Designing for machine intelligence (UX of AI)

Each design material comes with unique challenges. In the same way that designing a poster is different from designing a mobile app, designing AI-driven applications is different from designing apps.

While AI applications are not yet widespread, the interaction and user experience design challenges - ranging from interface solutions to the greater ethical challenges - are beginning to emerge through practice and research.

Below, we identify and introduce you to 9 challenges in designing for machine intelligence including examples and design patterns.

User Trust & Expectations

Explainability:Peeking inside the black box To build trust and understanding, communicate to your user how and why the system acts, generates, or concludes the output in a simplified and gradual way. Construct helpful mental models to make sense of the machine. Peeking inside the black box.  Related patterns: Exclude this activity, data swipe, disable intelligence

Managing expectationsand preventing disappointment From the beginning, set expectations about what the system can and can not do, and build interactions in ways that match the system’s abilities and limitations. Manage expectations to prevent disappointment. Related patterns: Exclude this activity, data swipe, disable intelligence

Failure-firstFail gracefully With less control over the ultimate outcome, design your interface assuming failure to minimize user frustration when the system produces faulty, imperfect, or even absurd results. Offer a way out and fail gracefully. Related patterns: Exclude this activity, data swipe, disable intelligence

User Autonomy & Control

Machine teachingUser feedback loop Design your interface to support your data strategy. Your users' implicit & explicit feedback on the algorithm’s performance is what drives its learning and improvement. Developing user-friendly ways of collecting this data is crucial to your solution becoming truly helpful. Besides the feedback loop, the interface can be a channel to collect (crowdsource if you will) data and label unlabelled data (examples). What you give is what you get. Related patterns: Exclude this activity, data swipe, disable intelligence

User autonomy/customization:One size does not fit all Algorithms can (mis)behave in a way that isn’t logically incorrect but doesn’t align with the users’ needs. To keep them from spiraling down, give the user a way to intervene with the course of the algorithm. Instead of trying to account for each variable, or neglecting differences, allowing them to tune the algorithm, express intent and exercise autonomy over their own data, model, and experience based on their individual and contextual needs. How might we give the user controls to tune the algorithm to their needs?Related patterns: Exclude this activity, data swipe, disable intelligence

Data privacy:Always listening The abundance of personal data across context and channels is both what enables amazing experiences, and what poses growing privacy and security risks. As creators of these systems we must take the necessary precautions in how we collect, store, communicate, and protect our users’ (sensitive) data from misinterpretation, leaking, hacking, and misuse.  Related patterns: Exclude this activity, data swipe, disable intelligence

Value Alignment & Fairness

Algorithmic etiquetteWhile (most) humans intuitively grasp basic morality and etiquette, machines do not. This requires designers to translate implicit human values and user needs into explicit objectives, labels, parameters, and trade-offs algorithms and data scientists can understand and optimize for in their model. Mapping needs to models. Related patterns: Exclude this activity, data swipe, disable intelligence

Computational inclusivityOur data and minds are infested with historical, cultural and implicit bias. We must actively seek out and eradicate these bias and identify gaps in inclusivity to make sure algorithms are fair to all and not perpetuating harmful patterns of discrimination (mostly against those who aren’t white, male, Western, and/or able-bodied). History can’t write the future. Related patterns: Exclude this activity, data swipe, disable intelligence

Ethics & (un)intended consequencesConsidering the scale of impact, we must prioritize ethical and responsible implementation and be one step ahead in anticipating and designing for indirect & unintended consequences of systems on individuals & society such as job disparity, deepfakes, divided information streams, technological dependency, loss of agency, misuse, political targeting, and more. To help, not harm, humanity.  Related patterns: Exclude this activity, data swipe, disable intelligence

These challenges are discussed in more detail in the e-book user experience design challenges of AI-driven interactions published in collaboration with Adyen and Awwwards and available for download here.

Explainability

To build trust and understanding, communicate to your user how and why the system acts, generates, or concludes the output in a simplified and gradual way. Making sense of the machine. Related patterns: Exclude this activity, data swipe, disable intelligence

Managing expectations

From the beginning, set expectations about what the system can and can not do, and guide behavior in ways that match the system’s abilities and limitations. Construct helpful mental models to prevent disappointment. Related patterns: Exclude this activity, data swipe, disable intelligence

Assuming failure

Assume failureWith less control over the ultimate outcome, design your interface assuming failure to minimize user pain when the system produces faulty, imperfect, or even absurd results. Offer a way out and fail gracefully. Related patterns: Exclude this activity, data swipe, disable intelligence