By William Leiss and Richard Smith

This forthcoming book presents a comprehensive analysis of the risks associated with Artificial Intelligence (AI), in the context of the impending transition of artificial intelligence from a passive tool to an autonomous actor.

The authors argue that while current AI provides significant benefits in medicine and science, there is an underlying risk that humans could lose control over the advanced AI systems (also known as frontier AI) which are now being created, leading to potentially catastrophic outcomes for humanity. The book’s chapter 9, which presents an original probabilistic safety assessment of loss of control, estimates a 1 in 8 chance of this eventuality occurring by the mid-21st century.

The second important underlying risk is the ongoing effort to create superintelligence, defined as a machine whose cognitive and operational capacities exceed that of humans by a wide margin. The authors argue that this development represents a credible risk of human extinction and call for an immediate global prohibition on developing superintelligent systems*.*

Finally, they contend that existing and proposed government regulations of AI are structurally inadequate because they treat AI as a controllable product rather than an adaptive, and sometimes deceptive, entity. The work emphasizes that internal self-control must be architecturally guaranteed before such powerful technologies are integrated into society.

The companion website will launch fully at publication, with ongoing commentary, an annotated bibliography, interactive tools, and resources for readers.

Publication details forthcoming. Expected publication date Spring 2027

Our book is the first to bring the well-established risk management approach to the issue of the catastrophic risks associated with Artificial Intelligence (AI), which have been identified by many AI professionals. This approach uses a formal methodology in the evaluation of risks, considering probability or likelihood, and consequences or outcomes, as separate aspects.* Separating the two influences the perceived severity of risks and the options for managing them—as well as judgments as to whether the level of risk is acceptable to the public. Using this approach, we find that certain catastrophic risks of AI are unacceptable at present, and therefore stringent regulations by governments are necessary in order to guarantee the safety of Frontier AI systems. We call upon the Government of Canada to promulgate such regulations and join with other countries to create and enforce an international treaty on AI safety.

NOTE: *This approach includes Quantitative Risk Assessment (QRA), and our book includes the first QRA study of an AI risk issue by our colleague Michael G. Tyshenko.

Contents

List of Figures List of Tables Preface

Part One: Artificial Intelligence

Introduction Chapter 1: The Good Chapter 2: The Bad Chapter 3: The Ugly Chapter 4: The Superhuman Machine

Part 2: Risk Policy and Analysis

Chapter 5: The Risk-Based Approach Chapter 6: The Control Problem Chapter 7: Catastrophic and Existential Risks Chapter 8: Risk Estimation Chapter 9: Quantitative Risk Assessment for Loss of Control (By Michael G. Tyshenko)

Part Three: Governance and Risk Management

Chapter 10: Regulation Chapter 11: Risk Mitigation Options Chapter 12: Heart of Darkness Chapter 13: Blaming the Victims Chapter 14: What Should the Public Do? Chapter 15: What Should Governments Do? Chapter 16: Risk Management Summary References Acknowledgements