TL;DR
This study investigates whether young adults (Gen Z) verify evidence provided by AI-generated health answers, finding that most do not, even when the answers are suboptimal. The research highlights a critical gap: while AI can provide answers, ensuring users engage with the evidence for better health decisions remains a significant UI/UX challenge.
What Problem Does This Paper Address?
The increasing adoption of AI-powered question-answering (QA) systems in healthcare raises concerns about misinformation, particularly "hallucinations" from Large Language Models (LLMs). While LLMs can be prompted to provide references, it's unclear if users actually use this evidence to verify the answers. This is especially critical for younger demographics, like Gen Z, who are highly susceptible to online health misinformation. The paper aims to understand how user interface (UI) and user experience (UX) design can encourage users to explore provided evidence and, consequently, make better health-related decisions. Existing research often focuses on minimizing cognitive load, but this study seeks to increase user engagement with evidence without overwhelming them.
Methods
The researchers designed an experiment using eight different UI layouts for presenting health-related questions and AI-generated answers with supporting evidence. These layouts varied in:
- Answer presentation: Full paragraph vs. sentence-by-sentence interleaved with references.
- Image placement: Horizontal vs. vertical scroll bars.
- Link augmentation: Text-only links vs. links with prominent icons.
These variations resulted in eight distinct interface types (A-H), presented in an 8x8 Latin Square design to control for order effects. Eight participants, fitting the vulnerable age range (18-34) and with varying health literacy levels, were recruited. After completing a health literacy assessment, participants were presented with eight health scenarios. They were instructed to read the scenario, explore the answer and evidence, and then make a health-related decision or answer a question. User interactions, including link clicks and time spent, were logged. A post-study survey and focus group aimed to gather qualitative feedback on user preferences and trust in the system.
Datasets
The study did not use pre-existing datasets for training AI models. Instead, it created its own "datasets" in the form of eight health scenarios. These scenarios were composed of:
- Questions: Derived from publicly available medical question collections.
- Reference answers: Sourced from reliable medical literature.
- Linked evidence: Relevant images and scientific publications.
The health literacy evaluation used a test designed by Schwartz, Woloshin, and Welch (2005), which has established attributes of reliability and validity for medical data interpretation.
Key Insights and Ideas
- The "Blind Acceptance" Phenomenon: A significant finding is the "alarming tendency" for young, well-educated users with fair health literacy to blindly accept AI-generated answers, leading to suboptimal health decisions. This underscores that simply providing evidence isn't enough.
- Evidence Exploration is Rare: Only about a third of participants explored the provided evidence. Those who did explore the evidence made better health-related decisions.
- UI Preferences vs. Actual Behavior: While participants expressed preferences for certain layouts (e.g., full answers with unobtrusive links and horizontal image scroll bars), none of the tested UI designs consistently motivated users to delve into the evidence. This suggests the UI/UX tested didn't sufficiently overcome user inertia.
- Trust is Key, and Often Implicit: Most participants trusted the AI-generated answers without verification. Only those who actively sought evidence expressed distrust, implying that the responsibility for ensuring accuracy is implicitly placed on the system by the user.
- Background Knowledge Drives Verification: The decision to seek supporting evidence was more strongly driven by users' existing knowledge, understanding, and confidence in the generated answers, rather than solely by the UI design.
Key Results
- Decision Quality: Participants who explored the evidence made better health-related decisions.
- Evidence Engagement: Only approximately one-third of participants explored the provided evidence.
- UI Preference: Interface Type C (text blob answer, horizontal image scroll bar, text links) was identified as the most popular layout.
- Time Spent: A moderately positive correlation was observed between the time spent reviewing answers/evidence and the quality of follow-up decisions.
- Trust: Seven out of eight participants indicated it did not matter how the answer was generated, suggesting a high level of implicit trust in the system.
Limitations and Future Work
Limitations: * Small Sample Size and Homogeneity: The pilot study involved only 8 users, all from a single age group (Gen Z). While this group is tech-savvy, their behavior might not be representative of the broader population. * Limited UI Variations: The study tested a specific set of UI/UX design choices; other variations might yield different results.
Future Work: * Larger and Diverse Samples: Expand experiments to include a more diverse and larger pool of users to improve generalizability. * Advanced Tracking: Incorporate eye-gaze tracking and qualitative feedback to gain deeper insights into user behavior. * More UI Variables: Experiment with additional design elements like color schemes and font emphasis. * Usability Assessment: Design a comprehensive questionnaire to assess system usability using metrics like the System Usability Scale (SUS). * Refining Engagement Strategies: Further research is needed to develop UI/UX designs that effectively encourage users to verify evidence and lead to optimal health-related decisions. * System Responsibility: The findings strongly suggest that the onus is on the AI systems to ensure the correctness and trustworthiness of the information they provide, as users are unlikely to verify it themselves.
Summary updated