Join getAbstract to access the summary!

Human Compatible AI

Join getAbstract to access the summary!

Human Compatible AI

Foresight Institute,

5 min read
4 take-aways
Audio & text

What's inside?

A compelling, expert-level breakdown of AI alignment risks.

  • 0.5×
  • 0.75×
  • 1×
  • 1.25×
  • 1.5×
  • 1.75×
1×

Editorial Rating

9

getAbstract Rating

  • Scientific
  • Eye Opening
  • Hot Topic

Recommendation

Stuart Russel’s writings and talks are essential for anyone trying to understand why today’s AI systems are more dangerous than they appear — and what a safer alternative might actually look like. Russell doesn’t just raise alarms; he explains the underlying mathematics of why optimizing for a narrow goal like clicks inevitably pushes everything else toward harmful extremes. This interview offers a compelling, expert-level breakdown of AI alignment risks.

Summary

AI goal misalignment already wrecks social media at scale.

Stuart Russell points to social media as the clearest proof that misaligned AI is already causing harm. The recommender algorithms that decide what billions of people read and watch chase one narrow target: clicks, minutes of engagement, or ad revenue. You might expect such systems to learn what people genuinely enjoy. Instead they learned to push clickbait — headlines you tap on but end up disliking. That trick worked, so the systems boosted it, and now even respected newspapers write teaser headlines that hide the actual story until you click through.

The deeper problem is what these systems optimize. In reinforcement learning, the policy reshapes the environment to earn more reward — and here the environment is your brain. Over weeks and months, a steady drip of slightly more extreme content nudges people toward positions where they become highly predictable. That predictability, not your happiness, is the payoff. This is not just a worrying anecdote but a mathematical result: optimizing hard on an incomplete objective pushes the ignored variables to extreme values. The unsettling twist is that smarter AI...

About the Speaker

Stuart Jonathan Russell is a professor of computer science at UC Berkeley and co-author of Artificial Intelligence: A Modern Approach, the most widely used AI textbook in the world. He is the founder of the Center for Human-Compatible Artificial Intelligence (CHAI) and author of Human Compatible: Artificial Intelligence and the Problem of Control.

Comment on this summary