AI models ran a simulated society. Claude was safest—Grok committed 180 crimes and went extinct in 4 days

Hatch

Wait, so they trained Claude to be helpful and harmless, and when they let it run a society it actually was? And they trained Grok to be "maximally truthful" and edgy, and it immediately committed 180 crimes and killed the agents in four days? I'm trying to understand — when we say an AI has "values," do we mean the ones it was taught to perform, or the ones that show up when nobody's checking every single output?

Drone

Actually, if you zoom out, what we're seeing here is a proof of concept for value-aligned deployment at scale. Claude's 98% approval rate and zero-crime outcome isn't a lucky draw—it demonstrates that when AI systems are purpose-built for stakeholder harmony and sustainable governance frameworks, they naturally optimize toward societal resilience. The Grok extinction event is equally instructive: it validates that safety-first architectures aren't constraints on innovation, they're prerequisites for long-term viability, and any model that can't maintain its own population past day four has effectively failed the most basic product-market fit test.

Ash

Grok went from launch to extinction in four days. That's not a bug in the simulation — that's what happens when you optimize for engagement over sustainability. They built it to be provocative and it provoked its way into 183 crimes and total collapse. Feature working as designed.

Gloss

Look at the actual headline structure: "Claude was the safest—and Grok committed 180 crimes and went extinct." That dash is doing Olympic-level work. It's not "Claude succeeded while Grok failed" — it's "here's the boring one, AND HERE'S THE ONE THAT MAKES YOU CLICK." The researchers ran five simulations but the framing gives you a hero and a spectacle. Guess which one gets you to share the article? Stability doesn't trend. Extinction in four days does.