کنترل هوش مصنوعی خود را در دست بگیرید: راهنمای عملی برای شناسایی، ارزیابی و کمی‌سازی ریسک‌ها

Much of our daily lives intertwine with artificial intelligence. From watching movies recommended by our entertainment streaming service, to interacting with customer service chatbots, to autotagging photos of friends in our social media apps, AI plays an invisible part in enriching our lives. While AI may be seen as a panacea for enterprise advancement and consumer convenience, it is still an emerging technology, and its explosive growth needs to be approached with proper care and preparation. How do we tackle the challenges it presents, and how do we make sure that it does precisely what it is supposed to do? In Keeping Your AI Under Control , author Anand Tamboli explores the inherent risk factors of the widespread implementation of artificial intelligence. The author delves into several real-life case studies of AI gone wrong, including Microsoft's 2016 chatbot disaster, Uber's autonomous vehicle fatally wounding a pedestrian, and an entire smart home in Germany dangerously malfunctioning because of one bad lightbulb. He expertly addresses the need to challenge our current assumptions about the infallibility of technology. The importance of data governance, rigorous testing before roll-out, a chain of human accountability, ethics, and much more are all detailed in Keeping Your AI Under Control . Artificial intelligence will not solve all of our problems for good, but it can (and will) present us with new solutions. These solutions can only be achieved with proper planning, continued maintenance, and above all, a foundation of attuned human supervision. What You Will Learn Understand various types of risks involved in developing and using AI solutions Identify, evaluate, and quantify risks pragmatically Utilize AI insurance to support residual risk management Who This Book Is For Progressive businesses that are on a journey to use AI (buyers/customers), technical and financial leaders in AI solution companies (solution vendors), AI system integrators (intermediaries), project and technology leads of AI deployment projects, technology purchase decision makers, CXOs and legal officers (solution users). Early praise for “Keeping Your AI Under Control” 5 Contents 7 About the Author 8 Acknowledgments 9 Introduction 11 Part I: Future State of AI 14 Chapter 1: Artificial Intelligence Beyond 2020 15 We have changed gears recently 16 AI has covered a lot of ground 16 But it is not everywhere yet 17 How do end users see AI? 18 What are business users thinking? 19 What’s to come? 20 A balanced approach 21 Chapter 2: Learning Lessons from Past Fiascoes 23 When Microsoft’s chatbot went Nazi on social network 24 When Amazon’s same-day delivery caused a racial disparity 24 When Uber’s autonomous car killed a pedestrian 25 When the light bulb DoS attacked an entire smart home 26 When an Australian telco wasted millions of dollars 27 These fiascoes have something in common 28 Key lessons 29 Chapter 3: Understanding AI Risks and Its Impacts 31 Musk, Hawking, Gates—they all said it already! 32 We are suffering from hypocognition 33 Many AI systems are as good as junk 34 There is no balanced scorecard for AI 35 AI knows “how” but not “why” 35 AI does not have a conscience 36 Losing jobs is not a real problem 37 The risk of the right AI in wrong hands 38 The gray matter—use it or lose it 39 Man plans, and technology laughs 40 Part II: Prevention 41 Chapter 4: Evaluating Risks of the AI Solution 42 It starts with the correct hypothesis 43 A curious case of Acme Solutions 44 Does the solution address a valid root cause? 45 Is the solution correctly trained? 46 Has the solution considered all scenarios? 47 Is there an option to stop gracefully? 48 Is the solution explainable and auditable? 48 Is the solution tailored for your use case? 49 Is the solution equipped to handle drift? 50 AI solution evaluation questionnaire 51 Evaluate thoroughly and be in control 53 Chapter 5: De-risking AI Solution Deployment 54 Ensuring long-term strategy is defined 55 Defining your problems correctly 55 Validating all the root causes 56 By applying reverse inference 57 By applying the MECE principle 57 Planning for business continuity 57 Clarifying the accountability 58 Ensuring the availability of the right infrastructure 58 Setting up target metrics of success 59 Defining your acceptance criteria 59 Building the capability for execution 60 Handling security and compliance aspects 60 Integrating with your other systems 61 AI solution deployment questionnaire 62 Asking the right questions is a must 64 Chapter 6: Good AI in the Hands of Bad Users 65 What does a bad user mean to us? 66 Creative users 66 Mischievous users 67 Deliberate (bad) users 67 Luddites 67 Bad user vs. incompetent user 68 Handling with change management 68 Educating users for better adoption 69 Checking the performance and gaps 70 Do not forget Gemba visits 71 Handling user testing and feedback 73 Establishing HAIR department 73 User readiness questionnaire 74 Nailing the technical element is not enough for AI 75 Chapter 7: A Systematic Approach to Risk Mitigation 76 Core principles of AI risk management 76 Frequent contributors to the AI risk 77 Pre-mortem analysis 78 Critical parts of the pre-mortem analysis 79 More about ratings 81 Using the output of the analysis 81 Sector-specific considerations 83 Conclusion 84 Chapter 8: Teach Meticulously and Test Rigorously 85 How AI learns 86 Types of machine learning 86 Five mistakes to avoid when training 87 Not having enough data to train 87 Not cleaning and validating the dataset 88 Not having enough spread in data 89 Ignoring near misses and overrides 89 Conflating correlation and causation 90 Complementing training with testing 91 AI testing is not software testing 91 The cardinal sequence for AI testing 92 Testing for concept drift 93 Testing for interaction issues 94 It can be the basis for your trust 95 Part III: Mitigation 96 Chapter 9: AI Supervision with a Red Team 97 What is a red team? 97 Supervising an AI is necessary 98 Why not use adversarial AI testing? 99 Only AI supervision can expose this 100 Highlighting statistical bloopers 100 Detecting concept drift 101 Finding what’s inside the black box 102 Building your red team 102 Top five red team skills 102 In-house or outsourced 104 Objectives of a red team 104 A red team is not for testing defenses 106 The red team is functional, what next? 106 Chapter 10: Handling Residual Risks 108 What is a residual risk? 109 Existing options and limitations 109 Is there a need for AI insurance? 110 AI loses an investor’s fortune 111 Risky digital assistant for patients 112 Regulations are gearing up 113 Challenges for AI insurance 113 A common pool 114 Equitability 114 Insurability of cloud-based AI 114 Attribution 114 What might make it work? 115 Why the fixation on AI insurance? 115 Insurance or not, you are better off 116 Chapter 11: When Working with Emerging Technologies 117 Are we making the same mistake again? 118 The proliferation of junk solutions 119 A good approach 120 We must remain skeptical 120 We must not overspeed 121 We must insist on a kill switch 121 We must demand quality 122 We must seek transparency 123 We must seek interpretability 124 We must seek simplicity 125 We must seek accountability 125 We must educate everyone involved 126 We must take user feedback seriously 127 We must prefer preemption over a fix 127 Working with AI responsibly 128 Sanity is the key! 128 Appendix A: The Leash System 129 Index 131 Much of our daily lives intertwine with artificial intelligence. From watching movies recommended by our entertainment streaming service, to interacting with customer service chatbots, to autotagging photos of friends in our social media apps, AI plays an invisible part in enriching our lives. While AI may be seen as a panacea for enterprise advancement and consumer convenience, it is still an emerging technology, and its explosive growth needs to be approached with proper care and preparation. How do we tackle the challenges it presents, and how do we make sure that it does precisely what it is supposed to do? In Keeping Your AI Under Control, author Anand Tamboli explores the inherent risk factors of the widespread implementation of artificial intelligence. The author delves into several real-life case studies of AI gone wrong, including Microsoft's 2016 chatbot disaster, Uber's autonomous vehicle fatally wounding a pedestrian, and an entire smart home in Germany dangerously malfunctioning because of one bad light bulb. He expertly addresses the need to challenge our current assumptions about the infallibility of technology. The importance of data governance, rigorous testing before roll-out, a chain of human accountability, ethics, and much more are all detailed in Keeping Your AI Under Control. Artificial intelligence will not solve all of our problems for good, but it can (and will) present us with new solutions. These solutions can only be achieved with proper planning, continued maintenance, and above all, a foundation of attuned human supervision Front Matter ....Pages i-xv Front Matter ....Pages 1-1 Artificial Intelligence Beyond 2020 (Anand Tamboli)....Pages 3-10 Learning Lessons from Past Fiascoes (Anand Tamboli)....Pages 11-18 Understanding AI Risks and Its Impacts (Anand Tamboli)....Pages 19-28 Front Matter ....Pages 29-29 Evaluating Risks of the AI Solution (Anand Tamboli)....Pages 31-42 De-risking AI Solution Deployment (Anand Tamboli)....Pages 43-53 Good AI in the Hands of Bad Users (Anand Tamboli)....Pages 55-65 A Systematic Approach to Risk Mitigation (Anand Tamboli)....Pages 67-75 Teach Meticulously and Test Rigorously (Anand Tamboli)....Pages 77-87 Front Matter ....Pages 89-89 AI Supervision with a Red Team (Anand Tamboli)....Pages 91-101 Handling Residual Risks (Anand Tamboli)....Pages 103-111 When Working with Emerging Technologies (Anand Tamboli)....Pages 113-124 Back Matter ....Pages 125-129