Ergonomics in AI: Designing and Interacting With Machine Learning and AI

Machine learning and artificial intelligence (AI) enable new types of autonomous systems that are changing our personal and professional lives. While there are plenty of stories about machine learning delivering the promise of a better future, such as autonomous vehicles for improving mobility, safety, and fuel efficiency, many examples have indicated great risks, such as bots on social media for spreading false information and manipulating public opinions. As machine learning approaches ubiquity in industrial systems and consumer products, Human Factors must innovate to support users in coping with emerging autonomous capabilities. The rise of machine learning technologies poses serious questions that have been discussed in the panels at the recent Annual Meetings of the Human Factors Ergonomics Society (Lau et al., 2018; Lau et al., 2019). How can we help users understand the autonomous capabilities developed through supervised or unsupervised learning? What kind of interactions could enhance cooperation between human and machine learning algorithms? We must also take advantage of machine learning techniques in advancing our own research and design science. How can we use machine learning in assessing human states and capabilities? How should we help incorporate human sensing into machine learning algorithms? In this special issue, we embrace the broad spectrum of research and design efforts that investigate machine learning for improving usability and safety of intelligent systems and consumer products. Our goal is to clarify the roles of Human Factors in contributing to a humanist perspective that considers the social, political, ethical and cultural factors of implementing AI into daily human–system interactions. At the same time, this special issue can only accommodate five articles, about one third of the submissions, after a rigorous peer-review process. So, what we have hoped to curate for the readers is an intellectually stimulating, short exhibit of our discipline in developing and applying next-generation AI. The first article in this special issue is a commentary by Hancock who envisions work to be eventually shared between self-evolving machines and humans. This vision of work challenges the Human Factors community to prepare for a future that requires designing interactions and user interfaces for machines whose behaviors we cannot fully anticipate and for work that we do not yet know. The second article by Zhang et al. speaks to the challenge of anticipating the consequences of machine learning in designing technology and making policy decisions. The authors use a speech recognition example to illustrate a violation of inclusivity in design. In their second example, they illustrate how a loan policy aimed at supporting a disadvantaged group ultimately harms the group in the long run. These examples raise questions about how Human Factors professionals can engage in a data-driven design process and how to develop “explainable” AI that presents the true behavior of the machine learning model to the user. Anthropomorphism is a much talkedabout concept to help design AI that humans can understand and interact with intuitively. Muller compares how deep neural networks and humans classify images to illustrate how their differences likely require appropriate interactivity to minimize the mismatch between how human and AI think about the intelligence of each other. The article highlights not only the need for interaction design but also the importance of understanding machine learning algorithms. The final two articles present empirical investigations on how ergonomic research can promote the appropriate use of machine learning tools. Wang et al. describe metrics of stability, robustness and sensitivity for aiding users to interpret prediction results of supervised learning algorithms. The authors illustrate how effective visualization of those metrics can improve decision making. Gilbank et al. describe a qualitative study with ten medical professionals using a machine learning–driven toxicity prediction tool that utilizes 10 years of historical data. The study presents the expectations and perspectives of medical professionals on AI, and the user interface design considerations for promoting trust and, ultimately use of the machine learning system. We hope that these five articles contribute to new thoughts and present exciting challenges. At the same time, we recognize that these articles only represent a speck of all the relevant Human Factors research and design issues related to machine learning and artificial intelligence. As we prepared this special issue, we realized that so much progress is needed to formulate design methods for machine learning. So, we agree with Hancock’s conclusion that our road ahead for ergonomics design of machine learning and artificial intelligence “promises to be a bumpy but exciting ride.”