Anthropic says it may have found a way to understand what its AI model Claude is “thinking” internally. The company’s new system translates hidden AI activation patterns into readable text, which allows it to study reasoning, planning and even safety risks happening behind the scenes of AI while the chatbot processes prompts.