pith. sign in

arxiv: 2604.00187 · v2 · submitted 2026-03-31 · 💻 cs.HC · cs.AI· cs.ET

Explainable AI for Blind and Low-Vision Users: Navigating Trust, Modality, and Interpretability in the Agentic Era

Abu Noman Md Sakib , Protik Dey , Zijie Zhang , Taslima Akter This is my paper

Pith reviewed 2026-05-08 02:24 UTC · model gemini-3-flash-preview

classification 💻 cs.HC cs.AIcs.ET
keywords Explainable AIBlind and Low-VisionHuman-AI InteractionAssistive TechnologyAgentic AITrustMultimodal Interfaces
0
0 comments X

The pith

Explainable AI must address user self-blame to support blind and low-vision autonomy.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Modern AI is moving from simple question-answering to agents that perform multi-step tasks, but for blind and low-vision users, these systems remain opaque black boxes. When these agents fail, users frequently blame their own actions or environment rather than the system's internal logic. This paper argues that for AI to be truly useful as an assistive tool, it must move beyond visual charts to conversational, multimodal explanations that clarify exactly where an error occurred. Establishing this functional trust is necessary to prevent users from abandoning helpful technology due to misinterpreted failures.

Core claim

The authors identify a modality gap in explainable AI, where the lack of non-visual feedback leads blind and low-vision users to experience self-blame for system errors. They find that as AI becomes more agentic—taking autonomous, multi-step actions—the risk of cascading, invisible errors increases, making real-time, conversational interpretability a requirement rather than a feature. The discovery highlights that trust in assistive AI is not just about accuracy, but about the user's ability to diagnose whether they or the machine caused a specific outcome.

What carries the argument

Blame-aware explanation design, a framework that explicitly distinguishes between user-error, such as poor camera framing, and system-error, such as misclassification, to mitigate the psychological burden of self-blame.

If this is right

  • Assistive AI will shift from static descriptions to interactive dialogues where users can interrogate the reasoning behind a decision.
  • System designers will prioritize correction mechanisms that allow users to intervene before an autonomous agent completes an irreversible multi-step task.
  • Future accessibility standards will likely require multimodal transparency, ensuring that any visual explanation has a functional audio or haptic equivalent.
  • The design of AI agents will include explicit uncertainty signals to prevent users from over-relying on potentially flawed autonomous decisions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The self-blame phenomenon likely extends to other non-technical populations who may feel intimidated or at fault when interacting with complex black-box systems.
  • Reducing self-blame could significantly lower the cognitive load and emotional friction associated with adopting new assistive technologies.
  • A blame-aware system could provide the empirical basis for new legal or ethical frameworks regarding AI accountability in personal assistive contexts.

Load-bearing premise

The qualitative findings from a specific group of users accurately represent the diverse psychological reactions and technical needs of the global blind and low-vision community.

What would settle it

A large-scale longitudinal study showing that providing detailed conversational explanations does not reduce the rate of user self-blame or increase the long-term adoption of AI agents among blind and low-vision users.

read the original abstract

Explainable Artificial Intelligence (XAI) is critical for ensuring trust and accountability, yet its development remains predominantly visual. For blind and low-vision (BLV) users, the lack of accessible explanations creates a fundamental barrier to the independent use of AI-driven assistive technologies. This problem intensifies as AI systems shift from single-query tools into autonomous agents that take multi-step actions and make consequential decisions across extended task horizons, where a single undetected error can propagate irreversibly before any feedback is available. This paper investigates the unique XAI requirements of the BLV community through a comprehensive analysis of user interviews and contemporary research. By examining usage patterns across environmental perception and decision support, we identify a significant modality gap. Empirical evidence suggests that while BLV users highly value conversational explanations, they frequently experience "self-blame" for AI failures. The paper concludes with a research agenda for accessible Explainable AI in agentic systems, advocating for multimodal interfaces, blame-aware explanation design, and participatory development.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 3 minor

Summary. This paper explores the unique requirements for Explainable AI (XAI) within the blind and low-vision (BLV) community, particularly as assistive technologies transition from single-query tools (e.g., OCR) to autonomous agents. Through a qualitative analysis of user interviews and a synthesis of contemporary literature, the authors identify a significant 'modality gap' in current XAI, which remains predominantly visual. A central finding is the phenomenon of 'self-blame,' where BLV users attribute AI failures to their own inability to provide high-quality input (like camera framing). The authors argue for a shift toward multimodal, conversational, and 'blame-aware' XAI frameworks to foster functional trust in multi-step agentic systems.

Significance. The paper addresses a critical and underserved intersection: accessibility and agentic AI. As AI systems take on higher levels of autonomy, the 'black box' problem is exacerbated for BLV users who cannot rely on visual cues to debug or verify system state. The identification of 'self-blame' as a psychological barrier to trust is a significant contribution that moves the conversation beyond technical accuracy toward human-centric UX. The proposed research agenda, particularly the call for 'blame-aware' design and participatory development, provides a timely roadmap for the HCI and AI communities.

major comments (3)
  1. [§3.2, 'The Self-Blame Phenomenon'] There is a significant risk of construct conflation between 'input-quality anxiety' and 'agentic reasoning failure.' The evidence provided for self-blame largely stems from scenarios where the user acts as a sensor operator (e.g., aiming a camera). In the 'Agentic Era' described in §4, where systems perform multi-step actions autonomously, the user's role shifts. The manuscript lacks evidence that self-blame persists when the user is no longer the primary data-gatherer. If the self-blame is purely a reaction to ambiguous feedback regarding physical actions, 'conversational explanations' of internal agent logic may not address the root cause, which is a lack of tactile or auditory feedback regarding sensor state.
  2. [§4.2, 'Task Propagation and Feedback Loops'] The claim that agentic systems require 'blame-aware' design is central, yet the manuscript does not specify how an agent should distinguish between a user's 'input error' and its own 'model error' in a way that reduces self-blame without offloading responsibility onto the user. Without a concrete mechanism to disambiguate these sources of error for the user, the proposed framework remains conceptual and lacks a path to falsifiability or implementation.
  3. [§5.1, 'Toward Blame-Aware XAI'] The proposed agenda advocates for multimodal interfaces but lacks discussion on the 'cognitive load overhead' of conversational XAI. For BLV users, who already manage significant auditory throughput (e.g., screen readers, environmental sounds), adding multi-step conversational explanations of agentic reasoning may be counter-productive. The paper needs to address the trade-off between transparency and cognitive saturation in the 'Agentic Era' context.
minor comments (3)
  1. [Introduction] The definition of the 'Agentic Era' is somewhat vague. It would benefit from a concrete distinction between 'Agentic AI' and 'Assistive Tools' based on the level of delegation or the temporal horizon of the tasks.
  2. [§3.1] The modality gap is well-articulated, but the paper would be strengthened by citing existing work on 'Non-Visual XAI' (NVXAI) to better contextualize the novelty of the 'blame-aware' component.
  3. [Figure 1 (if applicable) or overall structure] The manuscript would benefit from a diagram illustrating the 'self-blame loop' vs. a 'blame-aware feedback loop' to clarify how the proposed interventions disrupt negative user perceptions.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the reviewer for their insightful critique, particularly regarding the evolution of the 'self-blame' phenomenon as we transition into the agentic era. The reviewer’s observation that the user's role shifts from a 'sensor operator' to an 'orchestrator' is a vital distinction that allows us to sharpen our theoretical framework. We have addressed the concerns regarding construct conflation, the need for concrete error-disambiguation mechanisms, and the cognitive load overhead associated with conversational interfaces for BLV users. We believe these revisions significantly move the manuscript from a conceptual exploration toward a more actionable research agenda.

read point-by-point responses

  1. Referee: [§3.2, 'The Self-Blame Phenomenon'] There is a significant risk of construct conflation between 'input-quality anxiety' and 'agentic reasoning failure.' ... The manuscript lacks evidence that self-blame persists when the user is no longer the primary data-gatherer.

    Authors: The reviewer makes a profound point regarding the shifting locus of responsibility. We agree that current empirical evidence for self-blame is largely tied to 'physical alignment' (sensor operation). However, we argue that in agentic systems, self-blame is likely to migrate from 'how I held the camera' to 'how I framed the request' or 'the state I left my environment in' (instructional and contextual anxiety). We have revised §3.2 to explicitly distinguish between 'sensor-operator blame' and 'contextual-input blame.' We also added a theoretical discussion in §4 on how 'instructional ambiguity' in multi-step agents can trigger similar psychological barriers to trust, even when the user is not the primary data-gatherer. revision: yes

  2. Referee: [§4.2, 'Task Propagation and Feedback Loops'] The claim that agentic systems require 'blame-aware' design is central, yet the manuscript does not specify how an agent should distinguish between a user's 'input error' and its own 'model error'...

    Authors: We acknowledge that the original manuscript remained at a high level of abstraction regarding implementation. To address this, we have expanded §4.2 to propose a 'Tripartite Error Taxonomy' (Sensor, Interpretation, and Planning). We now describe concrete interaction patterns for disambiguation, such as 'Diagnostic Probes' where the agent can communicate confidence levels specifically for input quality (e.g., 'I am 20% sure of the objects in this view due to lighting') versus reasoning (e.g., 'I see the medications clearly but cannot find a safe path to the counter'). This provides a path toward implementation and falsifiability as requested. revision: yes

  3. Referee: [§5.1, 'Toward Blame-Aware XAI'] The proposed agenda advocates for multimodal interfaces but lacks discussion on the 'cognitive load overhead' of conversational XAI.

    Authors: This is a critical oversight in the previous draft. For BLV users, auditory throughput is a scarce resource. We have added a new subsection to §5.1 titled 'The Transparency-Efficiency Trade-off.' In it, we discuss 'Tiered Disclosure' (providing minimal cues by default and detailed explanations only on request) and 'Sonification for State Awareness.' By using non-speech audio cues to indicate agent 'busy-ness' or 'certainty,' we can provide continuous transparency without the cognitive saturation that comes from dense verbal explanations. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified in empirical interview methodology

full rationale

The paper follows a standard qualitative research trajectory: conducting interviews with a specific demographic (BLV users) to identify psychological and functional barriers in AI usage. The central 'self-blame' finding is presented as an emergent theme from the interview data rather than a pre-defined axiom or an effect forced by specific self-citations. The transition from observed user behavior to design recommendations ('blame-aware' interfaces) is a linear application of design thinking. While the skeptic notes that the cause of self-blame may be conflated (input-quality anxiety vs. agentic failure), this reflects a challenge in construct validity or generalizability—specifically, whether findings from sensory-tool interactions apply to agentic ones—rather than a circular derivation where the conclusion is smuggled into the premises or defined by the methodology. The core result is independently sourced from the study's subjects.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The study's conclusions rely on the internal validity of user interviews and the assumption that participant self-reporting accurately reflects their cognitive trust models.

axioms (2)
  • domain assumption Trust is a prerequisite for effective use of AI assistive technology.
    The paper assumes that without trust/explainability, the technology will not be adopted or used safely (Section 1).
  • domain assumption Self-reporting in interviews accurately reflects user cognitive states.
    The methodology relies on qualitative interviews where users describe their own internal reactions to AI failure (Section 3).
invented entities (1)
  • Blame-aware explanation design no independent evidence
    purpose: To mitigate user self-blame during AI failures by explicitly communicating system limitations and error sources.
    This is a proposed design framework introduced in the paper to address the discovered behavioral pattern.

pith-pipeline@v0.9.0 · 7223 in / 1639 out tokens · 41245 ms · 2026-05-08T02:24:24.262482+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages

  1. [1]

    I look at it as the king of knowledge

    Rudaiba Adnin and Maitraye Das. 2024. "I look at it as the king of knowledge": How Blind People Use and Understand Generative AI Tools. InProceedings of the 26th International ACM SIGACCESS Conference on Computers and Accessibility(St. John’s, NL, Canada)(ASSETS ’24). Association for Computing Machinery, New York, NY, USA, Article 64, 14 pages. doi:10.114...

    work page doi:10.1145/3663548.3675631 2024

  2. [2]

    Trying to Piece It Together

    Rahaf Alharbi, Angela D Cheong, Jaylin Herskovitz, Robin N. Brewer, and Sarita Schoenebeck. 2025. “Trying to Piece It Together”: Exploring Accessible Error Detection in Emerging Privacy Techniques With Blind People. InProceedings of the 27th International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS ’25). Association for Computing Machi...

    work page doi:10.1145/3663547.3746376 2025

  3. [3]

    Rahaf Alharbi, Pa Lor, Jaylin Herskovitz, Sarita Schoenebeck, and Robin N. Brewer

    work page

  4. [4]

    In Proceedings of the 26th International ACM SIGACCESS Conference on Computers and Accessibility(St

    Misfitting With AI: How Blind People Verify and Contest AI Errors. In Proceedings of the 26th International ACM SIGACCESS Conference on Computers and Accessibility(St. John’s, NL, Canada)(ASSETS ’24). Association for Computing Machinery, New York, NY, USA, Article 61, 17 pages. doi:10.1145/3663548.3675659

    work page doi:10.1145/3663548.3675659

  5. [5]

    Virginia Braun and Victoria Clarke. 2006. Using thematic analysis in psychology. Qualitative research in psychology3, 2 (2006), 77–101

    work page 2006

  6. [6]

    Shiye Cao and Chien-Ming Huang. 2022. Understanding User Reliance on AI in Assisted Decision-Making.Proc. ACM Hum.-Comput. Interact.6, CSCW2, Article 471 (Nov. 2022), 23 pages. doi:10.1145/3555572

    work page doi:10.1145/3555572 2022

  7. [7]

    Yu-An Chen, Zicheng Zhu, and Yi-Chieh Lee. 2025. Humans Help Conversational AI: Exploring the Impact of Perceived AI Deservingness on People’s Decisions to Help and Their Perceptions on AI Seeking Help. InProceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems (CHI EA ’25). Association for Computing Machinery, New...

    work page doi:10.1145/3706599.3720012 2025

  8. [8]

    Claggett, Robert E

    Elijah L. Claggett, Robert E. Kraut, and Hirokazu Shirado. 2025. Relational AI: Facilitating Intergroup Cooperation with Socially Aware Conversational Support. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 170, 22 pages. doi:10.1145/3706598.3713757

    work page doi:10.1145/3706598.3713757 2025

  9. [9]

    Sven Eckhardt, Niklas Kühl, Mateusz Dolata, and Gerhard Schwabe. 2025. A Survey of AI Reliance.ACM Comput. Surv.58, 6, Article 161 (Dec. 2025), 37 pages. doi:10.1145/3776528

    work page doi:10.1145/3776528 2025

  10. [10]

    Weisz, and Andreas Riener

    Upol Ehsan, Elizabeth A Watkins, Philipp Wintersberger, Carina Manger, Nina Hubig, Saiph Savage, Justin D. Weisz, and Andreas Riener. 2025. New Frontiers of Human-centered Explainable AI (HCXAI): Participatory Civic AI, Benchmarking LLMs, XAI Hallucinations, and Responsible AI Audits. InProceedings of the Extended Abstracts of the CHI Conference on Human ...

    work page doi:10.1145/3706599.3706713 2025

  11. [11]

    Vera Liao, Elizabeth Anne Watkins, Carina Manger, Hal Daumé III, Andreas Riener, and Mark O Riedl

    Upol Ehsan, Philipp Wintersberger, Q. Vera Liao, Elizabeth Anne Watkins, Carina Manger, Hal Daumé III, Andreas Riener, and Mark O Riedl. 2022. Human-Centered Explainable AI (HCXAI): Beyond Opening the Black-Box of AI. InExtended Abstracts of the 2022 CHI Conference on Human Factors in Computing Systems (New Orleans, LA, USA)(CHI EA ’22). Association for C...

    work page doi:10.1145/3491101.3503727 2022

  12. [12]

    Upol Ehsan, Philipp Wintersberger, Elizabeth A Watkins, Carina Manger, Gonzalo Ramos, Justin D Weisz, Hal Daumé Iii, Andreas Riener, and Mark O Riedl. 2023. Human-centered explainable AI (HCXAI): coming of age. InExtended abstracts of the 2023 CHI conference on human factors in computing systems. 1–7

    work page 2023

  13. [13]

    Jonggi Hong. 2018. Accessible Human-Error Interactions in AI Applications for the Blind. InProceedings of the 2018 ACM International Joint Conference and 2018 International Symposium on Pervasive and Ubiquitous Computing and Wearable Computers(Singapore, Singapore)(UbiComp ’18). Association for Computing Machinery, New York, NY, USA, 522–528. doi:10.1145/...

    work page doi:10.1145/3267305.3267321 2018

  14. [14]

    Explain What a Treemap is

    Gyeongri Kim, Jiho Kim, and Yea-Seul Kim. 2023. “Explain What a Treemap is”: Exploratory Investigation of Strategies for Explaining Unfamiliar Chart to Blind and Low Vision Users. InProceedings of the 2023 CHI Conference on Human Factors in Computing Systems(Hamburg, Germany)(CHI ’23). Association for Computing Machinery, New York, NY, USA, Article 805, 1...

    work page arXiv 2023

  15. [15]

    Scott M Lundberg and Su-In Lee. 2017. A unified approach to interpreting model predictions.Advances in neural information processing systems30 (2017)

    work page 2017

  16. [16]

    Nora McDonald, Sarita Schoenebeck, and Andrea Forte. 2019. Reliability and inter-rater reliability in qualitative research: Norms and guidelines for CSCW and HCI practice.Proceedings of the ACM on human-computer interaction3, CSCW (2019), 1–23

    work page 2019

  17. [17]

    Cecily Morrison, Martin Grayson, Rita Faia Marques, Daniela Massiceti, Camilla Longden, Linda Wen, and Edward Cutrell. 2023. Understanding Personalized Accessibility through Teachable AI: Designing and Evaluating Find My Things for People who are Blind or Low Vision. InProceedings of the 25th International ACM SIGACCESS Conference on Computers and Accessi...

    work page doi:10.1145/3597638.3608395 2023

  18. [18]

    Cliodhna O’Connor and Helene Joffe. 2020. Intercoder reliability in qualitative research: Debates and practical guidelines.International journal of qualitative methods19 (2020), 1609406919899220

    work page 2020

  19. [19]

    Bianca Giulia Sarah Schor, Emma Kallina, Jatinder Singh, and Alan Blackwell

    work page

  20. [20]

    Visibility into AI agents

    Meaningful Transparency for Clinicians: Operationalising HCXAI Research with Gynaecologists. InProceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency(Rio de Janeiro, Brazil)(FAccT ’24). Association for Computing Machinery, New York, NY, USA, 1268–1281. doi:10.1145/3630106. 3658971

    work page doi:10.1145/3630106 2024

  21. [21]

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedan- tam, Devi Parikh, and Dhruv Batra. 2017. Grad-cam: Visual explanations from Explainable AI for Blind and Low-Vision Users Human-centered Explainable AI Workshop (HCXAI) @ CHI 2026, April 13, 2016, Barcelona, Spain deep networks via gradient-based localization. InProceedings of the...

    work page 2017

  22. [22]

    Kamath, Aziz Zeidieh, Saairam Venkatesh, and Sean McCurry

    JooYoung Seo, Sanchita S. Kamath, Aziz Zeidieh, Saairam Venkatesh, and Sean McCurry. 2024. MAIDR Meets AI: Exploring Multimodal LLM-Based Data Visual- ization Interpretation by and with Blind and Low-Vision Users. InProceedings of the 26th International ACM SIGACCESS Conference on Computers and Accessibility (St. John’s, NL, Canada)(ASSETS ’24). Associati...

    work page doi:10.1145/3663548.3675660 2024

  23. [23]

    Ian René Solano-Kamaiko, Dibyendu Mishra, Nicola Dell, and Aditya Vashistha

    work page

  24. [24]

    InProceedings of the 2024 CHI Conference on Human Factors in Computing Systems(Honolulu, HI, USA)(CHI ’24)

    Explorable Explainable AI: Improving AI Understanding for Community Health Workers in India. InProceedings of the 2024 CHI Conference on Human Factors in Computing Systems(Honolulu, HI, USA)(CHI ’24). Association for Computing Machinery, New York, NY, USA, Article 399, 21 pages. doi:10.1145/ 3613904.3642733

    work page arXiv 2024

  25. [25]

    Xinru Tang, Ali Abdolrahmani, Darren Gergle, and Anne Marie Piper. 2025. Everyday Uncertainty: How Blind People Use GenAI Tools for Information Access. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 63, 17 pages. doi:10.1145/3706598.3713433

    work page doi:10.1145/3706598.3713433 2025