Take Action
Course
Curriculum
Download Textbook
1. Overview of Catastrophic AI Risks
0.1
Preface
1.1
Overview of Catastrophic AI Risks
1.2
Malicious Use
1.3
AI Race
1.4
Organizational Risks
1.5
Rogue AIs
1.6
Discussion of Connections Between Risks
2. AI Fundamentals
2.1
AI Fundamentals
2.2
Artificial Intelligence & Machine Learning
2.3
Deep Learning
2.4
Scaling Laws
2.5
Conclusion
3. Single Agent Safety
3.1
Single Agent Safety
3.2
Monitoring
3.3
Robustness
3.4
Alignment
3.5
Systemic Safety
3.6
Safety and General Capabilities
3.7
Conclusion
4. Safety Engineering
4.1
Safety Engineering
4.2
Risk Decomposition
4.3
Nines of Reliability
4.4
Safe Design Principles
4.5
Component Failure Accident Models and Methods
4.6
Systemic Factors
4.7
Tail Events and Black Swans
4.8
Conclusion
5. Complex Systems
5.1
Complex Systems
5.2
Introduction to Complex Systems
5.3
Complex Systems for AI Safety
5.4
Conclusion
6. Beneficial AI and Machine Ethics
6.1
Beneficial AI and Machine Ethics
6.2
Law
6.3
Fairness
6.4
The Economic Engine
6.5
Wellbeing
6.6
Preferences
6.7
Happiness
6.8
Social Welfare Functions
6.9
Moral Uncertainty
7. Collective Action Problems
7.1
Collective Action Problems
7.2
Game Theory
7.3
Cooperation
7.4
Conflict
7.5
Evolutionary Pressures
7.6
Conclusion
8. Governance
8.1
Governance
8.2
Timelines
8.3
Distribution
8.4
Compute Governance
8.5
Corporate Governance
8.6
National Governance
8.7
International Governance
8.8
Conclusion
9. Appendices
9.1
Acknowledgements
10.1
App. A: Normative Ethics
11.1
App. B: Utility Functions
12.1
App. C: Reinforcement Learning
13.1
App. D: Long-Tailed and Thin-Tailed Distributions
14.1
App. E: Evolutionary Game Theory
15.1
App. F: Other Cooperation Mechanisms
16.1
App. G: Intrasystem Conflict Causes
All
Section
Appendix
3.3

Robustness

AI systems are vulnerable to adversarial examples, Trojans and other attacks. We need progress on improving the robustness of models to ensure that powerful AI systems are not misused or stolen.

No items found.

Review Questions

How can the use of inaccurate proxies lead AI systems to cause harm?

Answer:

Proxies that fail to capture important values we care about can result in models exploiting those gaps and taking undesired actions that negatively impact people.

View Answer
Hide Answer

What is an example that illustrates Goodhart's law in proxies?

Answer:

In Hanoi, paying for rat tails to control the population led people to just cut off tails, increasing the rat population over time as the proxy and goal became inversely related.

View Answer
Hide Answer

Why could reliance on AI systems to evaluate other AIs be risky?

Answer:

AI evaluators could be vulnerable to exploitation from adversarial attacks or proxy gaming, undermining their ability to accurately assess other AI systems.

View Answer
Hide Answer
0.1
1.1
1.2
1.3
1.4
1.5
1.6
2.1
2.2
2.3
2.4
2.5
3.1
3.2
3.3
3.4
3.5
3.6
3.7
4.1
4.2
4.3
4.4
4.5
4.6
4.7
4.8
5.1
5.2
5.3
5.4
6.1
6.2
6.3
6.4
6.5
6.6
6.7
6.8
6.9
7.1
7.2
7.3
7.4
7.5
7.6
8.1
8.2
8.3
8.4
8.5
8.6
8.7
8.8
9.1
10.1
11.1
12.1
13.1
14.1
15.1
16.1
Get Involved
CourseFacilitateTake ActionSubmit Feedback
Textbook Resources
Table of ContentsDownload TextbookContent for Curriculum
Terms of ServicePrivacy Policy

Cookies Notice: This website uses cookies to identify pages that are being used most frequently. This helps us analyze data about web page traffic and improve our website. We only use this information for the purpose of statistical analysis and then the data is removed from the system. We do not and will never sell user data. Read more about our cookie policy on our privacy policy. Please contact us if you have any questions.

© 2024 Center for AI Safety
Built by Osborn Design Works