UX for AI-Powered Products: Balancing Magic and User Trust
Download as PPTX, PDF0 likes131 views
The document discusses user trust in AI-powered products, highlighting that many users are unaware of AI's influence on applications and express concerns about its risks, with 82% wanting to learn more about AI. It outlines strategies for calibrating user trust through explainable AI (XAI) techniques and emphasizes the importance of timely explanations tied to user actions. The paper also offers insights on overcoming distrust and overtrust in AI systems while suggesting methods to validate XAI through user testing.
UX for AI-Powered Products: Balancing Magic and User Trust
- 1. UX for AI-powered products: Balancing Magic and User Trust by Kateryna Korovkina
- 2. About Me Curator at Projector Institute guiding students to kickstart their career in UX Design Manager at Eleks UK building our design expertise in the UK and Ireland 11+ years in Product Design helping clients build human-centered experiences for complex products Reach out on LinkedIn: Kateryna Korovkina
- 3. Are people aware they are interacting with AI-powered products? do not know that AI enables common applications they use* 41% * According to a 2023 global study by KPMG and the University of Queensland (17,193 respondents across 17 countries)
- 4. Do people have trust in AI? are wary about the use of AI in digital products 82% want to learn more about AI 61% At the same time 50% of respondents agree that the benefits of AI outweigh the risks 66% of people feel optimistic about the future of AI. Fear and worry about AI dominate in Australia, Canada, France, and Japan, with people in France the most worried, and outraged about AI.
- 5. EU Principles and practices for design of Trustworthy AI
- 6. EU Principles and practices for design of Trustworthy AI
- 7. Calibrating user trust in AI systems AI capability User trusr Distrust Overtrust Calibrated trust People trust AI when they shouldn’t People reject to use AI when they should
- 8. AI-powered Google Maps nav. planner excludes certain available routes based on complex calculations Fixed algorithm of the Transport for London journey planner covers all available options Both distrust and overtrust cause mistakes in decision making
- 9. Calibrating trust via explainable AI (XAI) strategies High Low Low High Confidence of AI Risk and impact Provide explanations Enable higher user control Show benefits Avoid lengthy explanations Provide explanations, confidence scores Enable user to teach AI via feedback Provide explanations and confidence scores Enable user to overtake control and AI to learn STRATEGY 1 STRATEGY 3 STRATEGY 2 STRATEGY 4
- 10. 1 MITIGATE DISTRUST Communicate value for the user and include subtle explainability cues about AI decisions in low-stakes scenarios to build user trust High Low Low High Confidence of AI Risk and impact Provide explanations Enable higher user control Show benefits Avoid lengthy explanations Provide explanations, confidence scores Enable user to teach AI via feedback Provide explanations and confidence scores Enable user to overtake control and AI to learn STRATEGY 1 STRATEGY 3 STRATEGY 2 STRATEGY 4
- 11. 2 MITIGATE DISTRUST + ENABLE FEEDBACK Enable user feedback when it can improve the system’s confidence scores. Acknowledge the feedback, act upon it or let users know when adjustments will happen. High Low Low High Confidence of AI Risk and impact Provide explanations Enable higher user control Show benefits Avoid lengthy explanations Provide explanations, confidence scores Enable user to teach AI via feedback Provide explanations and confidence scores Enable user to overtake control and AI to learn STRATEGY 1 STRATEGY 3 STRATEGY 2 STRATEGY 4
- 12. 3 MITIGATE OVERTRUST + ENABLE CONTROL Give the user details about why a prediction was made in a high stakes scenario, even if the confidence score is relatively high High Low Low High Confidence of AI Risk and impact Provide explanations Enable higher user control Show benefits Avoid lengthy explanations Provide explanations, confidence scores Enable user to teach AI via feedback Provide explanations and confidence scores Enable user to overtake control and AI to learn STRATEGY 1 STRATEGY 3 STRATEGY 2 STRATEGY 4
- 13. 4 MITIGATE OVERTRUST + ENABLE FULL CONTROL Enable user to overtake control in high- stakes scenarios when confidence is low (e.g., suspecting a fraudulent transaction via AI algorithms) High Low Low High Confidence of AI Risk and impact Provide explanations Enable higher user control Show benefits Avoid lengthy explanations Provide explanations, confidence scores Enable user to teach AI via feedback Provide explanations and confidence scores Enable user to overtake control and AI to learn STRATEGY 1 STRATEGY 3 STRATEGY 2 STRATEGY 4
- 14. When and how should we calibrate trust? Timing of explanations is highly dependent on the context and requires a thorough understanding of the user journey 01 Initially explain what the system can do Communicate the product’s capabilities and limitations clearly to set expectations and do this early. 02 Tie explanations to user actions People learn faster when they can see a response to their actions right away, because then it’s easier to identify cause and effect. 03 Experiment to optimise for maximum attention Increase perceived value of explanations, reduce their complexity to make sure users don’t skip them; include elements of friction for the top priority scenarios.
- 15. Choosing XAI methods based on experience mapping and research
- 16. User testing to validate effectiveness of explanations • “On this scale, show me how trusting you are of this recommendation.” • “What questions do you have about how the app came to this recommendation?” • “What, if anything, would increase your trust in this recommendation?”
- 17. Validating XAI in functional products via qual and quant UX methods
- 18. Explainability design patterns 1. Narrative-based explanations in consumer-grade apps 2. Saliency maps overlapping medical images as a graphic representation method for black-box AI models 3. “What If”, explainable AI tool by Google, allowing users to understand how a model works. It enables to explore how individual features affect the model's predictions
- 19. List of resources 1. EU Ethics Guidelines for Trustworthy AI 2. People + AI Research by Google 3. Explainable recommendations and calibrated trust: two systematic user errors 4. Microsoft Guidelines for Human-AI Interaction 5. Trust in Artificial Intelligence: A Global Study by KPMG and University of Queensland
- 20. Thank you for your attention! Have a question? Write to kateryna.korovkina@eleks.com Reach out to me on Linkedin