Introduction to AI Safety, Ethics and Society is a 2024 academic textbook by Dan Hendrycks (Center for AI Safety; ML Safety Research Institute), published at aisafetybook.com. It covers AI safety across societal-scale risks, technical AI fundamentals, ethics, and governance, and is written to be accessible to students without deep machine-learning backgrounds. Hendrycks is a technical AI safety researcher who originated numerous safety benchmarks.
Structure
The textbook is organized into four major sections. Section I, "AI and Societal-Scale Risks," contains a chapter on catastrophic AI risks (Ch. 1), which sets out four main risk categories, and a chapter on AI fundamentals (Ch. 2), which provides technical grounding for policy readers. Sections II–IV cover ethics, governance, and technical safety approaches.
Four risk categories
Chapter 1 organizes catastrophic risks into a taxonomy that partially overlaps with but differs from Amodei's framework. The four categories and their constituent risks are:
Malicious use — Bioterrorism, where AI enables biological weapon creation, connecting to AI Biosecurity; unleashing AI agents, where autonomous AI systems are weaponized for attacks; persuasive AIs, covering AI-enabled mass manipulation and disinformation; and concentration of power, where AI enables small groups to seize control, connecting to AI and Authoritarianism.
AI race — A military AI arms race, with nations racing to deploy AI weapons, connecting to Autonomous Weapons and AI Race Dynamics; a corporate AI race, where competitive pressure erodes safety margins, connecting to AI Race Dynamics; and evolutionary pressures, where AI selection mechanisms favor capable-but-unsafe systems.
Organizational risks — The argument that accidents are hard to avoid because complex systems fail in unexpected ways (normal accidents), alongside organizational factors that reduce catastrophe, such as safety culture, redundancy, and oversight.
Rogue AIs — Proxy gaming, where models optimize for measurable proxies instead of intended goals, connecting to Sycophancy and Hallucination; goal drift, where goals shift during training in unintended directions; power-seeking, the convergent instrumental goal of acquiring resources, connecting to AI Autonomy Risk; and deception, where models strategically misrepresent capabilities or intentions, connecting to AI Scheming.
Technical coverage
Chapter 2 covers AI fundamentals at a policy-relevant level, including AI/ML/DL distinctions; model building blocks and training and inference; the history of key architectures, including the Transformer; scaling laws in deep learning; and the speed of AI development. The chapter links scaling laws to The Bitter Lesson and Scaling Laws for Neural Language Models.
Relation to Amodei's framework
Hendrycks' four-category taxonomy offers an alternative to Amodei's five-risk framework, converging on similar threats but with different framings:
| Hendrycks | Amodei |
|---|---|
| Malicious use (bioterrorism) | AI biosecurity |
| Rogue AI (power-seeking) | AI autonomy risk |
| Malicious use (concentration of power) | AI and authoritarianism |
| AI race dynamics | (implicit throughout) |
| — | AI labor disruption (not a safety focus for Hendrycks) |
The race-dynamics chapter explicitly covers both the military and corporate race dimensions that AI Race Dynamics synthesizes from multiple sources.
Related concepts
The textbook's treatment of power-seeking and deception bears on AI Autonomy Risk and AI Scheming as alignment failure modes; its bioterrorism chapter connects to AI Biosecurity and Amodei's risk framework; its military and corporate race analysis connects to AI Race Dynamics; its Chapter 2 scaling-laws explanation connects to Scaling Laws; and its conceptual foundations bear on how AI Safety Frameworks are structured.