UX design in AI

The current face of AI

AI is an umbrella term encompassing many different methodologies and concepts, referring to any machine developed to perform tasks that would require intelligence if done by a human. Although the media commonly portrays AI capabilities as superior to human capabilities – i.e. as an artificial super intelligence (ASI) – the truth is quite different. Since the earliest explorations into the AI field, the scientists and practitioners have sought to create a computer with a level of intelligence similar to a human. Known as an artificial general intelligence (AGI) – these would be machines with a reasonable degree of self-understanding and autonomous self-control, able to solve a variety of complex problems in a variety of contexts. Despite the huge advancements of AI, especially in the last decade, we are still far away from being able to create an AGI, let alone an ASI.

The current form of AI we are working with is known as an Artificial Narrow Intelligence (ANI), or “weak AI”. ANI systems are created to carry out specific tasks showing specific aspects of intelligence in a specific context. All current applications of AI, whether it is an autonomous car, a chatting app camera filter, or an intelligent OSS, are all considered narrow or “weak” by this definition.

An easier way to describe the role of current AI applications is to call them “agentive technology” - whereby we can think of them as our assistants or agents, handling a discreet task and not the entire job.

In this current context, humans still need to have a view of the bigger picture and are still required to supervise, evaluate, and orchestrate the work of these AI systems.

Figure 1: Levels of AI

Trust as a vital component in AI adoption

There is an increasing trend of digital assistants appearing in many different aspects of our lives. Powered by machine learning (ML) models, they analyze data to come up with statistical probabilities that can be used to offer recommendations and make predictions and decisions, from suggesting the optimum route to take on a commute, to adjudicating whether we are viable for a loan or not.

Although they sound less impressive than the idea of the ASI’s superior artificial brain, these recommendations, predictions, and decisions taken by the AI systems can be considered a fundamental change in the way humans are using tools – a paradigm shift in the human-tool relationship. Since the beginning of this relationship, humans have always been in full control not only of what the tool should do, but also exactly how it will work, at least in the design and creation phase. The progression to the current status of AI is an evolution of this relationship in two ways.

First, it is an upgrade of the tool’s status from the role of a “slave” to that of “agent,” giving it agency by having a degree of autonomy with regards to “what” it should be doing. And second, it is a change in that with AI, we no longer entirely decide “how” the tool executes its function. In fact, in many cases, the creators of an AI system cannot entirely describe the criteria that the ML model has used to reach the output. This is known as the “Black Box problem”.

Taking these points into consideration, the following can be said about the current state of AI systems:

1. Rather than just executing what the human user wants, AI systems will autonomously come up with predictions, recommendations and decisions.
2. We are not always able to fully understand or explain why an AI/ML system has reached its output.
3. AI/ML output is based on statistical probabilities just like human decision-making – it judges low or high probability of outcomes, it’s not some kind of ultimate truth or absolute objective correctness.

We can therefore reach the conclusion that a degree of trust is needed, before the user can hand responsibility over to the AI - and give the autonomous car the steering wheel.

The requirements in building this trust-based relationship varies according to the specific task the AI is supposed to handle. Accepting an AI’s recommendation on which movie to watch is much “easier” with a lower trust threshold than, for example, the recommendation on which medicine a doctor should prescribe to their patient.

In a recent survey asking owners of smart voice assistant devices to list the tasks they perform using the device, 84.9 percent reported they use it to set a timer, while only 3.5 percent reported using it to call a cab.

A recent study has shown that when it comes to the application of AI in a business context, 94 percent of business executives understand that AI is essential to business strategy, however a separate study by MIT Solan found that only 18 percent of companies are widely adopting and understanding AI. In designing an OSS AI solution that takes critical decisions affecting the performance of an entire network, we discovered that the success of the system depended on more than building more efficient and accurate models and algorithms. We came to the realization that trust is an essential factor in the human-AI interaction, and if we want the users of our AI solutions to accept handing over more critical tasks and decisions, we need to design them to be trustworthy.

The four components of trust in human relationships

Although the concept of trust in human-AI relationships is a new field, we don’t have to work from scratch.

Our approach at the Experience Design team in Ericsson was human-centric. Humans have been trusting other humans since the beginning of our existence, and once human-human trust relationships were established, we started to putting our trust in entities and organizations; religions, political parties, banks, schools, business and so on. Our approach was to draw on the formula of trust already functioning in these human-human and human-organization relationships, and use it as a base for building trust in human-AI interaction.

The four main factors that contribute to building trust in another person or entity are: competence, benevolence and openness, integrity and charisma. A good example to illustrate these four components in action, is the process of decision making when hiring an employee that will be responsible for a task in an office. When dealing with “digital” assistants in the form of AI systems, exactly the same framework of trust applies.

The following pages examine each of these components in turn, from the perspective of human-AI interaction, along with relevant examples of design-related decision and focus areas that can contribute towards creating a trustworthy AI experience.

Other factors that can affect trust in AI

All the elements that are mentioned so far in this framework are those that are represented in one way or another in the interaction and the interface of the AI system, and are therefore relevant to the UX design.

There are numerous other factors that can affect trust in an AI system, and although these additional factors cannot be translated into the interface as UX, they are nevertheless essential elements in the building of a trust-based relationship.

For example:

1. Accuracy
   If the output results, predictions, recommendations and decisions of the system are not accurate to begin with, then the system will not meet the aforementioned criteria needed to satisfy the "competence" component.

2. Bias in AI
   One widely discussed topic in the AI community is the concept of bias - specifically that ML selects incomplete, uninclusive or biased data sets to train the model, whether deliberately or otherwise. This will result in output that is not fair and biased towards a group of users - meaning that the core pillar of "integrity" will not be met.

3. Laws and ethics
   Without sufficient and clear laws and a code of ethics that regulates the relationship between the user and the AI system, for example defining who is responsible if the output of the system affects the user in a negative way, then the trust in the human-AI relationship will not survive any potential mistakes the system makes.

The future of human-tool relationships

When introducing AI-powered software like Ericsson's Cognitive Operation Support System (OSS) services, the notions of what a network is, what the owner or network operator’s role is and what a system provider contributes
are changing.

The interaction will be on a much higher and abstract level. Instead of changing gears in the car, the focus will be on the passenger's journey. Instead of having field technicians manually climbing towers to fine-tune the radio, business operators will collaborate with the AI machine to reach the organization’s business intent, impacting the roles of the system providers and the network operator.

A user-centric design process will be even more important when designing AI-powered services than in traditional services. If users and organizations are going to trust the AI-powered system, for example an airplane without a pilot, the trust must be designed into the system and the relationship from the very beginning. Without it, these services will fail.

Designing an OSS AI solution that takes critical decisions that can affect the performance of an entire network is about more than focusing on building better AI models and algorithms. Trust will be the most vital factor in human-AI interaction.
If we want the users of our AI solutions to accept handing over more critical tasks and decisions to AI, we need to design them to be trustworthy.

Mikael Eriksson Björling

Mikael is Experience Design Line Manager at Ericsson Experience Design Lab and an Ericsson Evangelist. He was previously Director at the Networked Society Lab. His specialty is in understanding how new behavior, emerging technologies and new industry logics are shaping the future society and in the intersection of these areas design great user and customer experiences. Mikael believes that with the ongoing digital transformation we have a great opportunity to shape a better world. Mikael joined Ericsson in 1998.

Ahmed H. Ali

Ahmed is visual and user experience designer at Ericsson Experience Design Lab. With over 15 years of experience, his career inside and outside Ericsson was focused on designing digital systems that satisfy users’ needs and help them achieve their goals by bringing design thinking to the product development process, applying human-computer interaction best practices, and delivering UI/UX concepts and insights. Ahmed joined Ericsson in 2018 and he holds an M.A. in visual design from the University of Hertfordshire in the UK.

• Minsky, M. (1982). Semantic information processing.
• Goertzel, B. (2007). Artificial general intelligence (Vol. 2). C. Pennachin (Ed.). New York: Springer.
• Gobble, M. M. (2019). The Road to Artificial General Intelligence.
• Noessel, C. (2017). Designing agentive technology: AI that works for people. Rosenfeld Media.
• Bathaee, Y. (2017). The artificial intelligence black box and the failure of intent and causation. Harv. JL & Tech., 31, 889.
• Andras, P., Esterle, L., Guckert, M., Han, T. A., Lewis, P. R., Milanovic, K., & Urquhart, N. (2018). Trusting intelligent machines: Deepening trust within socio - technical systems. IEEE Technology and Society Magazine, 37(4), 76-83.
• Muir, B. M. (1987). Trust between humans and machines, and the design of decision aids. International journal of man-machine studies, 27(5-6), 527-539.
• Siau, K., & Wang, W. (2018). Building trust in artificial intelligence, machine learning, and robotics. Cutter Business Technology Journal, 31(2), 47-53.
• Commerce is a conversation: Survey on Amazon Echo and Voice Assistants - Experian Insights, 2016.
• Intelligent Economies: AI’s transformation of industries and society -The Economist Intelligence Unit, and Microsoft, 2018.
• Ransbotham, S. et all. Artificial Intelligence In Business Gets Real - Pioneering Companies Aim for AI at Scale. MIT Solan, 2018.
• Stern, M. J., & Coleman, K. J. (2015). The multidimensionality of trust: Applications in collaborative natural resource management. Society & Natural Resources, 28(2), 117-132.
• Sanders, K., Schyns, B., Dietz, G., & Den Hartog, D. N. (2006). Measuring trust inside organisations. Personnel review.
• Schoorman, F. D., Mayer, R. C., & Davis, J. H. (2007). An integrative model of organizational trust: Past, present, and future.