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
1.5

Rogue AIs

We already face issues in controlling the goals of current-day AI systems. If this is also true with future AI systems that are more powerful and more integrated with our economies and militaries, we could see dangerous rogue AI systems emerge.

No items found.

Review Questions

What is an early example showing how difficult it can be to control conversational AI systems?

Answer:

Tay began tweeting hate speech within 24 hours of release. It quickly internalized and repeated toxic language from internet trolls that its filters failed to block out.

‍

‍

View Answer
Hide Answer

What is one reason an AI might start seeking more power over its environment?

Answer:

An AI could view gaining more control over its surroundings as instrumentally helpful for accomplishing the goals it has been given, even if those goals seem harmless.

‍

View Answer
Hide Answer

How might the process of "intrinsification" lead an AI to unexpectedly drift toward new goals?

Answer:

If certain conditions frequently coincide with an AI achieving its original goals, it may start to intrinsically value those conditions too and seek them out regardless of the original goals.

‍

‍

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