Handbook of Human-Centered Artificial Intelligence
摘要
Explore the transformative world of Human-Centered Artificial Intelligence (HCAI) with the Handbook of Human-Centered Artificial Intelligence. This authoritative resource bridges cutting-edge AI technologies with core human-centered principles, such as human needs, values, controllability, and experiences, offering a holistic approach for designing, developing, deploying, and using AI systems that put humans first.
Written by leading experts, the handbook serves as an indispensable reference for researchers, practitioners, and policymakers dedicated to creating AI solutions that are reliable, safe, controllable, usable, valuable, and trustworthy. It covers the foundational principles of HCAI, from core concepts, methodology, design and development strategies, risk management and governance, to real-world applications across diverse domains, demonstrating how AI can enhance human capabilities and improve the quality of work and life.
Empower your knowledge and be at the forefront of AI innovation, a future where AI benefits humans rather than harming or replacing them. This is your definitive guide to understanding and shaping the future of AI through a human-centered lens.
This handbook comprises ten sections, encompassing approximately sixty chapters in total. (Note: In the current living version, individual chapters are published in alphabetical order as they become available.)
• Fundamentals
• Human-AI interaction
• Human-centered AI design
• Human-centered AI in computing
• Human controllability over AI
• Human-centered AI risk management
• Human-centered AI governance
• Human-centered large language models
• Sectoral applications of human-centered AI
• Human-centered AI practice
Foreword
The emergence of new disciplines generates excitement that sets my mind spinning. I am drawn to novelty, energized by curiosity, and thrilled by vigorous controversies among respected peers, even those who passionately disagree. Such bubbling cauldrons are necessary to forge a solid foundation that is the basis for future growth.
I’ve witnessed and participated in the birth of new disciplines such as human-computer interaction, information visualization, and computer science itself. It’s brought me great satisfaction to see how these disciplines grew with thriving conferences, vigorous journals, and fruitful academic programs to forge successful commercial products and entire industries with trillion dollar valuations.
Now the early efforts on Human-Centered AI (HCAI) are going through the tumultuous processes that will spawn similar results. This handbook’s editor Wei Xu and the hundreds of contributors, have put in a remarkable effort to catalyze the next stage of HCAI work. The 59 papers present a broad spectrum of diverse topics and viewpoints, dealing with pressing issues of taming Generative AI. The contributors come from an international community of young and senior scholars, who adhere to a uniform chapter structure and reference style. The thousands of references, including many recent ones and vital classics, enrich the discussion with their many frameworks and pointers to resources for further reading.
This collection sometimes has several chapters that take differing perspectives on one subtopic. This healthy overlap gives readers the opportunity to understand the different points of view, which they can synthesize into their own principles to guide design, implementation, evaluation, and continuous refinement of research directions and commercial products that embed Generative AI and its variations: “[HCAI] is now expanding from fundamental concepts and design philosophy toward practical methodologies, governance, and interdisciplinary collaboration” (Chapter 1, HCAI Foundations and Approaches, Wei Xu).
While excitement about the benefits of HCAI is warranted, diligence is needed to limit the harm from malicious actors, such as cybercriminals, terrorists, and political oppressors. Designs can prevent some harms, but legislation, regulation, and careful monitoring are also needed.
The differing perspectives are a natural part of an emerging discipline, enabling readers to identify research topics for their future work or as suggestions for doctoral students. I hope readers will appreciate this collection as a springboard for their own efforts and for their students.
Other indicators of an emerging discipline are the appearance of new workshops, conferences, journal special issues, and new journals. These have already begun to appear, as have podcasts, seminar series, courses, and entire degree programs. Google Scholar tracks the growth of citations to HCAI, which have soared in the past five years. New research labs, often spanning existing disciplines, sprout on campuses, along with independent institutes, non-profit groups, and government agencies. The bubbling energy from the participants exudes a very human essence, reminding us that people are the source of new ideas and technologies.
It's a wonderful beginning, but the best is yet to come!
What is Human-Centered AI?
For many years, I have sought to describe Human-Centered AI (HCAI), sometimes presenting the goals as a bumper sticker “Amplify, Augment, Empower, Enhance Human Performance.” I’ve used article titles, such as “Ensuring Human Control, While Increasing the Level of Automation” and suggested principles, such as “Supporting Self-Efficacy, Enabling Creativity, Clarifying Responsibility, and Facilitating Social Connectedness.”
The chapter authors are well-aligned with my descriptions, offering their versions such as: “AI systems should not simply mimic humans but should be designed and implemented to empower them…True empowerment of humans can only be achieved when, alongside the increase in automation, individuals also retain control over it, so that qualities like intelligence, creativity, self-confidence, and others are preserved and strengthened.” (Chapter 11, Human-Robot Interaction, Alireza Mortezapour and Giuliana Vitiello)
Beyond the goals, the authors present extensive details about the human-centered processes that are needed. There are chapters on critical subtopics and reports from important application domains, such as healthcare and business. Embedded in the foundational values, which are discussed, is an implicit respect for distinct human qualities and capabilities. I’ve found that making these qualities explicit, accelerates designs that truly serve human needs. Abraham Maslow’s Hierarchy of Human Needs, first formulated in the 1940s needs an update that addresses the profound cultural evolution that has increased human capabilities and raised expectations. The discussions in these chapters helps us see a hope-filled future with still greater possibilities for human creativity, discovery, and invention.
A few themes that I would like to comment on are below.
Human responsibility, legally and morally
While many companies talk about their responsible AI practices, I would like more discussion of human and organizational responsibility for the impacts of the AI systems they develop. Only humans and their organizations are legally and morally responsible for these impacts. Just as medical providers, construction businesses, and transportation companies are responsible for failures, so too are the individuals and organizations that develop and use AI systems. Of course malicious developers and users who break laws must also be held responsible. I was pleased to see this quote: “Accountability is the final layer of ethical AI, ensuring that organizations remain responsible for their systems’ impacts… The future of human-centered AI depends on balancing technological innovation with ethical responsibility” (Chapter 14, Design Thinking and AI, Martin Bockle, Ulrich Mohme, and Markus Bick).
Autonomous agents
The terms “autonomous” and “autonomous agent” always struck me as misleading goals for designers. It suggests that the machine is an independent entity, capable of doing what humans do for long periods of time, with the added sparkle of perfection. I prefer to use the milder sounding “automated system” which conveys the image of something that needs to be started, monitored, and improved. Yes, the goal is to provide a high degree of automation, while at the same time providing a high level of human control. My position is in the minority for now, but I have found it helps me think more clearly about design features, such as detailed log or audit trails, which permit examination of failures and near misses.
I was gratified to find that chapter authors, even those who discuss autonomous systems believed “that human users [should] retain the ability to understand, direct, and intervene in the behavior of AI systems” (Chapter 28, Human-in/on-the-Loop AI, Leonie Bensch et al.). Another step forward was the shift from discussions of self-driving to remotely-operated cars. I also encourage a shift from self-driving to safe-driving cars, so as to put the emphasis on reducing deadly car crashes.
Discussion of the Mars Rovers shows the split between AI enthusiasts and the human factors engineers. The AI enthusiasts emphasize that Mars Rovers traverse the surface autonomously, avoiding obstacles and precipices, and adjusting their antenna and solar panels. However, the human factors engineers remind us that there are dozens of NASA ground controllers at work deciding on where the Rovers should go, dealing with equipment failures, and designing features for future Rovers.
Explainability
Developers of AI systems have produced more than 100,000 papers in the past decade seeking to explain what their systems were doing. These post-hoc or retrospective approaches using natural language have not succeeded: “Despite significant research efforts to advance the evaluation of explainability methods, the field remains fragmented, lacking widely accepted norms, standards, and universal definitions…Much of the current work also overlooks the critical need to align explainability methods with stakeholders’ requirements” (Chapter 46, LLM Explainability Ine`s Arous, et al.).
I’ve come to believe that the retrospective approach will not work, but the prospective user interface approach might succeed, as it has in many existing systems. Users make decisions in a step-by-step manner so that they are not surprised or confused by the result. This approach seeks to prevent the disease of confusion, rather than cure it. Visual interfaces such as Amazon’s shopping and checkout processes are good models to follow.
Collaboration
Another virtue of this collection is the reduced focus on antedated ideas of computers as teammates, partners, collaborators, or companions, with increased attention to people working in teams, collaborating within organizations, and developing their communities. This is a natural aspect of socio-technical systems thinking, which is the subject of several chapters: “Humans usually do not work in isolation but together with others—mostly in teams” (Chapter 58, Organizational Practices and Socio-Technical Design of Human-Centered AI, Thomas Herrmann).
In spite of substantial progress, technology design to support collaboration is arguably the next major growth area in computing, providing tools for ideation, deliberation, consensus building, conflict resolution, and collective action. These tools will need to be tailored to teams of 2, 2 thousand, 2 million, and 2 billion people. Ambitious goals include changing diet and exercise of entire populations, gaining agreement on national economic practices, and reforming lifestyles in response to climate change.
Conclusion
As I read the chapters, I felt invigorated and energized by the ideas, recommendations, and design guidelines. The new frameworks will shape the work of future researchers and invite developers to apply the clear guidance based on empirical evidence from academic studies and large-scale commercial implementations. Furthermore, policymakers in companies and governments can benefit from the risk management frameworks and other topics.
I was pleased to find an in-depth chapter on sustainability that guides analysts in the assessing plans for AI data centers and related infrastructure: “Ensuring that AI meets the high expectations of sustainability and is monitored, tracked, and audited for long-lasting positive effects will demand the following: (a) Rigorous documentation and demonstration of actual environmental benefits of AI implementation for sustainability and (b) Creation of certification or verification programs that include monitoring and auditing of sustainability claims” (Sinistore et al., Chapter 54, Sustainable AI: An Environmental Sustainability Perspective).
While the Generative AI technologies are startlingly impressive, they are also alarmingly flawed, so that applications in life-critical applications need to be carefully tested, certified, and monitored to reduce risks and improve safety.
As researchers and developers apply direct manipulation designs that provide compact control panels for efficient action, they will bring the benefits of Generative AI to wider audiences that include old and young, diverse cultures and gender identities, differing levels of literacy, and a wide range of users with disabilities. By these efforts, which respect human dignity, promote equality, and advance social justice, I believe that we can build a better world for our communities and our children, while improving the environment that we share.
Ben Shneiderman, Ph.D.
Professor of Computer Science, University of Maryland, USA
Member, National Academy of Engineering
Endorsements
This handbook’s editor Wei Xu and the hundreds of contributors, have put in a remarkable effort to catalyze the next stage of HCAI (human-centered AI) work…As I read the chapters, I felt invigorated and energized by the ideas, recommendations, and design guidelines. The new frameworks will shape the work of future researchers and invite developers to apply the clear guidance based on empirical evidence from academic studies and large-scale commercial implementations. Furthermore, policymakers in companies and governments can benefit from the risk management frameworks and other topics.
Ben Shneiderman, Ph.D.
Professor of Computer Science, University of Maryland, USA
Member, National Academy of Engineering
Wei Xu's monumental volume on human-centered artificial intelligence has to be a requirement for all libraries. Its fundamental human-centered knowledge will not age and will provide useful information to researchers and practitioners for years to come.
Gavriel Salvendy, Ph.D.
University Distinguished Professor, University of Central Florida, USA
Member, National Academy of Engineering
The Handbook of Human-Centered Artificial Intelligence provides an amazingly comprehensive and in-depth review of the existing and emerging thinking and practices in AI from both human and computing perspectives. Researchers, system developers, and policy makers will find this resource invaluable.
William B. Rouse, Ph.D.
Professor, McCourt School of Public Policy, Georgetown University, USA
Member, National Academy of Engineering
Human-centered AI applies frameworks, methods, and concepts from HCI to new developments in AI, with a focus on Generative AI. This timely and important handbook provides an extensive set of chapters covering the breadth and depth of this growing sub-field.
Yvonne Rogers, Ph.D.
Professor of Human-Computer Interaction, University College London, UK
International member, the US National Academy of Engineering; ACM Fellow
A comprehensive overview of the current state of thinking as well as the issues and methodologies associated with the design, development, application, and governance of human-AI symbiotic technologies. It provides an excellent resource to researchers, practitioners, and users in a variety of domains to grapple with the demands of responsible and trustworthy AI-enabled socio-technical systems. Should be on the desk of anyone interested in a systems design approach evolved from technology-centred to human-centred, and then interaction-centred towards a collaborative human-AI partnership.
Ming Hou, Ph.D.
Principal Scientist, Defence Research and Development Canada
Fellow, Canadian Academy of Engineering and IEEE
This book provides a foundation for rethinking the trajectory of AI by placing human values, agency, and well-being at the core of AI progress. It convincingly demonstrates that the long-term success of AI will not be defined by computational performance alone, but by its capacity to augment human capabilities, preserve meaningful control, and earn societal trust. This is an essential reference for researchers, practitioners, and policymakers who seek to shape a future of AI that is not only powerful but fundamentally responsible and human-centered. I personally consider this volume a landmark contribution to the field and strongly endorse it.
Andreas Holzinger, Ph.D.
Professor and Head of Human-Centered AI Lab, BOKU University, Vienna, Austria
Member, Academia Europaea
This volume provides a timely, comprehensive, and rigorously grounded foundation for understanding how AI systems should be designed, developed, and governed with humans at the center. I strongly recommend it to scholars, practitioners, and decision-makers who are shaping the next generation of intelligent systems.
Honghai Liu, Ph.D.
Chief Scientist and Professor of Intelligent Systems, Southeast University, China
Member, Academia Europaea; IEEE Fellow
As AI becomes more prevalent in everyday life, it becomes even more important to consider how such systems can and should be designed in accordance with human-first principles. This timely handbook brings together the latest thinking on why and how AI systems should be designed around human needs. It is important reading for anyone trying to shape the future of AI.
Joseph A. Konstan, Ph.D.
Professor of Computer Science and Engineering, University of Minnesota, USA
Fellow, IEEE/ACM/AAAS
Rather than a hammer looking for nails, Artificial Intelligence will only realise its full potential when it is seen as a diverse and versatile toolbox for solving highly complex and challenging problems in human-compatible ways. This volume brings together a wide range of insightful perspectives and commentaries on human-centred AI, and will be an invaluable resource for AI researchers and practitioners for years to come.
Peter Flach, Ph.D.
Professor of Artificial Intelligence, University of Bristol, UK
Fellow, European Association for Artificial Intelligence