(2022) Ben Shneiderman, Oxford University Press, £20, hrdbk, xii + 377pp, ISBN 978-0-192-84529-0
Following twelve pages of Preface and Acknowledgements, the main text runs for two hundred and eighty pages (including a short Personal Epilogue). The remainder consists of copious Notes, Bibliography, and Name and Subject Indices.
The book takes the form of a discussion on the development of Artificial Intelligences (AIs) which take account of people (i.e. human-centred) as against developing AIs with no thought for human aspects. It is divided into five sections, each of several chapters and each ending with a Summary, and finally there is a Skeptic’s Corner chapter. ‘What Is Human-Centered Artificial Intelligence?’ opens the discussion, in which he looks at rationalism and empiricism, compares people with computers, and considers the possible impact on employment as we increase our use of automated systems, AIs, and robots to help us perform our tasks and run our industries. ‘Human-Centred AI Framework’ then looks at defining reliable, safe, and trustworthy systems, using human-centred AI frameworks, and follows up by examining guidelines and examples.
‘Design Metaphors’ moves on to the goals of AI research and considers science and innovation goals, intelligent agents and supertools, teammates and tele-bots, assured autonomy and control centres, and social robots and active appliances. This is followed by ‘Governance Structures’, which considers bridging the gap between ethics and practice, reliable systems based on sound software engineering practices, and safety cultures through business management strategies, before continuing on to consider trustworthy certification by independent oversight and the role of government regulations and intervention.
Finally, ‘Where Do We Go From Here?’ looks to the future, such as driving human-centred AI forward and assessing trustworthiness, and considering the likes of caring for and learning from our older adults.
Exactly what constitutes an AI is the subject of much public discussion but as far as the author is concerned an AI is any system which makes decisions and takes actions on our behalf. One of the simplest AIs is the humble domestic thermostat; without any input from ourselves, it turns on the heating when our homes get cool and turns it off when they are warm enough. It has a human aspect though - we get to set and change the desired temperature, and some such devices will also tell us the current temperature. On the other hand, other, much more powerful AIs pour over vast amounts of data, categorise it, and make decisions such as which job candidates should be invited for interview or how machines such as cars and aeroplanes should respond to circumstances, but without any human involved in the decision-making process.
The author explains that there are two approaches to AI design: rationalism and empiricism. The rationalist approach looks purely at logic and assumes that if the rules are correctly ascertained and perfect algorithms applied then a system will always work flawlessly. This approach has lead to many excellent systems; however, when an unexpected event happens (such as an unforeseen circumstance or a failed component) it can lead to disaster. The well known problems with Boeing’s new 737 MAX aircraft are often used as examples, as are some of the failures of Tesla’s Autopilot driver-assistance feature. The empiricist approach, on the other hand, is that not everything can be predicted and that things can go wrong, and this should be taken into account in the design. A simple example is to add a Cancel button for when the system is taking the user in an undesired direction.
An example of human-centred AI that the author often makes use of is the modern digital camera. It automatically performs a huge amount of work for us; it checks each pixel, compares it with its neighbouring pixels, establishes colours, colour casts, lighting conditions, straight and curved lines, focussing, depth of field, and so on, but allows the user to control the overall picture. Yet if the user wishes to do so they can select the shutter speed, the aperture, the exposure (perhaps over or under exposing for the effect they want), the focus and depth of field, and the colour rendition (perhaps enforcing the warm oranges of a sunset or the crisp brightness of a snow scene under a cloudless sky).
As non-fiction books go, this is not the most exciting of reads. It is not telling the story of anything and there is nothing to carry the reader along. It is very much a text book, the sort of thing I read as part of my studies. However, for anyone working in the field of AI, intending to work in AI, or merely studying the field in either a professional or amateur way, this has a lot to say - and it needs saying.
Be in no doubt, the use of AIs is increasing and will continue to do so, affecting many walks of life. A recent BBC documentary, for example, pointed out the errors made in poor, AI-based systems that accessed the performance of employees, assessed people’s moods and motivations, etc., and made the important point that an AI system can easily have built-in but unrecognised biases - and ultimately that is bad for everyone. There is already much concern, for example, about the inability of facial recognition systems to work well with darker skin colours. It is therefore imperative people remain in control of the systems we create and that these systems help people rather than make life more difficult - and for that AIs must be centred on human requirements. This book covers much ground on why this is necessary and how such human-centring can be built in and maintained, and why it should be a matter of course that AIs should always be human-centred to the necessary degree.
Furthermore, the author stresses the need for accountability: that all AIs should include audit trails which allow their assessments and decisions to be monitored and examined, akin to the black boxes that are fitted to commercial aircraft. This would ensure that they can be trusted and that any mistakes are quickly caught and avoided in future. He also stresses the necessity to look out for near misses, i.e. things which nearly went wrong, and again he quotes the airline industry as the use of black boxes and the requirement to report near misses has been so effective in making air travel so very safe.
The author is both optimistic and realistic about the abilities of AIs to augment and enhance our lives, but concerned that AIs must be of benefit to us through being human-centred. It makes for, as they say, interesting reading!
Peter Tyers
(Note on spelling: the book uses American spellings, which are sometimes quoted, whereas this reviewer uses English spellings.)
[Up: Non-Fiction Index | Top: Concatenation]