The User Experience on How Young Adults Decide When to Verify AI Recommendations and Voice Assistants: A Qualitative Study
UI/UX | UX Research | Workshops | Branding | Overall Design Works
Understanding when young people feel the need to verify AI recommendations versus following them directly, and the underlying decision-making factors
Duration
Q1-Q4
Q1-Q4
Team
Michelle Cedeño
Michelle Cedeño
My Role
- Engaged with stakeholders and identify their interests and requirements in the problem space
- Led the interview protocol design and IRB application process in collaboration with Computer Science and Psychology teams
- Recruited participants and conducted eight interviews
- Preprocessed data: converted and cleaned the transcripts
- Collaborated with a researcher from the Informatics department to conduct inductive thematic analysis on the qualitative data
- Periodically presented findings and insights to both internal stakeholders
Methods
Figma,
Midjourney,
Maze,
After Effects
Figma,
Midjourney,
Maze,
After Effects
Introduction
Large language model-based conversational AI assistants like ChatGPT or Google Gemini are very popular now, but they treat every suggestion the same, whether you're asking about a million-dollar investment or how long to cook pasta. Users get frustrated when simple questions come with unnecessary verification steps, while important decisions lack proper fact-checking. We explore what factors make users want more or less fact-checking for different AI suggestions, making AI both more trustworthy for big decisions and more convenient for everyday use.
Project Objective
Investigate the verification attitude patterns of young adults for AI recommendations. It will help develop evidence-based design principles for context-aware verification systems that automatically calibrate verification levels based on users personal need. The ultimate goal is to improve user satisfaction and platform trust while reducing session abandonment.
Problem Statement
Research Question: When do young people feel the need to verify AI recommendations versus follow them directly? And Why?
Research Timeline
Outline scope, define research objectives, and align research teams
Design interview protocol, secure ethics approval, and recruit participants
Conduct data collection, perform analysis, and hold weekly check-ins
Deliver findings, provide recommendations, and propose paper writing plan
Research MethodIn-depth Interview as Method
We selected in-depth interviews as our optimal research method after evaluating the alternative methods such as online survey, focus groups, and diary studies. Our team, working closely with teams from Psychology and Informatics department of Indiana University, determined that while behavioral surveys could provide larger samples, they lack the follow-up capabilities needed to understand nuanced verification decision-making processes. In contrast, in-depth interviews provide deep mental model insights and rich contextual understanding that other methods cannot capture. While time-intensive with smaller samples, we selected this approach as it would reveal the nuanced decision-making processes behind AI verification behaviors.
Participant Recruitment
We employed purposive sampling in collaboration with Informatics and Psychology teams to recruit 14 participants aged 18-25 with demonstrated AI experience.
- Recruitment strategy: University networks, social media, and snowball sampling coordinated with outreach teams
- Sample rationale: Data saturation occured after 12 interviews (validated with research partners). Concluded after 14 pre scheduled interviews
- Inclusion criteria: Recent AI chatbot usage for recommendations within past month, screened with open-ended experience questions
- Quality assurance: Verified participant authenticity and experience levels with help of the research partners
Interview Protocol
We conducted semi-structured Zoom interviews, with each 40-45 minute session following our carefully designed protocol.
- Protocol development: Collaborated with Psychology researchers experienced in designing questions for young adults
- Key questions: Open ended memory anchoring such as "Walk us through your last AI recommendation experience" and "How did you decide to verify or trust that advice?"
- Pilot testing: Refined protocol with design team feedback after testing with 2 participants
- Technology coordination: Ensured secure recording and transcription capabilities upon proper consent from the participants
Data Analysis Approach
I worked with one researcher from Psychology team to employ inductive coding and thematic analysis.
- Coding process: Two researchers independently coded interviews and evaluated inter rater reliability after coding same two participant data
- Collaboration method: Regular analysis sessions with the stakeholders to interpret findings
- Framework development: Cross-functional teams identified three-stage decision framework through collaborative analysis
- Validation: Computer Science, Psychology, and Informatics teams triangulated findings in light of prior studies and business alignments
Key Findings: Verification Decision Framework
Young adults follow a consistent three-step decision process when determining whether to verify AI recommendations:
Stake Assessment → Domain Familiarity → Action Decision
This framework reveals that verification behavior is predictable based on two key factors:
Stake Assessment → Domain Familiarity → Action Decision
This framework reveals that verification behavior is predictable based on two key factors:
- What's at stake
- How familiar users are with the domain
Verification Attitude Matrix
User Stories & Insights
Design Recommendations for Potential External Stakeholders
- Query Analysis Engine: Develop ML models to automatically detect risk indicators and domain complexity in real-time user queries
Modular Verification API: Build dynamic verification components that assemble appropriate responses based on users' stake/familiarity assessment
Management
- Phased Rollout Strategy: Implement verification features gradually starting with high-stakes domains, allowing user feedback integration
Performance Metrics Dashboard: Establish KPIs tracking verification satisfaction and trust scores across different scenarios
Leadership
- Competitive Differentiation: Position intelligent verification as core advantage, emphasizing trust in market positioning
Strategic Investment: Prioritize verification system as foundational investment enabling premium pricing and subscription growth
Business Impact
Reduced Friction Improve User Satisfaction
Context-aware verification will eliminate unnecessary friction, providing minimal verification for familiar queries. This will create faster task completion, higher satisfaction scores, increased daily users, and better retention as AI would feel intuitive rather than bureaucratic.
Context-aware verification will eliminate unnecessary friction, providing minimal verification for familiar queries. This will create faster task completion, higher satisfaction scores, increased daily users, and better retention as AI would feel intuitive rather than bureaucratic.
Necessary Verification Increase Safety Perception
Robust verification for high-stakes scenarios will build trust and will position platforms as responsible. Comprehensive citations and uncertainty indicators for critical advice would increase user confidence, generate positive word-of-mouth, and create competitive differentiation
Robust verification for high-stakes scenarios will build trust and will position platforms as responsible. Comprehensive citations and uncertainty indicators for critical advice would increase user confidence, generate positive word-of-mouth, and create competitive differentiation
Increased reliability Would Drive Subscription Growth
Balanced verification would drive subscription conversion and retention. Users experiencing appropriately calibrated responses will likely upgrade to premium services, recommend the platform, and maintain long-term subscriptions, justifying premium pricing and driving revenue growth.
.
Future Research Directions
Behavioral Observation
Our interview findings are based on self-reported behaviors, which may not reflect actual verification actions. Direct observation would validate whether users truly follow the stake/familiarity framework we identified and reveal unconscious verification patterns that users cannot articulate in interviews.
User Segmentation
Our purposive sampling recruited participants based solely on AI experience, but verification behaviors may vary significantly across user segments with different professional backgrounds, technical expertise, cultural contexts, and demographic diversity. Segmented recruitment targeting specific user groups (students, professionals, casual users, power users) across diverse cultural backgrounds, ethnicities, and socioeconomic levels would reveal whether the stake/familiarity framework applies consistently across different populations or requires culturally-sensitive and demographically-tailored customization.Longitudinal Validation
Our research captured verification behaviors at a single point in time, but trust and verification patterns likely evolve as users gain AI experience. Understanding this progression would inform onboarding strategies and adaptive verification systems that adjust to user expertise over time.