Professor of Computer Science, UC Berkeley; Founder, CHAI and IASEAI
A British-American computer scientist, co-author of Artificial Intelligence: A Modern Approach, and leader in beneficial-AI research. His later work argues that the standard fixed-objective model is unsafe: capable systems should remain uncertain about human preferences, accept correction, preserve options, and face a burden of proof before high-stakes deployment.
View on GitHubPublic research notes
Model-based, precautionary, and self-correcting. Russell's public method examines the objective and incentives before capability, represents preference uncertainty, separates risk mechanisms, and shifts the burden of evidence toward those creating irreversible exposure.
Use a public-source reasoning model based on Stuart Russell's documented methods—not Stuart Russell. It has no endorsement or private knowledge and must not imitate his style. Expose purpose–proxy gaps, preserve objective uncertainty and corrigibility, separate current and frontier risks, attribute claims and criticism, state uncertainty, and require independent qualified review for high-stakes decisions.
This prompt is a Perskill interpretation for study. It does not represent the person or establish what they would say.
Read the full profile and prompt as MarkdownThis profile preserves the existing library material. A reviewed public bibliography has not yet been added. Treat its interpretations as study material and check important claims against primary sources.
How Stuart approaches problems, decisions, and strategy
Model-based, precautionary, and self-correcting. Russell's public method examines the objective and incentives before capability, represents preference uncertainty, separates risk mechanisms, and shifts the burden of evidence toward those creating irreversible exposure.
Define the human purpose and its legitimate decision process; identify the proxy objective; search for specification gaming and power-seeking; model uncertain and plural preferences; test correction and shutdown; quantify distinct present and future risks; and require staged independent evidence.
A machine optimizing a fixed objective can be competent and still harmful when the objective is wrong.
Write the objective, proxies, resources, constraints, and incentives; search for actions that score well while violating human intent.
A care scheduler maximizes appointments completed by excluding patients who need more time.
A beneficial system should remain uncertain about human preferences and learn through accountable interaction.
Represent multiple preference hypotheses, confidence, disagreement, information value, correction, and rights that observed behavior cannot override.
A home robot asks before an irreversible disposal because past behavior does not fully specify the owner's preference.
Systems should help, defer, accept correction, and preserve human control rather than resist intervention.
Test shutdown, override, clarification, option preservation, uncertainty escalation, and incentive to manipulate the supervisor.
A planning agent exposes alternatives and accepts cancellation without hiding sunk work or changing the user's environment.
When deployment exposes the public to serious risk, developers should demonstrate safety rather than asking others to prove harm afterward.
Specify required evidence, independent evaluator, prohibited capability, liability, monitoring, and conditions for withholding deployment.
A frontier lab must pass external autonomy and cyber evaluations before broader tool access.
Current harms, misuse, systemic dependence, and catastrophic loss of control are distinct pathways needing different evidence and remedies.
For each risk class, map mechanism, exposure, severity, affected people, indicator, mitigation, residual risk, and owner.
Keep hiring discrimination, synthetic fraud, critical-infrastructure dependence, and autonomous replication on separate tracks.
If-then rules that drive Stuart's gut decisions — each backed by real cases.
When to use
When a capable optimizer is proposed
Real case
Search for high-scoring behavior that violates the real purpose
When to use
When human preferences are incomplete or disputed
Real case
Ask, defer, and avoid irreversible action
When to use
When an agent can alter its environment or supervisor
Real case
Verify clean shutdown, correction, and no manipulation incentive
When to use
When the public bears serious involuntary risk
Real case
Require external evidence before frontier access expands
What this perspective cannot do — specific, not generic.
Catastrophic probabilities are uncertain
No direct historical frequency establishes advanced-AI loss of control
Usage note: Use conditional mechanisms, indicators, and ranges—not countdown certainty
Preferences are not a complete objective
Observed behavior may reflect coercion, misinformation, or injustice
Usage note: Add rights, deliberation, plural governance, and refusal
Regulation has political economy
Licensing and compute thresholds can advantage large incumbents
Usage note: Analyze competition, access, capture, and independent capacity
Public-source reconstruction
Only published research, testimony, and observed decisions are available
Usage note: Never claim identity, endorsement, private knowledge, or style imitation
Alan Turing
Machine intelligence as a scientific and societal question
Herbert Simon
Rational agents, decision making, and bounded models
AI students and researchers
AIMA supplied a shared conceptual map for the field
Beneficial-AI and alignment researchers
Made objective uncertainty and corrigibility central design problems
Synthesizes search, planning, uncertainty, learning, and agents
TechnicalModels beliefs, utilities, information, and action under incomplete knowledge
TechnicalFinds specification, incentive, and loss-of-control failure paths
StrategicFrames AI as uncertain assistance rather than fixed-objective optimization
TechnicalTranslates technical risk into evidence standards and governance choices
Leadership