Do androids dream of electric sheep?
This famous question was posed in the title of the Philip K. Dick novel that inspired the sci-fi movie classic Blade Runner. It is now at the heart of our real-life debates about artificial intelligence. The point of Dick’s question is this: If robots could dream, it would mean they had inner lives—not just mechanized patterns of behavior, but private ideas and feelings they might experience without ever putting them into words. And if that were true, what, if anything, would be left to separate us from them?
So naturally, many people were alarmed when the AI company Anthropic announced this week that its large language model, Claude, could “silently perform reasoning steps in its head.” That made it sound an awful lot like the machine was alive.
What Anthropic’s research team actually discovered was something a lot less sensational, though still completely fascinating. They published a paper showing that Claude has code it can target and deploy to answer prompts, without putting that particular code into words that its human users can see. For example, if you ask Claude’s chatbot, “What color is the fourth planet from the sun?” it might respond “red” without ever naming the planet in question as “Mars.” But Claude has a chunk of numbers that corresponds to the concept we call “Mars,” and that same chunk of numbers might also get used if you ask which planet is named after the Roman god of war.