Everyday Uncertainty: How Blind People Use GenAI Tools for Information Access

Research Background and Problem

Identified Issues or Challenges

• With the widespread introduction of generative artificial intelligence (GenAI) tools into accessibility technologies, blind individuals face new opportunities and challenges in accessing information.
• Outputs generated by GenAI tools may contain errors, hallucinations, and biases. Additionally, the evolving capabilities of these tools introduce uncertainties for blind users.
• Blind individuals routinely encounter common accessibility issues, such as difficult-to-navigate websites, inaccessible documents, and societal misconceptions about disabilities.
• Few studies have explored how blind individuals utilize these GenAI tools across diverse tasks and contexts.

Importance

• GenAI tools have the potential to enhance access to non-visual information through large language models (LLMs) and multimodal models (LMMs). However, the limitations of these tools impact the independence of blind individuals and may increase the risks associated with technology use.
• Since blind individuals rely on information access for daily tasks, learning, and social interactions, it is crucial to study how they address the challenges posed by generative AI outputs to propose improvements.

Research Motivation and Related Work

• While existing research has focused on the technical limitations of GenAI tools (e.g., data quality issues and model biases), little work has examined how blind individuals adapt their strategies to cope with the uncertainties introduced by these tools.
• The issue of uncertainty in AI research is often explored from a computational perspective (e.g., quantifying model prediction uncertainty), with limited attention to its application and understanding in human contexts.
• This study aims to provide a comprehensive understanding of how blind individuals utilize tools (including GenAI, human assistance, and other social-material resources) to navigate these challenges.

Solution

Methods or Solutions

• The study introduces the concept of “everyday uncertainty”, which reflects the mindset of blind individuals as they navigate and adapt to multiple uncertainties while using GenAI tools to access information. These uncertainties include tool imperfections, the diversity of human support, and the subjectivity of information itself.
• Through qualitative research, the study elucidates how blind individuals adjust their strategies to incrementally refine their understanding when faced with erroneous, incomplete, or subjective information.

Innovations of the Solution

• Adopting a systemic perspective, the study views GenAI tools as part of a multimodal workflow for blind individuals rather than as isolated technological products.
• Reframes uncertainty from a quantifiable problem to a lived practice, positioning it as an integral part of blind individuals' confidence and adaptability in navigating information environments.
• Expands the understanding of uncertainty in the HCI (Human-Computer Interaction) field, treating it as a positive driver for analysis and design rather than a problem to be solved.

Implementation Steps and Key Techniques

1. Data Collection and Preliminary Study: The research team observed and analyzed natural interactions between researchers and tools such as ChatGPT, Gemini, and Be My AI to contextualize the problem.
2. Participant Interviews: Conducted structured interviews with 20 blind screen reader users to understand how they use GenAI tools and adjust their strategies.
3. Data Analysis: Employed reflexive thematic analysis to identify thematic practices and patterns of uncertainty management in information access.

Research Outcomes

Specific Findings

• Introduction of the Concept of Everyday Uncertainty: Defined as the ongoing process by which blind individuals question and explore errors, incompleteness, and subjectivity in information.
• Thematic Analysis identified four key coping strategies:
  1. Extracting meaningful cues from high-error outputs: For example, using document structure or leveraging common errors as clues for verification.
  2. Incrementally updating information and understanding: Treating information as dynamic rather than “final and correct,” and continuously refining and supplementing understanding.
  3. Acknowledging the subjectivity of information and exploring alternative perspectives: Relying on AI, online resources, and human assistance to form diverse understandings.
  4. Adapting to the political environment of information access: Adjusting request strategies within societal norms and tool limitations.

Comparative Advantages Over Existing Solutions

• Introduces a “systemic and collaborative perspective” that goes beyond studying tools in isolation to examining all factors involved in the information access pathway.
• Provides a deeper understanding of how blind individuals employ critical, creative, and adaptive approaches to navigate errors in generative AI outputs, surpassing traditional analyses focused solely on “accuracy/inaccuracy.”
• Offers practical design insights and user behavior findings that can inform improvements in AI models and user experience design.

Experimental or Evaluation Results

• The study confirms that blind individuals are accustomed to tolerating and managing certain errors and uncertainties in information access. For instance, many participants used generative AI to create outlines to save reading time or extracted keywords to enhance search efficiency.
• Real-life examples from interviews (e.g., using AI to describe personal photos or appliance components) further illustrate the adaptability and dynamism of “everyday uncertainty.”

Limitations and Future Directions

• Limitations:
  • The study sample is concentrated in the United States, lacking a global perspective.
  • Heavy reliance on self-reports and interviews; future research could incorporate more objective behavioral data.
• Future Directions:
  1. Explore designs that enable seamless switching between multiple AI tools and human resources within an information ecosystem.
  2. Further investigate how AI tools can transparently convey model limitations in specific usage contexts.
  3. Examine policy and institutional approaches to standardize and control technology as a means of addressing the root causes of uncertainty.

Conclusion

This study analyzes how blind individuals use generative AI tools to navigate everyday uncertainties in information access, proposing a new perspective that integrates uncertainty as an organic component of the process. From design and strategy standpoints, the research provides actionable pathways for developing more flexible and adaptive accessible AI tools, while also contributing fresh theoretical insights to the fields of Human-Computer Interaction and accessibility.