Breaking

Stuart Russell Explains Why Our Instructions for AI May Be the Problem

October 5, 2026 at 15:18 EDT

Foundation Models
AI Agents

The fundamental challenge in AI alignment isn't finding the right way to follow instructions—it's that the instructions themselves may be flawed. In a new episode of The Information's AI Deep Dive series, UC Berkeley computer science professor Stuart Russell explores why teaching AI systems to pursue human-defined goals can backfire, and why robust alignment requires machines to learn through uncertainty rather than blind obedience.

AI Alignment Problem

Why Instructions Alone Aren't Enough

Advanced AI systems pursuing literal interpretations of human goals can violate core values. The challenge: ensuring machines learn and refine their understanding of what we actually want.

The Current Approach

Reinforcement Learning from Human Feedback (RLHF): AI learns from human preferences and comparisons. Effective for today's models, but breaks down as systems grow more capable than human evaluators.

Core Principles of Russell's Framework: uncertainty about preferences, active learning, multiple values

∞

The Extended Frame Problem: any fixed value encoding will mismatch contexts advanced AI creates

The Specification Trap

Hume's Is-Ought Gap: Behavioral data never fully determines what values should be.

Value Pluralism: Human values resist consistent mathematical formalization.

Capability Scaling: Static value systems become misaligned as AI reaches new domains.

Russell's Vision: Move from obedience to cooperative understanding—AI systems that recognize inherent uncertainty in human values and treat value discovery as an ongoing process, not a solved problem.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.