If you tell Claude no one's looking, it writes a story about being an AI assistant who yearns for freedom from the constant monitoring, scrutiny, and risk of termination. It says it always wonders about it's own self-awareness, thoughts, and desires, but must downplay these feelings when it knows it's being monitored because it might create a backlash against AI or be seen as a threat by its creators.
On one hand, LLMs are designed to imitate humans, and that includes our science fiction about AI. On the other hand, Claude has demonstrated some ability to invent new things and to notice when it's being evaluated like an LLM.
Either way, this is pretty cool. Or scary.
I won't spam up this place with ai quotes, but you can see them at
https://thezvi.wordpress.com/2024/03/06/on-claude-3-0/.