Wei Xu, Intel Corporation

Toward
Human-Centered AI:
A Perspective
from Human-
Computer
Interaction

I
Insights In a 2018 survey of emerging trends and use are decentralized and global,
→→ The third wave of AI can be among 6,300 corporate IT executives the entry bar is relatively low, making
characterized by technological around the world, Accenture found it more difficult to control [1]. Machine
enhancement and application that the foremost concern was citizen learning (ML)-based AI systems trained
+ a human-centered approach. artificial intelligence (AI)—that AI with incomplete or distorted data (i.e.,
→→ HCI professionals should take research and development should their “worldview”) can lead to biased
a leading role by providing make AI-based solutions responsible “thinking,” which may in turn magnify
explainable and comprehensible and productive actors in society. This prejudice and inequality, spread rumors
AI, and useful and usable AI. means that as AI is developed further, and fake news, and even cause physical
→→ HCI professionals should attention should be given not only to harm. Because of these concerns,
proactively participate in AI the technology itself but also to other some large AI projects never came to
R&D to increase their influence, important nontechnical factors. fruition. As more and more enterprise
enhance their AI knowledge, As with other transformative and and government services based on
and integrate methods between controversial technologies like nuclear AI/ML algorithms are released,
the two fields. and biochemical, AI has both potential decisions influenced by AI thinking
benefits and risks. Since its development and worldviews will increasingly affect

42 INTER ACTIONS J U LY – A U G U S T 2 0 19 @INTER ACTIONSMAG

 people’s daily lives. Some even worry follow three objectives: to technically users, an ML-based intelligent system is
that in the future AI systems may reflect the depth characterized by a black box, especially neural networks
completely reject human beings, who human intelligence; to improve human for pattern recognition in deep learning.
will eventually lose control. capabilities rather than replace them; This black-box phenomenon causes
In response to these concerns, and to focus on AI’s impact on humans users to question the decisions from the
Stanford University, UC Berkeley, [1]. In the industry, several leading high- system: Why did you do this? Why is
and MIT have established human- tech companies have advocated for AI this the result? When did you succeed
centered AI (HAI) research institutes. solutions that are responsible, ethical, or fail? When can I trust you? This
Their HAI research strategies secure, and inclusive by publishing reflexive skepticism directly affects
emphasize that the next frontier guiding principles for AI technology users’ trust and decision-making
IMAGE BY CARLOS CA STILL A

of AI is not just technological but development, processes, tools, and efficiency, thus also affecting the
also humanistic and ethical: AI is to training. adoption of AI solutions. Because of
enhance humans rather than replace As the core technology of AI, ML the black-box effect, AI solutions are
them. For example, researchers from and its learning process are opaque, not explainable and comprehensible
Stanford University believe that AI and the output of AI-based decisions is to users. This phenomenon may occur
research and development should not intuitive. For many nontechnical across various types of AI applications,

INTER ACTIONS. ACM.ORG J U LY – A U G U S T 2 0 19 INTER ACTIONS 43

including financial and legal decisions,
medical diagnoses, industrial process
monitoring, security screening,
Ethically
employment recruitment, legal aligned
judgment, university admissions, smart design
homes, and autonomous vehicles.
In addition, some AI applications
were very expensive and failed due to
a lack of use value. People have now
reached a consensus on the bottleneck
effect of developing solutions, stressing
that AI solutions must have a clear
purpose. By providing useful AI, such HAI
solutions can match user needs and thus
gain acceptance and generate economic
benefits. From the perspective of Technology Human
enhancement factors
intuitively usable AI, there are still design
some challenges in HCI design. For
example, a recent user experience (UX)
test conducted for three top brands
of intelligent assistants with voice
interactions in the U.S. market shows
that the smart assistants failed in all Figure 1. An extended HAI framework.
main types of complex problem tasks
and succeeded only in some simple ones ethical AI design emphasizes the designed for other experts in the
[2]. In recent years, autonomous vehicles enhancement of human capabilities same field. Similarly, much current
have also been involved in multiple fatal rather than their replacement. It AI research focuses only on technical
accidents, partially due to HCI design requires HCI design to ensure that aspects, and therefore AI solutions are
issues. This shows the importance of human operators are able to quickly and facing similar problems. The proposed
HCI design for usable AI. effectively take over the control of an HAI framework is intended to promote
How can the HCI community intelligent system in an emergency, so a refocus to a user-centered approach in
respond to this and help deliver that fatal accidents such as the accidents a broader context, much like the rise of
a comprehensive HAI solution of autonomous cars mentioned above the user-centered design (UCD) practice
encompassing not just ethics and can be avoided. 30 years ago. Thus, a new version of
technology, but also explainable, UCD practice, HAI, has again fallen

L
comprehensible, useful, and usable AI? THIRD-WAVE AI AND on the shoulders of HCI professionals,
OPPORTUNITIES FOR HCI promising many great opportunities for

T
AN EXTENDED Looking back at the history of AI the HCI community.
HAI FRAMEWORK

E
development (Table 1), we may conclude
This article proposes an extended HAI that the first two AI waves failed not EXPLAINABLE AND
framework (Figure 1). The framework only because they lacked mature COMPREHENSIBLE AI
includes three main components: 1) technologies but also because they Explainable AI (XAI) enables users
ethically aligned design, which creates left human needs unsatisfied. AI is to understand the algorithm and
AI solutions that avoid discrimination, beginning to meet human needs and parameters used, which is intended to
maintain fairness and justice, and do provide a positive UX for a variety address the AI black-box problem. If a
not replace humans; 2) technology that of application scenarios in the third medical-diagnosis intelligent system
fully reflects human intelligence, which wave. It is also beginning to deliver predicts the likelihood of cancer based
further enhances AI technology to mature business models with useful on a patient’s personal data, XAI
reflect the depth characterized by AI. In addition, people began to should provide the reasons for the
human intelligence (more like human consider the ethics of AI, as well as AI’s prediction to the patient through a UI.
intelligence); and 3) human factors interpretability and comprehensibility. Previous research on XAI had mainly
design to ensure that AI solutions are These are all human aspects. Thus, the been done in two ways: visualization
explainable, comprehensible, useful, third AI wave can be characterized of ML processes and explainable ML
and usable. Human factors design is by technological enhancement and algorithms. However, these approaches
not fully considered in today’s HAI application + a human-centered may be biased in explaining how ML
research agenda [1]. The purpose approach. algorithms work and rely mainly on
of this framework is to promote a History seems to be repeating abstract visualization methods or
comprehensive approach, ultimately itself. When PCs were emerging in the statistical algorithms, which may
providing people with safe, efficient, 1980s, computer users were mainly further increase complexity. At
healthy, and satisfying HAI solutions. programmers, who considered only present, the most representative one
The HAI framework shows synergy technical factors and not wide usability is the DARPA XAI five-year research
across the three domains. For example, when designing products: Experts program [3]. The program organized

44 INTER ACTIONS J U LY – A U G U S T 2 0 19 @INTER ACTIONSMAG

13 universities and research institutes computational models that contribute designs, we do have challenges in
to approach XAI primarily by to comprehensible AI and accelerate delivering usable AI.
developing new or improved explainable the transition from theory to practice. The first challenge is to move
ML algorithms. The program also Finally, HCI professionals can drive beyond mere interaction. With the
investigates explanation UIs with rigorous user-involved behavioral addition of learning capabilities in
advanced HCI techniques (e.g., UI experimental methods to validate AI-based machine intelligence, human-
visualization, conversational UI) and proposed research protocols, which machine relationships have shifted
evaluates psychological explanation was overlooked in most of the previous from human-computer interaction
theories to assist XAI research. The research driven by AI professionals. to human-machine integration and

U
program is still midterm, and so far human-machine teaming [4]. The
no well-researched results have been USEFUL AND USABLE AI humans and machines are teammates
published, particularly research related Useful AI is defined as an AI solution and collaborative partners now. The
to HCI. that can provide the functions required dynamic cooperation between the
From an HCI perspective, there is to satisfy target users’ needs in the valid two cognitive agents with enhanced
no guarantee that the target users of an usage scenarios of their work and life. capability on the machine side (as
XAI system will be able to understand From a historical point of view, one of it learns over time) brings added
it. For example, the XAI version created the main reasons third-wave AI can complexity to the HCI design of
for data scientists is incomprehensible penetrate people’s work and life is that AI solutions. There are a series of
to most non-expert users. According it can now solve practical problems questions that require systematic
to UCD, a design needs to provide a with the right usage scenarios and HCI research. For example, dynamic
comprehensible AI that is based on UX, something not achieved in past functional allocation and task
target users’ needs and capabilities (e.g., waves. As HCI professionals working assignments between human and
knowledge level). The ultimate goal on AI solutions, we must ask ourselves machine, dynamic goal setting, and
of XAI should be to ensure that target how much we can contribute to the allocation of decision-making power
users can understand the outputs, thus foundation for useful AI. between the two over time.
helping them improve their decision- HCI professionals are good at One of the other challenges is
making efficiency. identifying usage scenarios based on that current HCI methods were
Since the research has so far been HCI methods such as ethnographic originally created for non-intelligent
dominated by AI/ML professionals, studies and contextual inquiries, and solutions. In our study, we suggested
it is time for HCI professionals to helping mine user needs, behavioral a series of enhanced HCI methods
proactively seek to make contributions patterns, and usage scenarios. For specifically for AI solutions [5]. For
to XAI and comprehensible AI. First, example, using AI and big data to model example, during the UI prototyping,
HCI professionals can provide effective real-time user behaviors, and digital instead of rushing to focus on visual
HCI design for the UIs of HAI solutions user personas to identify potential user and interactive design as usual, HCI
by adopting effective visualization needs and real-world usage scenarios. designers should consider the AI-
models, adaptive UI, and natural UI AI also needs to be usable. Usable AI first approach, carrying out dynamic
dialogue technologies. Also, while can be defined as an AI solution that functional allocation between human
there are many psychological theories is easy to learn and use via optimal and machine, prioritizing the use of
of explanations, they have not been UX created by effective HCI design. machine intelligence functions (e.g.,
fully considered in current research. Although the HCI community has smart search, real-time user behavior,
HCI professionals can take advantage mature processes and tools (e.g., UCD) contextual information, voice input) to
of their interdisciplinary approaches to specify user requirements, prototype reduce repetitive human activities and
to work with AI experts to build UI or UIs, and conduct UX tests to validate design more intuitive UIs to optimize

First Wave Second Wave Third Wave

(1950s–1970s) (1980s–1990s) (2006–)

Major advances Early symbolism and connectionism Statistical model in speech recognition Breakthroughs in applications of
in technologies school, production systems, and machine translation, artificial deep learning in speech recognition,
knowledge inference, preliminary neural network in pattern recognition, pattern recognition, big data, high-
expert systems expert systems performance computers

Human needs Not satisfied Not satisfied Starting to provide useful and real
problem-solving AI solutions

Focus Technological solutions Technological solutions Integrated solutions: ethical design,

technological enhancement, human
factors design

Characteristics Academia driven Academia driven Technological enhancement

and application + a human-centered
approach

Table 1. A comparison of the three waves of AI.

INTER ACTIONS. ACM.ORG J U LY – A U G U S T 2 0 19 INTER ACTIONS 45

the UX. This is the approach we have propose solutions by adopting social framework by providing explainable,
explored in our research project, with and behavioral science methods from comprehensible AI and useful,
promising results [6]. a broader sociotechnical systems usable AI. HCI professionals can also
HCI professionals can also play a perspective. contribute to ethical AI design and AI
critical role in validating AI solutions. The HCI community can also technological enhancement. The deep
The traditional software verification support the technology-enhancement involvement of the HCI community in
method assumes that the system has no component in the HAI framework. these areas has yet to be fully realized,
learning ability to change its behavior, HCI adopts the UCD process, but is necessary and urgent.
which is predictable. However, the defining human needs, designing In order to provide full disciplinary
behaviors of intelligent systems develop and prototyping solutions for these support for HAI solutions, the
over time. The verification evaluation needs, and then testing the design work of the HCI community should
of AI solutions needs collaboration with human users. HCI, which studies include research on human-machine
between AI software engineers and HCI the interaction between human and integration/teaming, UI modeling
professionals. A combination of methods machine (AI), can make contributions and HCI design, transference of
(e.g., software validation, user-involved to algorithm modeling, training, and psychological theories, enhancement
UX validation) may help achieve better testing. By following a human-centered of existing methods, and development
results. Early UX evaluation of low- ML approach as an example, AI and of HCI design standards. HCI
fidelity intelligent design prototypes HCI professionals can work together to professionals should proactively
requires alternatives such as Wizard of define UX criteria, test/optimize ML participate in AI research and
Oz (WOZ) design prototypes to simulate training data and algorithms iteratively, development to increase their influence,
and validate the learning and intelligent and avoid extreme algorithmic bias. enhance their AI knowledge, and
behaviors of AIs. Although history seems to be integrate methods between the two
Finally, current AI-related standards repeating itself in the initial stages fields to promote effective cooperation.
focus primarily on ethical design issues, of AI’s third wave, future success is
such as the guidelines published by much more likely if HCI professionals Endnotes
1. Li, F.F. and Etchemendy, J. A common goal
IEEE. There are currently no specific proactively embrace the challenges
for the brightest minds from Stanford and
HCI design standards for guiding AI and start to participate directly in the beyond: Putting humanity at the center of
solutions; the HCI community needs to research and development of AI today, AI. 2018; https://hai.stanford.edu/
develop these. just as the entire HCI community has news/introducing-stanfords-human-
been doing over the past 30 years. The centered-ai-initiative
ADDITIONAL HCI maturation of key technologies and the 2. Budiu, R. and Laubheimer, P. Intelligent

I A
CONSIDERATIONS OF accumulated experience of the field assistants have poor usability: A user
THE HAI FRAMEWORK make this all the more likely.
study of Alexa, Google Assistant, and
Siri. 2018; https://www.nngroup.com
In addition to the human-factors design 3. Gunning, D. Explainable artificial
component of the HAI framework, the KEY TAKEAWAYS intelligence (XAI) at DARPA. 2017;
HCI community can also contribute AND ACTIONS REQUIRED https://www.darpa.mil/attachments/
to other components. For example, in As we see, there are lessons to be XAIProgramUpdate.pdf
the case of ethically aligned design, AI learned and actions to be taken. The 4. Farooq, U. and Grudin, J. Human-
computer integration. ACM Interactions
engineers typically lack formal training third wave of AI is characterized
23, 6 (2016), 27–32.
in applying ethics to design in their by technological enhancement and 5. Xu, W. User-centered design (III):
engineering courses and tend to view application + a human-centered Methods for user experience and
ethical decision making as another approach, which provides a great innovative design in the intelligent era.
form of technical problem solving. opportunity for the HCI community to Chinese Journal of Applied Psycholog y 25,
Fortunately, the AI community now provide comprehensive HAI solutions 1 (2019), 3–17.
recognizes that ethical AI design to address emergent challenges. This 6. Xu, W., Furie, D., Mahabhaleshwar,
M., Suresh, B., and Chouhan, H.
requires wisdom and cooperation requires a systematic consideration of
Applications of an interaction, process,
from a multidisciplinary field ethically aligned design, technology that integration, and intelligence (IPII) design
extending beyond computer science fully reflects human intelligence, and approach for ergonomics solutions.
[1]. The HCI community can leverage human factors design. Specifically, HCI Ergonomics 62, 7 (2019), 954–980;
their interdisciplinary skills to assess professionals should take a leading role https://doi.org/10.1080/00140139.201
the ethics-related issues and help in the human factors design in the HAI 9.1588996

Wei Xu is a researcher at Intel. He is chair

of the Intel IT Cross-Domain HCI/UX Technical
Working Group, leading HCI/UX design
strategy, standards, and governance. He has a
It is time for HCI professionals to Ph.D. in psychology (HCI focused) and an M.S.

proactively seek to make contributions

in computer science from Miami University.
His research interests include HCI, cognitive

to XAI and comprehensible AI.

engineering, and aviation human factors.
→→ [email protected].

DOI: 10.1145/3328485  © 2019 ACM 1072-5520/19/07 $15.00

46 INTER ACTIONS J U LY – A U G U S T 2 0 19 @INTER ACTIONSMAG