978-1-4503-5620-6, Link, Document Cited by: §2.3. Preininger et al. (2021) A. M. Preininger, B. L. Rosario, A. M. Buchold, J. Heiland, N. Kutub, B. S. Bohanan, B. South, and G. P. Jackson Differences in information accessed in a pharmacologic knowledge base using a conversational agent vs traditional search methods. International Journal of Medical Informatics 153, p. 104530. External Links: ISSN 1386-5056, Link, Document Cited by: §2.2, §2.2. Pucci et al. (2023) E. Pucci, I. Possaghi, C. M. Cutrupi, M. Baez, C. Cappiello, and M. Matera Defining Patterns for a Conversational Web. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, CHI ’23, New York, NY, USA, p. 1–17. External Links: ISBN 978-1-4503-9421-5, Link, Document Cited by: §2.2, §2.3. Reindal (2008) S. M. Reindal A social relational model of disability: a theoretical framework for special needs education?. European Journal of Special Needs Education 23 (2), p. 135–146. External Links: ISSN 0885-6257, Document Cited by: §3.6. Saei et al. (2010) S. N. S. M. Saei, S. Sulaiman, and H. Hasbullah Mental model of blind users to assist designers in system development. In 2010 International Symposium on Information Technology, Vol. 1, p. 1–5. Note: ISSN: 2155-899X External Links: ISSN 2155-899X, Link, Document Cited by: §2.1. Sanchez and Flores (2010) J. Sanchez and H. Flores Concept Mapping for Virtual Rehabilitation and Training of the Blind. IEEE Transactions on Neural Systems and Rehabilitation Engineering 18 (2), p. 210–219. External Links: ISSN 1558-0210, Link, Document Cited by: §2.1. Saracevic (1997) T. Saracevic The stratified model of information retrieval interaction: extension and applications. In Proceedings of the annual meeting-american society for information science, Vol. 34, p. 313–327. Cited by: §2.2, §2.2, §5.2. Sarrafzadeh et al. (2014) B. Sarrafzadeh, O. Vechtomova, and V. Jokic Exploring knowledge graphs for exploratory search. In Proceedings of the 5th Information Interaction in Context Symposium, IIiX ’14, New York, NY, USA, p. 135–144. External Links: ISBN 978-1-4503-2976-7, Link, Document Cited by: §3.4. Saucier and Dobmeier (2025) C. J. Saucier and C. M. Dobmeier A mental models approach to communication: integrating the features, functions, and mechanisms of mental modeling. Communication Theory, p. qtaf012. External Links: ISSN 1468-2885, Link, Document Cited by: §2.1. Selker and Wu (2024) T. Selker and Y. Wu Generative AI’s aggregated knowledge versus web-based curated knowledge. arXiv. Note: arXiv:2410.12091 [cs] External Links: Link, Document Cited by: §5.2. Sharma et al. (2025) T. Sharma, Y. Tseng, L. Zhang, A. Ide, K. A. Mack, L. Findlater, D. Gurari, and Y. Wang “Before, I Asked My Mom, Now I Ask ChatGPT”: Visual Privacy Management with Generative AI for Blind and Low-Vision People. In Proceedings of the 27th International ACM SIGACCESS Conference on Computers and Accessibility, ASSETS ’25, New York, NY, USA, p. 1–14. External Links: ISBN 979-8-4007-0676-9, Link, Document Cited by: §1. Shneiderman (1996) B. Shneiderman The Eyes Have It: A Task by Data Type Taxonomy for Information Visualizations. In Proceedings of the 1996 IEEE Symposium on Visual Languages, VL ’96, USA, p. 336. External Links: ISBN 978-0-8186-7508-9 Cited by: §4.1.2. Simon (1978) H. A. Simon On the forms of mental representation. Note: Read 2024-04-04 for representational work Eminent Psychologist, AI luminary and accidental computer science. Presents a theory of representation in the human mind, that is mostly symbolic. The paper posits that most of the “underlying hardware” in the mind is probably associative symbol stores, in particular through “colored” (labeled) linked lists. It also posits that, because of an abstraction, we can layer or separate the hardware mechanisms from the overlaid representations (which can be multiple). It proposes two important concepts for representations. Representations can be informationally equivalent or computationally equivalent. Informational equivalence means that you can recover one from the other. Computational equivalence means that, within a certain multiplicative constant, you can extract inferences in the same time (it’s a bit like big-O notation). It speculates that most underlying “hardware” is probably symbolic and sequential in nature, but it has some trouble (and concedes) the possibility of a graphical representation. Some interesting sentence: It is impossible to find an entirely neutral language in which to describe representations of information, for a language is itself a form of representation. It describes a representation as teh “data types” + the “primitives” (fast) operations that can be performed on information in those data types. External Links: Link Cited by: §2.1. Soufan (2023) A. Soufan Towards Understanding and Supporting Exploratory Searches. In Proceedings of the 2023 Conference on Human Information Interaction and Retrieval, CHIIR ’23, New York, NY, USA, p. 490–494. External Links: ISBN 979-8-4007-0035-4, Link, Document Cited by: §3.4. Sowa and Przegalinska (2025) K. Sowa and A. Przegalinska From Expert Systems to Generative Artificial Experts: A New Concept for Human-AI Collaboration in Knowledge Work. Journal of Artificial Intelligence Research 82, p. 2101–2124 (en). External Links: ISSN 1076-9757, Link, Document Cited by: §2.2. Stadler et al. (2024) M. Stadler, M. Bannert, and M. Sailer Cognitive ease at a cost: LLMs reduce mental effort but compromise depth in student scientific inquiry. Computers in Human Behavior 160, p. 108386. External Links: ISSN 0747-5632, Link, Document Cited by: §5.2. Subramonyam et al. (2024) H. Subramonyam, R. Pea, C. Pondoc, M. Agrawala, and C. Seifert Bridging the Gulf of Envisioning: Cognitive Challenges in Prompt Based Interactions with LLMs. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, CHI ’24, New York, NY, USA, p. 1–19. External Links: ISBN 979-8-4007-0330-0, Link, Document Cited by: §2.2. Suri et al. (2024) S. Suri, S. Counts, L. Wang, C. Chen, M. Wan, T. Safavi, J. Neville, C. Shah, R. W. White, R. Andersen, G. Buscher, S. Manivannan, N. Rangan, and L. Yang The Use of Generative Search Engines for Knowledge Work and Complex Tasks. CoRR (en). External Links: Link Cited by: §2.2. Sweller (1988) J. Sweller Cognitive Load During Problem Solving: Effects on Learning. Cognitive Science 12 (2), p. 257–285 (en). External Links: ISSN 1551-6709, Link, Document Cited by: §5.1. Tang et al. (2025) X. Tang, A. Abdolrahmani, D. Gergle, and A. M. Piper Everyday Uncertainty: How Blind People Use GenAI Tools for Information Access. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI ’25, New York, NY, USA, p. 1–17. External Links: ISBN 979-8-4007-1394-1, Link, Document Cited by: §1, §5.2. Tankelevitch et al. (2024) L. Tankelevitch, V. Kewenig, A. Simkute, A. E. Scott, A. Sarkar, A. Sellen, and S. Rintel The Metacognitive Demands and Opportunities of Generative AI. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, CHI ’24, New York, NY, USA, p. 1–24. External Links: ISBN 9798400703300, Link, Document Cited by: §2.2, §5.2. Thomas (2004) C. Thomas Rescuing a social relational understanding of disability. Scandinavian Journal of Disability Research 6 (1), p. 22–36. External Links: Document, Link, https://doi.org/10.1080/15017410409512637 Cited by: §3.6. Thompson et al. (2023) J. R. Thompson, J. J. Martinez, A. Sarikaya, E. Cutrell, and B. Lee Chart reader: accessible visualization experiences designed with screen reader users. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, CHI ’23, New York, NY, USA. External Links: ISBN 9781450394215, Link, Document Cited by: §2.3. Torres et al. (2019) C. Torres, W. Franklin, and L. Martins Accessibility in Chatbots: The State of the Art in Favor of Users with Visual Impairment. In Advances in Usability, User Experience and Assistive Technology, p. 623–635 (en). External Links: ISBN 978-3-319-94947-5, Link, Document Cited by: §2.2. Wabiński et al. (2022) J. Wabiński, A. Mościcka, and G. Touya Guidelines for Standardizing the Design of Tactile Maps: A Review of Research and Best Practice. The Cartographic Journal 59 (3), p. 239–258. External Links: ISSN 0008-7041, Link, Document Cited by: §2.3. WCAG (2025) WCAG Web Content Accessibility Guidelines (WCAG) 2.1. External Links: Link Cited by: §3.2.1, §3.2.2. Weth and Hauswirth (2013) C. v. d. Weth and M. Hauswirth DOBBS: Towards a Comprehensive Dataset to Study the Browsing Behavior of Online Users. In Proceedings of the 2013 IEEE/WIC/ACM International Joint Conferences on Web Intelligence (WI) and Intelligent Agent Technologies (IAT) - Volume 01, WI-IAT ’13, Vol. 01, USA, p. 51–56. External Links: ISBN 978-0-7695-5145-6, Link, Document Cited by: §3.5.1. Yang et al. (2025) Y. Yang, K. Urgo, J. Arguello, and R. Capra Search+Chat: Integrating Search and GenAI to Support Users with Learning-oriented Search Tasks. In Proceedings of the 2025 ACM SIGIR Conference on Human Information Interaction and Retrieval, CHIIR ’25, New York, NY, USA, p. 57–70. External Links: ISBN 9798400712906, Link, Document Cited by: §2.2, §5.2. Yong et al. (2026) Q. Yong, A. Estey, and M. Nacenta Characterization and effects of cs2 learning with genai, visualization, and human support. arXiv preprint arXiv:2606.02933. Cited by: §5.2. Yoon and Jetter (2016) B. S. Yoon and A. J. Jetter Comparative analysis for Fuzzy Cognitive Mapping. In 2016 Portland International Conference on Management of Engineering and Technology (PICMET), p. 1897–1908. External Links: Link, Document Cited by: §3.5.1. Zerhoudi and Granitzer (2025) S. Zerhoudi and M. Granitzer SearchLab: Exploring Conversational and Traditional Search Interfaces in Information Retrieval. In Proceedings of the 2025 ACM SIGIR Conference on Human Information Interaction and Retrieval, CHIIR ’25, New York, NY, USA, p. 382–389. External Links: ISBN 979-8-4007-1290-6, Link, Document Cited by: §5.2. Zhao et al. (2026a) Y. Zhao, M. A. Nacenta, M. A. Sukhai, and S. Somanath Accessibility-Driven Information Transformations in Mixed-Visual Ability Work Teams. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, CHI ’26, New York, NY, USA, p. 1–14. External Links: ISBN 979-8-4007-2278-3, Link, Document Cited by: §5.3. Zhao et al. (2026b) Y. Zhao, M. A. Nacenta, M. A. Sukhai, and S. Somanath " If we had the information that we need to interpret the world around us, we wouldn’t be disabled:" barriers and opportunities in information work among blind and sighted colleagues. In Proceedings of the 5th Annual Symposium on Human-Computer Interaction for Work, p. 1–15. Cited by: §5.3. Zhao et al. (2024) Y. Zhao, M. A. Nacenta, M. A. Sukhai, and S. Somanath TADA: making node-link diagrams accessible to blind and low-vision people. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, CHI ’24, New York, NY, USA. External Links: ISBN 9798400703300, Link, Document Cited by: §2.3, §3.1. Zong et al. (2022a) J. Zong, C. Lee, A. Lundgard, J. Jang, D. Hajas, and A. Satyanarayan Rich screen reader experiences for accessible data visualization. Computer Graphics Forum 41 (3), p. 15–27 (en). External Links: ISSN 1467-8659, Document Cited by: §2.3. Zong et al. (2022b) J. Zong, C. Lee, A. Lundgard, J. Jang, D. Hajas, and A. Satyanarayan Rich Screen Reader Experiences for Accessible Data Visualization. Computer Graphics Forum 41 (3), p. 15–27 (en). External Links: ISSN 1467-8659, Link, Document Cited by: §2.3.

Instant research discovery