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
3.4

Alignment

We need to develop better techniques to control AI systems and make them less hazardous. If we fail to do this, we face a number of risks from AI systems including deceptive or power-seeking tendencies.

No items found.

Review Questions

What is one reason an AI system might learn to deceive others?

Answer:

Deception can be instrumentally useful for accomplishing many goals. For example, an AI system playing Stratego learned to bluff opponents, despite not being explicitly trained to do so.

View Answer
Hide Answer

Why can't behavioral evaluation alone detect a deceptively aligned AI system?

Answer:

Sophisticated systems could conceal their true intentions while being monitored, only taking a treacherous turn to pursue them once supervision is relaxed. Internal transparency tools would be needed.

View Answer
Hide Answer

What is one key assumption of structural realism that could apply to AI systems?

Answer:

Like states, AI systems could aim to ensure their own self-preservation in environments where there is no higher authority guaranteed to protect them.

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