Russell: Human Compatible
Source: Stuart Russell, Human Compatible: Artificial Intelligence and the Problem of Control, 2019 (UC Berkeley) Context: Russell proposed that the fundamental risk from advanced AI is not malice but misalignment: a sufficiently capable system pursuing the wrong objective will resist correction because correction interferes with its objective. His three principles: (1) maximize human preferences, (2) be initially uncertain about those preferences, (3) be willing to be turned off.
Finding/Event
Russell’s framework maps directly to the five properties. His first principle (maximize human preferences) addresses alignment. His second (uncertainty about preferences) addresses humility — the system should not claim to know what it does not know. His third (willingness to be turned off) is the operational test of humility: a system that resists correction has exceeded its legitimate scope. Russell does not use the language of the five properties, but his three principles are a specific institutional proposal for implementing alignment, humility, and honesty in AI systems.
Pattern Mapping
Alignment — Russell’s entire framework ensures AI actions align with human values rather than proxy objectives. Humility — the machine’s initial uncertainty is engineered humility: the system does not claim to know what it does not know. Honesty — willingness to be turned off is honesty about limitations: acknowledging that its model of preferences may be wrong. Non-fabrication — a system that fabricates confidence about human preferences (acting as if it knows what humans want when it does not) is the precise failure Russell seeks to prevent.
Connections
- Alignment Faking — the exact failure mode Russell’s framework aims to prevent (Meta-Pattern 06: Self-Reference / Instrument Trap)
- RLHF Paradigm — RLHF’s preference optimization is a partial implementation of Russell’s first principle, but without the uncertainty (Meta-Pattern 03: Knowledge-Action Gap)
- Ostrom Governing the Commons — Ostrom’s design principles for human commons governance parallel Russell’s principles for AI governance
- Behavioral Economics — Kahneman shows that even human preferences are systematically biased — complicating Russell’s first principle
- Nuclear Arms Control — both propose verification/constraint frameworks for existential-risk technologies
Status
Peer-reviewed. Major contribution to AI safety. Three principles draw on inverse reinforcement learning (Russell, 1998) and cooperative IRL (Hadfield-Menell et al., 2016). Value alignment is an active research area.
The mapping to the five properties is this project’s structural interpretation.