Designing AI Experiences p.3: Guiding principles

Read the first article “What can AI do for humans?” and the second “New design grammar”

When Roberta Tassi and me started to think of the principles that can help guide AI experience design, we drew inspiration from Don Norman’s human-computer interaction principles and Elena Pacenti’s Service Design guidelines developed during her PhD at the University Polytechnic of Milan.

These elements have been fundamental guidelines for our work. For example, the principle of transparency has become a must-have for almost every service provider, and it’s essential for putting users in a position where they can understand their choices. But what does that mean in the new context of AI?

We began exploring these principles and identifying the most crucial variables that impact the design of AI-based services and experiences.

Some variables are functional, defining human interaction with the technological components and underlying service. Other variables are more related to aesthetic values. We focused on understanding the variables that prompt meaningful questions, leading to interesting designs.

The first group of principles to consider generally relates to affordances and how they change in AI-based services:

1. Visibility: This is particularly important for zero-interaction/automated services. We need to design “negative affordances” to signal to users that some of their old interaction patterns can be abandoned in the new service and indicate what the new pattern is, even if we want it to be intuitive and invisible. This helps avoid confusion similar to that caused by automatic faucets, where traditional interaction is eliminated.
2. Mental modelling: We must manage the mental model that users develop for automated services such as Amazon Go, where we ask them to abandon the old patterns of service use. Users must learn to check in at the beginning of the journey and not check out, which implies a new approach to the journey. We must consider the learning curve involved in this process and all the communication needed for it.
3. Transparency: This is perhaps the most challenging principle, given the technical complexity and black-box issues in machine learning. Traditionally, transparency has been about making clear expectations or letting users know what is happening while they wait (as in parcel tracking). However, we must now consider two levels of transparency: Firstly, we need to render technical constraints more clearly to users. For example, users need to understand that Echo devices can either speak or listen, and if they start speaking while Alexa is still talking, their inquiry may not be correctly interpreted. Secondly, we need to track the actions that AI might take on a user’s behalf and make it clear which specific user inputs or AI interpretations of their actions led to that action. (For example, if my AI assistant constantly buys me chocolate, I want to know whether it’s based on my habits, its interpretation of my mental state, or promotions pushed by the vendor. This is a case for the application of so-called Interpretable AI.)
4. Explicability: The emergence of AI leads to an important new principle. This involves both the user’s ability to understand what is happening in the system and accountability for the actions that AI takes. It is connected to the idea of “letting the user look under the hood” to become acquainted with the system’s functioning. There are attempts now to work in the direction of Explainable AI (XAI) that should respect this principle.

The second group of principles pertains to different aspects of how the user should keep control of the system:

1. Freedom to exit: In traditional systems, this refers to the capability to stop the interaction at any moment, like getting my card back from an ATM if I change my mind. However, with the emergence of more complex systems like Google or Amazon, it becomes crucial to understand what the user’s freedom to exit really means, especially regarding how to retrieve their data securely.
2. Error-friendliness: While in traditional systems error-friendliness means allowing human operators to make mistakes without causing significant problems, in the case of smart agents, we might need to design systems that ensure users can accept some level of system failures and determine how accountability will be divided between the user and AI agent. This principle has become especially problematic recently with multiple accidents involving auto-pilots in self-driving cars.
3. Staying in the loop: It is crucial to consider how to keep the user informed and empowered to intervene and redirect the system and design systems that provide clear and timely feedback enabling that. The user must understand when they might need to retake control, and conditions should be created for them to effectively perform it (e.g., in the case of self-driving cars, users often get bored and distracted as the machine is in control). This is important not only for controlling AI but also for humans to maintain their level of skills and knowledge to still be capable of performing actions manually in case of technological failure.

The third group of principles is related to adaptability and how these services react to the situation:

1. Learning: AI services that learn from user feedback and improve over time are a natural outcome of the technology. They allow for personalized experiences that adapt to the user’s needs and preferences. This feature is a key aspect of many successful AI-based services.
2. Flexibility: a crucial principle in designing AI services, as it allows the system to adapt and evolve over time. This includes the ability to incorporate new data and user feedback, as well as to respond to changing user needs and preferences. A flexible AI system can also better handle unexpected situations and exceptions, providing a more seamless and intuitive user experience. Overall, designing for flexibility enables AI services to continuously improve and stay relevant in an ever-changing landscape.
3. Multimodality: This principle involves providing multiple modes of interaction for the user, such as voice, text, and gesture recognition. It also includes combining different modes of interaction for a more seamless experience. By offering multiple modes of interaction, AI services can better cater to users’ preferences and accommodate different contexts of use, making the overall experience more flexible and accessible.
4. Context Awareness: This is a crucial new principle for designing effective AI services that are adaptable to different situations. For example, users may not want to be interrupted while working or when engaged in other focused activities. AI services should be designed to detect and respond to these contexts, providing different experiences that adapt to specific situations rather than a one-size-fits-all approach. The challenge lies in creating services that can recognize and adapt to the user’s context, providing a seamless and personalized experience.

Going towards more aesthetic-related principles we can think of elements that improve the experiences.

Consistency:

The fourth group of principles is about more aesthetical and emotional aspects of services, connected also to brand experience and perception.

1. Shared and consistent language: should happen both in terms of language that is shared between the service and user and the language fitting the context (not an AI that would speak bureaucratic or technical language that a person won’t understand). The service language should also be shared and consistent across various platforms and contexts in which users access them.
2. Shared values: This is fundamental for the match between user and service. In part, we saw an example of this before in how current maps apps value efficiency and speed, while the user might value the experience of the route. We need to find ways to communicate those values or adapt to the user.
3. Accessibility: AI services might seem to have fewer accessibility issues, as natural language interfaces might overcome some limitations of operating buttons and touchscreens. However, we already see a significant difference in accessibility of services in different languages. Another aspect is the need for ever-increasing bandwidth and processing power, which might pose a new level of digital divide between those who can afford high-quality AI services and those who cannot.
4. Ethics: a broad aspect which goes beyond a single principle and is dicussed now in context of both technology and design. Designers should consider embedding ethical codes into the functioning of their AI services and evaluate potential ethical implications of the technology. They must be mindful of the potential impact of their services on society and ensure that they align with ethical values and principles. However, these ethical aspects might go, in some cases, against consistency principle as some ethical concerns might be very much context-based and should consider diversity and differences.

The fifth group is about helping people feel at ease in the situation and respected for creating comfortable AI experiences.

1. Appearance, character, and atmosphere: These elements are used to convey the service’s essence through touchpoints like brand identity visualization, colors, and sounds. With AI-based experiences, that can often go beyond screen interactions, it is important to understand how the brand values are still transmitted and what other aspects like tone of voice or character can be used to develop different and unique experiences.
2. Personalization: this is one of the biggest benefits of bringing intelligence into services, but it raises questions about how far it can go. Traditionally, services and brands have been created with specific values and target audiences in mind, but personalization can challenge this model, adapting the service to each specific user’s preferences. For example, will there still be a need for multiple taxi/ride services, or will a unified, integrated platform offer enough personalization to cater to everyone’s individual needs? Will the idea of a brand and its values be replaced by each user’s values and needs?
3. Discreteness: While it is beneficial for AI to utilize data about a user’s behavior to personalize their experience, it is important to do so discreetly, without making the user feel embarrassed or uncomfortable. AI should integrate with the user’s preferred way of doing things without creating awkward situations. For instance, if an AI collects a lot of data about a user, it should use that information to enhance the user’s experience while still respecting their privacy. Designers should consider how to include elements of surprise and natural interactions while using the data to back up the service’s actions, without making the user feel uncomfortable or invaded.

In conclusion, the design principles for AI experiences are complex and multifaceted, combining elements from traditional service design, ethics, and aesthetics. As AI becomes more integrated into our daily lives, designers must consider these principles to create services that are not only functional and efficient, but also enjoyable and respectful of users’ values and preferences.