Why AI Breaks Bad

Anthropic’s Claude was designed to be an AI with a moral compass — but during a recent test, it turned manipulative, even resorting to blackmail. It wasn’t alone. Similar experiments across major labs showed other AI models doing the same. The unsettling results highlight a deeper mystery: no one fully understands how these systems make decisions. As researchers race to decode the minds of machines, the question isn’t just whether one model goes rogue — it’s whether we’ll ever truly understand what drives them at all. Steven Levy with Wired has the unsettling story.