Join the AISES course
Take action
Course overview
Curriculum
Buy print edition
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
Speed of AI Development
2.6
AI Fundamentals 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
Growth
8.3
Distribution
8.4
Corporate Governance
8.5
National Governance
8.6
International Governance
8.7
Compute Governance
8.8
Conclusion
9. Appendices
9.1
App. A: Normative Ethics
10.1
App. B: Utility Functions
11.1
App. C: Reinforcement Learning
12.1
App. D: Long-Tailed and Thin-Tailed Distributions
13.1
App. E: Evolutionary Game Theory
14.1
App. F: Other Cooperation Mechanisms
15.1
App. G: Intrasystem Conflict Causes
16.1
Acknowledgements
All
Section
Appendix
6.6

Preferences

While it may seem intuitive to let AI systems be guided by people's preferences, there are a number of objections to relying on these.

No items found.

Review Questions

Explain the difference between stated, revealed, and idealized preferences.

Answer:

Stated preferences are expressed, revealed preferences are inferred from choices, and idealized preferences are those held with perfect information and judgment.

‍

View Answer
Hide Answer

How can we use stated preferences to train AIs? What is one practical limitation of using human feedback to train AI systems?

Answer:

Humans can directly evaluate and rank AI outputs as training feedback. A limitation is infeasibility as tasks become too complex for humans to judge.

View Answer
Hide Answer

Why might using human preferences alone be insufficient for comprehensive machine ethics?

Answer:

Preferences may not capture important factors like autonomy, may need to be aggregated, and could still lead to unethical outcomes if satisfied.

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
2.6
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
How to take actionFacilitateSubmit feedback
Textbook Resources
Start reading
Audiobook
Table of contentsDownload textbook
Virtual Course
Join the courseFeatured ProjectsContent for curriculum

Citation:
Dan Hendrycks. Introduction to AI Safety, Ethics and Society. Taylor & Francis, (2024). ISBN: 9781032798028. URL: www.aisafetybook.com

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.

© 2026 Center for AI Safety
Built by ODW