20 Years of AI Design Summary: A Guide to Infusing AI into Products
-
As AI interaction becomes increasingly popular, many articles have explored its design. In this piece, the designer summarizes lessons learned from 20 years of AI design, presenting 18 interaction design guidelines that offer valuable design principles, hoping to provide useful insights.
Today, many products incorporate AI-powered features. We carry AI systems in various forms on our mobile devices daily, such as activity trackers, email filters, autocomplete, and social network feed ads. Virtual assistants like Siri, Google Home, and Cortana help us with simple tasks. Recommendation systems suggest our next purchase on Amazon or guide us to the next Netflix series we might enjoy.
The Nature of AI-Infused Products
A common characteristic of AI-infused products is that, unlike other systems, they rely on probabilities and thus involve uncertainty. AI-powered products are inherently inconsistent because they learn over time. They may also respond differently depending on lighting conditions, ambient noise, human accents, and other situational factors. This unpredictability can confuse users, reduce their trust, and ultimately lead them to abandon the product.
AI Interaction Design Guidelines
To ensure good design, Amershi and colleagues recommend several AI interaction guidelines. These design principles synthesize over 20 years of lessons in the field. They gathered 168 design recommendations from white papers, reviews, news articles, editorials, and academic papers, consolidating them into the following 18 guidelines.
DG1: Clarify What the System Can Do
Help users understand the capabilities of the AI system. For example, be transparent about what data is tracked and how it is processed.
DG2: Clarify How Well the System Performs
Help users understand how often the AI system might make mistakes.
DG3: Context-Aware Timing of Services
The AI system should consider when to act or interrupt the user based on the current task and context.
DG4: Display Relevant Information
Show information relevant to the user's current task and environment.
Searching for the movie title "Weathering With You" should display relevant screening times and locations.
DG5: Align with Social Norms
Provide an experience that aligns with users' social and cultural expectations. For example, voice assistants should use appropriate language when interacting with families with children.
DG6: Mitigate Social Biases
Prevent the AI system's language and behavior from reinforcing harmful stereotypes and biases.
DG7: Support Efficient Invocation
Simplify how users request services from the AI system.
Users can directly say the wake word "Alexa" to start interacting with Amazon Echo.
DG8: Support Efficient Dismissal
Simplify how users can reject or ignore unwanted services.
DG9: Support Efficient Correction
Simplify methods for editing or recovering from errors. For example, after setting a reminder via a voice assistant, users should be able to find and manually edit it on their phone.
DG10: Handle Ambiguity Gracefully
When uncertain about the user's intent, engage in disambiguation or gracefully scale back the AI system's services.
Autocomplete should offer scrollable suggestions rather than automatically filling in words for the user.
DG11: Explain Why the System Behaves a Certain Way
The AI system's behavior should be transparent, explaining why it acts as it does.
Google Maps displays "fastest route" as the reason for choosing a specific path.
DG12: Remember Recent Interactions
Maintain short-term memory and allow users to reference it effectively. For example, search engines should understand subsequent queries based on previous ones.
DG13: Learn from User Behavior
Continuously learn from user behavior to personalize the experience.
DG14: Update and Adapt Carefully
Avoid disruptive changes when updating or adjusting the AI system's behavior.
When users select a song, avoid suddenly updating all top recommendations. Keep the top selections stable and update the rest.
DG15: Encourage Detailed Feedback
Enable users to provide feedback indicating their preferences during regular interactions. For example, if the AI system fails to mark an email as important, users should be able to do so manually.
DG16: Communicate Consequences of User Actions
Immediately update or inform users how their actions will affect the AI system's future behavior.
Instagram notifies users that hiding an ad will help improve future ad quality.
DG17: Provide Global Controls
Allow users to globally customize what the AI system monitors and how it behaves. For example, users should be able to toggle location sharing on and off.
DG18: Notify Users About Changes
Inform users when the AI system adds or updates features. For example, the AI system could offer tutorials for new functionalities.
These guidelines are derived from Amershi et al.'s AI Interaction Guidelines published at CHI 2019.