AI Unexplainable, Unpredictable, Uncontrollable (Roman V. Yampolskiy) (Z-Library)

📄 File Format: PDF

💾 File Size: 3.3 MB

📖 Read Online ⬇️ Download

Registered users can read the full content for free

Register as a Gaohf Library member to read the complete e-book online for free and enjoy a better reading experience.

📄 Page 1

CHAPMAN & HALUCRC ARTIFICIAL INTELLIGENCE AND ROBOTICS SERIES • S�r�Fr:n��Zr�up A CHAPMAN & HALL BOOK Al Unexplainable, Unpredictable, Uncontrollable Roman V. Yampolskiy, PhD

📄 Page 2

AI Delving into the deeply enigmatic nature of artificial intelligence (AI), AI: Unpredictable, Unexplainable, Uncontrollable explores the various reasons why the field is so challenging. Written by one of the founders of the field of AI safety, this book addresses some of the most fascinating questions facing humanity, including the nature of intelligence, consciousness, values, and knowledge. Moving from a broad introduction to the core problems, such as the unpre- dictability of AI outcomes or the difficulty in explaining AI decisions, this book arrives at more complex questions of ownership and control, conduct- ing an in-depth analysis of potential hazards and unintentional consequences. The book then concludes with philosophical and existential considerations, probing into questions of AI personhood, consciousness, and the distinction between human intelligence and artificial general intelligence (AGI). Bridging the gap between technical intricacies and philosophical musings, AI: Unpredictable, Unexplainable, Uncontrollable appeals to both AI experts and enthusiasts looking for a comprehensive understanding of the field, while also being written for a general audience with minimal technical jargon.

📄 Page 3

Chapman & Hall/CRC Artificial Intelligence and Robotics Series Series Editor: Roman Yampolskiy Topological Dynamics in Metamodel Discovery with Artificial Intelligence From Biomedical to Cosmological Technologies Ariel Fernández A Robotic Framework for the Mobile Manipulator Theory and Application Nguyen Van Toan and Phan Bui Khoi AI in and for Africa A Humanist Perspective Susan Brokensha, Eduan Kotzé, Burgert A. Senekal Artificial Intelligence on Dark Matter and Dark Energy Reverse Engineering of the Big Bang Ariel Fernández Explainable Agency in Artificial Intelligence Research and Practice Silvia Tulli and David W. Aha An Introduction to Universal Artificial Intelligence Marcus Hutter, Elliot Catt, and David Quarel AI: Unpredictable, Unexplainable, Uncontrollable Roman V. Yampolskiy For more information about this series please visit: https://www.routledge.com/ Chapman–HallCRC-Artificial-Intelligence-and-Robotics-Series/book-series/ARTILRO

📄 Page 4

AI Unexplainable, Unpredictable, Uncontrollable Roman V. Yampolskiy, PhD

📄 Page 5

First edition published 2024 by CRC Press 2385 NW Executive Center Drive, Suite 320, Boca Raton FL 33431 and by CRC Press 4 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN CRC Press is an imprint of Taylor & Francis Group, LLC © 2024 Roman V. Yampolskiy Reasonable efforts have been made to publish reliable data and information, but the author and publisher cannot assume responsibility for the validity of all materials or the conse- quences of their use. The authors and publishers have attempted to trace the copyright holders of all material reproduced in this publication and apologize to copyright holders if permission to publish in this form has not been obtained. If any copyright material has not been acknowledged please write and let us know so we may rectify in any future reprint. Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, repro- duced, transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and recording, or in any information storage or retrieval system, without written permission from the publishers. For permission to photocopy or use material electronically from this work, access www. copyright.com or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. For works that are not available on CCC please contact mpkbookspermissions@tandf.co.uk Trademark notice: Product or corporate names may be trademarks or registered trademarks and are used only for identification and explanation without intent to infringe. ISBN: 978-1-032-57627-5 (hbk) ISBN: 978-1-032-57626-8 (pbk) ISBN: 978-1-003-44026-0 (ebk) DOI: 10.1201/9781003440260 Typeset in Palatino by KnowledgeWorks Global Ltd.

📄 Page 6

To my friend, Jaan Tallinn, the man who did more for the world than you will ever know.

📄 Page 7

(This page has no text content)

📄 Page 8

vii Contents Acknowledgments .................................................................................................xi Author ................................................................................................................... xiii 1 Introduction .....................................................................................................1 1.1 Introduction ...........................................................................................1 1.2 The AI Control Problem .......................................................................2 1.3 Obstacles to Controlling AI .................................................................3 1.4 Defining Safe AI ....................................................................................4 1.5 On Governability of AI ........................................................................5 1.6 Conclusions ............................................................................................6 1.7 About the Book ......................................................................................7 References .........................................................................................................8 2 Unpredictability ............................................................................................ 11 2.1 Introduction to Unpredictability ...................................................... 11 2.2 Predictability: What We Can Predict – A Literature Review ............................................................................ 13 2.3 Cognitive Uncontainability ............................................................... 15 2.4 Conclusions .......................................................................................... 16 References ....................................................................................................... 17 3 Unexplainability and Incomprehensibility ............................................ 21 3.1 Introduction .........................................................................................22 3.2 Literature Review ................................................................................22 3.3 Unexplainability ..................................................................................25 3.4 Incomprehensibility ............................................................................ 27 3.5 Conclusions ..........................................................................................30 Notes ................................................................................................................30 References ....................................................................................................... 31 4 Unverifiability ...............................................................................................36 4.1 On Observers and Verifiers ...............................................................36 4.2 Historical Perspective ......................................................................... 37 4.3 Classification of Verifiers ................................................................... 39 4.4 Unverifiability ......................................................................................42 4.5 Unverifiability of Software ................................................................43 4.5.1 Unverifiability of Artificial Intelligence .............................44 4.6 Conclusions and Future Work ..........................................................45

📄 Page 9

viii Contents Notes ................................................................................................................46 References .......................................................................................................46 5 Unownability ................................................................................................. 51 5.1 Introduction ......................................................................................... 51 5.1.1 Proposals for Establishing Ownership ............................... 52 5.2 Obstacles to Ownership ..................................................................... 52 5.3 Conclusions ..........................................................................................54 References .......................................................................................................55 6 Uncontrollability .......................................................................................... 57 6.1 Introduction ......................................................................................... 59 6.2 AI Control Problem .............................................................................60 6.2.1 Types of Control Problems ...................................................60 6.2.2 Formal Definition ................................................................... 62 6.3 Previous Work ..................................................................................... 69 6.3.1 Controllable ............................................................................ 69 6.3.2 Uncontrollable ........................................................................72 6.4 Proving Uncontrollability .................................................................. 78 6.5 Multidisciplinary Evidence for Uncontrollability of AI ................ 81 6.5.1 Control Theory ....................................................................... 82 6.5.2 Philosophy ..............................................................................84 6.5.3 Public Choice Theory ............................................................85 6.5.4 Justice (Unfairness) ................................................................86 6.5.5 Computer Science Theory ..................................................... 87 6.5.6 Cybersecurity .........................................................................88 6.5.7 Software Engineering ...........................................................88 6.5.8 Information Technology ....................................................... 89 6.5.9 Learnability ............................................................................ 89 6.5.10 Economics ...............................................................................90 6.5.11 Engineering ............................................................................90 6.5.12 Astronomy ..............................................................................90 6.5.13 Physics ..................................................................................... 91 6.6 Evidence from AI Safety Research for Uncontrollability of AI ....................................................................................................... 91 6.6.1 Value Alignment .................................................................... 93 6.6.2 Brittleness ................................................................................ 94 6.6.3 Unidentifiability ..................................................................... 95 6.6.4 Uncontainability .................................................................... 96 6.6.5 Uninterruptability ................................................................. 97 6.6.6 AI Failures ............................................................................... 98 6.6.7 Unpredictability ..................................................................... 98 6.6.8 Unexplainability and Incomprehensibility ........................99 6.6.9 Unprovability ....................................................................... 100 6.6.10 Unverifiability ...................................................................... 101 6.6.11 Reward Hacking .................................................................. 103

📄 Page 10

ixContents 6.6.12 Intractability ......................................................................... 103 6.6.13 Goal Uncertainty .................................................................. 104 6.6.14 Complementarity ................................................................. 105 6.6.15 Multidimensionality of Problem Space ............................ 106 6.7 Discussion .......................................................................................... 106 6.8 Conclusions ........................................................................................ 109 Notes .............................................................................................................. 112 References ..................................................................................................... 112 7 Pathways to Danger .................................................................................... 128 7.1 Taxonomy of Pathways to Dangerous AI ...................................... 128 7.1.1 On Purpose – Pre-Deployment .......................................... 128 7.1.2 On Purpose – Post-Deployment ........................................ 130 7.1.3 By Mistake – Pre-Deployment ........................................... 131 7.1.4 By Mistake – Post-Deployment .......................................... 132 7.1.5 Environment – Pre-Deployment ........................................ 133 7.1.6 Environment – Post-Deployment ...................................... 133 7.1.7 Independently – Pre-Deployment ..................................... 133 7.1.8 Independently – Post-Deployment .................................... 133 7.2 Conclusions ........................................................................................ 134 References ..................................................................................................... 135 8 Accidents ...................................................................................................... 139 8.1 Introduction ....................................................................................... 139 8.2 AI Failures .......................................................................................... 140 8.2.1 Preventing AI Failures ........................................................ 146 8.3 AI Safety ............................................................................................. 148 8.4 Cybersecurity vs. AI Safety ............................................................. 149 8.5 Conclusions ........................................................................................ 151 Notes .............................................................................................................. 151 References ..................................................................................................... 154 9 Personhood .................................................................................................. 158 9.1 Introduction to AI Personhood ....................................................... 158 9.2 Selfish Memes .................................................................................... 159 9.3 Human Indignity .............................................................................. 160 9.4 Legal-System Hacking ...................................................................... 161 9.5 Human Safety .................................................................................... 162 9.6 Conclusions ........................................................................................ 165 Notes .............................................................................................................. 166 References ..................................................................................................... 166 10 Consciousness ............................................................................................. 170 10.1 Introduction to the Problem of Consciousness ............................. 170 10.2 Test for Detecting Qualia ................................................................. 173

📄 Page 11

x Contents 10.3 Computers Can Experience Illusions, and so Are Conscious.................................................................................... 176 10.3.1 Qualia Computing ............................................................... 177 10.4 Purpose of Consciousness ............................................................... 178 10.4.1 Qualia Engineering ............................................................. 179 10.5 Consciousness and Artificial Intelligence ..................................... 180 10.6 Conclusions and Conjectures .......................................................... 182 Note ................................................................................................................ 184 References ..................................................................................................... 184 11 Personal Universes ..................................................................................... 195 11.1 Introduction to the Multi-Agent Value Alignment Problem ............................................................................................... 195 11.2 Individual Simulated Universes ..................................................... 196 11.3 Benefits and Shortcomings of Personalized Universes ............... 199 11.4 Conclusions ........................................................................................200 Note ................................................................................................................ 201 References ..................................................................................................... 201 12 Human ≠ AGI .............................................................................................. 206 12.1 Introduction ....................................................................................... 206 12.2 Prior Work .......................................................................................... 207 12.3 Humans Are Not AGI ...................................................................... 209 12.4 Conclusions ........................................................................................ 211 Note ................................................................................................................ 213 References ..................................................................................................... 213 13 Skepticism .................................................................................................... 217 13.1 Introduction to AI Risk Skepticism ................................................ 217 13.2 Types of AI Risk Skeptics ................................................................. 219 13.2.1 Skeptics of Strawman .......................................................... 221 13.3 Arguments for AI Risk Skepticism ................................................222 13.3.1 Priorities Objections ............................................................223 13.3.2 Technical Objections ...........................................................225 13.3.3 AI Safety-Related Objections .............................................227 13.3.4 Ethical Objections ................................................................228 13.3.5 Biased Objections .................................................................229 13.3.6 Miscellaneous Objections ...................................................230 13.4 Countermeasures for AI Risk Skepticism ..................................... 232 13.5 Conclusions ........................................................................................234 Notes ..............................................................................................................235 References .....................................................................................................236 Index ..................................................................................................................... 243

📄 Page 12

xi Acknowledgments I would like to thank a great number of people for helping me, sharing their ideas, commenting on my work, supporting my research, or simply inspir- ing my thinking in many ways. Some of them I had the pleasure of meet- ing in person, and others are virtual presences on my computer, but ideas are substrate independent, so they are all equally amazing. I am confident that I forgot many important people simply because I am not superintelligent and my memory is terrible. I apologize in advance for any such omissions. I acknowledge contributions of many great scientists by simply citing their work as that is the greatest recognition of scientific contribution anyone can wish for. A huge thank you goes to Max Tegmark, Ian Goodfellow, Kenneth Regan, Edward Frenkel, Sebastien Zany, Søren Elverlin, Melissa Helton, Anna Husfeldt, Thore Husfeldt, David Kelley, David Jilk, Scott Aaronson, Rob Bensinger, Seth Baum, Tony Barrett, and Alexey Turchin. Last but certainly not least I would like to thank Jaan Tallinn and Survival and Flourishing Fund, and Elon Musk and the Future of Life Institute for partially funding my work on AI Safety.

📄 Page 13

(This page has no text content)

📄 Page 14

xiii Author Dr. Roman V. Yampolskiy is a tenured associate pro- fessor in the Department of Computer Science and Engineering at the Speed School of Engineering, University of Louisville. He is the founding and cur- rent director of the Cyber Security Lab and an author of many books including Artificial Superintelligence: A Futuristic Approach, Editor of AI Safety and Security and The Technological Singularity. During his tenure at UofL, Dr. Yampolskiy has been recognized as: Distinguished Teaching Professor, Professor of the Year, Faculty Favorite, Top 4 Faculty, Leader in Engineering Education, Top 10 of Online College Professor of the Year, and Outstanding Early Career in Education award winner among many other honors and distinctions. Yampolskiy was promoted to a Senior member of IEEE and AGI; Member of Kentucky Academy of Science, and he is a former research advisor for MIRI and GCRI. Roman Yampolskiy holds a PhD degree from the Department of Computer Science and Engineering at the University at Buffalo. He was a recipient of a four-year NSF (National Science Foundation) IGERT (Integrative Graduate Education and Research Traineeship) fellowship. Before beginning his doc- toral studies, Dr. Yampolskiy received a BS/MS (High Honors) combined degree in Computer Science from Rochester Institute of Technology, NY, USA. After completing his PhD dissertation, Dr. Yampolskiy held a position of an Affiliate Academic at the Center for Advanced Spatial Analysis, University of London, College of London. He had previously conducted research at the Laboratory for Applied Computing at the Rochester Institute of Technology and at the Center for Unified Biometrics and Sensors at the University at Buffalo. Dr. Yampolskiy is an alumnus of Singularity University (GSP2012) and a visiting fellow of the Singularity Institute (Machine Intelligence Research Institute). Dr. Yampolskiy’s main area of interest is AI Safety. Dr. Yampolskiy is an author of over 200 publications including multiple journal articles and books. His research has been cited by 1000s of scientists and profiled in popular magazines both American and foreign. Dr. Yampolskiy’s research has been featured 10,000+ times in numerous media reports in 40+ languages. Twitter: @romanyam

📄 Page 15

(This page has no text content)

📄 Page 16

1DOI: 10.1201/9781003440260-1 1 Introduction* * Parts of this chapter have been previously published as On Governability of AI by Roman Yampolskiy. AI Governance in 2020 a Year in Review. June, 2021 and On Defining Differences Between Intelligence and Artificial Intelligence by Roman V. Yampolskiy. Journal of Artificial General Intelligence 11(2), 68-70. 2020. 1.1 Introduction Rapid advances in artificial intelligence (AI) over the past decade have been accompanied by several high-profile failures [1], highlighting the importance of ensuring that intelligent machines are beneficial to humanity. This realiza- tion has given rise to the new subfield of research known as AI Safety and Security [2], which encompasses a wide range of research areas and has seen steady growth in publications in recent years [3–10]. However, the underlying assumption in this research is that the problem of controlling highly capable intelligent machines is solvable, though no rig- orous mathematical proof or argumentation has been presented to demon- strate that the AI control problem is solvable in principle, let alone in practice. In computer science, it is standard practice to first determine whether a prob- lem belongs to a class of unsolvable problems before investing resources in trying to solve it. Despite the recognition that the problem of AI control may be one of the most important problems facing humanity, it remains poorly understood, poorly defined, and poorly researched. A computer science problem could be solvable, unsolvable, undecidable, or partially solvable; we don’t know the actual status of the AI control problem. It is possible that some forms of con- trol may be possible in certain situations, but it is also possible that partial control may be insufficient in many cases. Without a better understanding of the nature and feasibility of the AI control problem, it is difficult to deter- mine an appropriate course of action [11]. Potential control methodologies for artificial general intelligence (AGI) have been broadly classified into two categories: Methods based on capa- bility control and motivational control [12]. Capability control methods aim to limit the damage that AGI systems can cause by placing them in constrained environments, adding shutdown mechanisms or trip wires. Motivational control methods attempt to design AGI systems to have an innate desire not to cause harm, even in the absence of capacity control

📄 Page 17

2 measures. It is widely recognized that capacity control methods are, at best, temporary safety measures and do not represent a long-term solution to the AGI control problem [12]. Furthermore, it is likely that motivational con- trol measures should be integrated at the design and training phase, rather than after deployment. 1.2 The AI Control Problem We define the problem of AI control as: How can humanity remain safely in con- trol while benefiting from a superior form of intelligence? This is the fundamental problem in the field of AI Safety and Security, which aims to make intelli- gent systems safe from tampering and secure for all stakeholders involved. Value alignment is currently the most studied approach to achieve security in AI. However, concepts such as safety and security are notoriously dif- ficult to test or measure accurately, even for non-AI software, despite years of research [13]. At best, we can probably distinguish between perfectly safe and as safe as an average person performing a similar task. However, society is unlikely to tolerate machine errors, even if they occur with a frequency typical of human performance or even less frequently. We expect machines to perform better and will not accept partial safety when dealing with such highly capable systems. The impact of AI (both positive and negative [3]) is strongly related to its capability. With respect to possible existential impacts, there is no such thing as partial safety. An initial understanding of the control problem may suggest designing a machine that accurately follows human commands. However, because of pos- sible conflicting or paradoxical commands, ambiguity of human languages [14], and perverse instance creation problems, this is not a desirable form of control, although some ability to integrate human feedback may be desirable. The solution is thought to require AI to act in the capacity of an ideal advisor, avoiding the problems of misinterpretation of direct commands and the pos- sibility of malevolent commands. It has been argued that the consequences of an uncontrolled AI could be so severe that even if there is a very small chance of a hostile AI emerging, it is still worthwhile to conduct AI Safety research because the negative utility of such an AI would be astronomical. The common logic is that an extremely high (negative) utility multiplied by a small chance of the event still results in a large disutility and should be taken very seriously. However, the reality is that the chances of a misaligned AI are not small. In fact, in the absence of an effective safety program, that is the only outcome we will get. So the statistics look very compelling in support of a major AI Safety effort. We are looking at an almost guaranteed event with the potential to cause an existential catastrophe. This is not a low-risk, high-reward scenario, but a AI

📄 Page 18

3Introduction high-risk, negative-reward situation. No wonder many consider this to be the most important problem humanity has ever faced. The outcome could be prosperity or extinction, and the fate of the universe hangs in the balance. A proof of the solvability or non-solvability of the AI control problem would be the most important proof ever. 1.3 Obstacles to Controlling AI Controlling an AGI is likely to require a toolbox with certain capabilities, such as explainability, predictability, and model verifiability [15]. However, it is likely that many of the desired tools are not available to us. • The concept of Unexplainability in AI refers to the impossibility of providing an explanation for certain decisions made by an intelligent system that is 100% accurate and understandable. A complementary concept to Unexplainability, Incomprehensibility of AI addresses the inability of people to fully understand an explanation provided by an AI. We define Incomprehensibility as the impossibility to fully comprehend any 100% accurate explanation for certain decisions of intelligent systems, by any human being [16]. • Unpredictability of AI, one of the many impossibility outcomes in AI Safety, also known as Unknowability, is defined as our inabil- ity to accurately and consistently predict what specific actions an intelligent system will take to achieve its goals, even if we know the ultimate goals of the system [17]. It is related to but not the same as the Unexplainability and Incomprehensibility of AI. Unpredictability does not imply that better-than-random statisti- cal analysis is impossible; it simply points to a general limitation on how well such efforts can work, particularly pronounced with advanced generally intelligent systems in novel domains. • Non-verifiability is a fundamental limitation in the verification of mathematical proofs, computer software, intelligent agent behavior, and all formal systems [18]. It is becoming increasingly obvious that just as we can only have probabilistic confidence in the correctness of mathematical proofs and software implementations, our ability to verify intelligent agents is at best limited. Many researchers assume that the problem of AI control can be solved despite the absence of any evidence or proof. Before embarking on a quest to build controlled AI, it is important to demonstrate that the problem can be solved so as not to waste valuable resources. The burden of proof is on

📄 Page 19

those who claim that the problem is solvable, and the current absence of such proof speaks loudly about the inherent dangers of the proposal to develop AGI. In fact, Uncontrollability of AI is very likely to be the case, as can be demonstrated by reduction to the problem of human control. There are many open questions to consider regarding the issue of controllability, such as: Can the control problem be solved? Can it be done in principle? Can it be done in practice? Can it be done with a sufficient level of accu- racy? How long would it take to do it? Can it be done in time? What are the energy and computational requirements to do it? What would a solution look like? What is the minimum viable solution? How would we know if we solved it? Does the solution scale as the system continues to improve? We argue that unconstrained intelligence cannot be controlled and con- strained intelligence cannot innovate. If AGI is not properly controlled, no matter who programmed it, the consequences will be disastrous for everyone, probably its programmers in the first place. No one benefits from uncontrolled AGI. There seems to be a lack of published evidence to conclude that a less intel- ligent agent can indefinitely maintain control over a more intelligent agent. As we develop intelligent systems that are less intelligent than we are, we can maintain control, but once such systems become more intelligent than we are, we lose that ability. In fact, as we try to maintain control while designing advanced intelligent agents, we find ourselves in a Catch-22, since the control mechanism needed to maintain control must be smarter or at least as smart as the agent over which we want to maintain control. A whole hierarchy of intelligent systems would need to be built to control increasingly capable sys- tems, leading to infinite regress. Moreover, the problem of controlling such more capable intelligence only becomes more challenging and more obvi- ously impossible for agents with only a static level of intelligence. Whoever is more intelligent will be in control, and those in control will be the ones with the power to make the final decisions. As far as we know, as of this moment, no one in the world has a working AI control mechanism capable of scaling to human-level AI and eventually beyond, or even an idea for a prototype that might work. No one has made verifiable claims to have such a technology. In general, for anyone claiming that the problem of AI control is solvable, the burden of proof is on them. Currently, it appears that our ability to produce intelligent software far outstrips our ability to control or even verify it. 1.4 Defining Safe AI In “On Defining Artificial Intelligence” Pei Wang presents the following definition [19]: “Intelligence is the capacity of an information-processing sys- tem to adapt to its environment while operating with insufficient knowledge 4 AI

📄 Page 20

5Introduction and resources” [20]. Wang’s definition is perfectly adequate, and he also reviews definitions of intelligence suggested by others, which have by now become standard in the field [21]. However, there is a fundamental difference between defining intelligence in general or human intelligence in particular and defining AI as the title of Wang’s paper claims he does. In this chap- ter, I would like to bring attention to the fundamental differences between designed and natural intelligences [22]. AI is typically designed for the explicit purpose of providing some benefit to its designers and users and it is important to include that distinction in the definition of AI. Wang only once, briefly, mentions the concept of AI Safety [12, 23–26] in his article and doesn’t bring it or other related con- cepts into play. In my opinion, definition of AI which doesn’t explicitly men- tion safety or at least its necessary subcomponents, such as controllability, explainability [27], comprehensibility, predictability [28], and corrigibility [29], is dangerously incomplete. Development of AGI is predicted to cause a shift in the trajectory of human civilization [30]. In order to reap the benefits and avoid pitfalls of such power- ful technology, it is important to be able to control it. Full control of intelligent system [31] implies capability to limit its performance [32], for example, set- ting it to a particular level of IQ equivalence. Additional controls may make it possible to turn the system off [33], and turn on/off consciousness [34, 35], free will, autonomous goal selection, and specify moral code [36] the system will apply in its decisions. It should also be possible to modify the system after it is deployed to correct any problems [1, 37] discovered during use. An AI system should be able, to the extent theoretically possible, explain its decisions in a human-comprehensible language. Its designers and end users should be able to predict its general behavior. If needed, the system should be confinable to a restricted environment [38–40], or operate with reduced computational resources. AI should be operating with minimum bias and maximum transparency; it has to be friendly [41], safe, and secure [2]. Consequently, we propose the following definition of AI which compli- ments Wang’s definition: “Artificial Intelligence is a fully controlled agent with a capacity of an information-processing system to adapt to its environ- ment while operating with insufficient knowledge and resources”. 1.5 On Governability of AI In order to make future AIs beneficial for all of humanity, AI governance initiatives attempt to make AI governable by the world’s governments, international organizations, and multinational corporations collaborating on establishing a regulatory framework and industry standards. However, direct governance of AI is not meaningful, and what is implied by the term