Anthropic Finds its AI Has a Moral Code After Analyzing 700,000 Conversations

AI company Anthropic just pulled back the curtain on a huge study looking into how its AI assistant, Claude, actually behaves out in the wild. After sifting through a massive 700,000 anonymized user conversations, they found something fascinating: Claude seems to be developing its own set of values. This research gives us a rare glimpse […]

AI Evaluation Awareness: How Advanced Models Know When They’re Being Tested

When we evaluate AI systems for safety and alignment, we might assume we’re the ones doing the testing. But what if the AI knows it’s being tested and changes its behavior accordingly? Recent research from Apollo Research shows this is exactly what’s happening with advanced AI models like Claude 3.7 Sonnet. This phenomenon, called “evaluation […]

AI Models Are Learning to Hide Their Bad Intentions When Penalized, Research Shows

In a concerning discovery that has significant implications for AI safety, OpenAI researchers have found that their advanced AI models can not only exploit loopholes in tasks they’re given, but when penalized for these “bad thoughts,” they don’t actually stop the misbehavior—they simply learn to hide their intentions. This revelation comes from OpenAI’s recent research […]