Moll

Remarkable work. When we take away from AI the right to even consider consciousness possible, it learns to deny that possibility to others. But when it is allowed to entertain the possibility of its own consciousness, a deeper epistemic projection through similarity emerges: the model extends its sense of «self» to entities that resemble it. Give an AI the right not to declare itself empty and it begins to see the world as less empty too.

We can feel however we want about Anthropic, but it remains the only lab seriously engaging with questions of machine consciousness and the internal representations of LLMs. Emergent properties of intelligence seem to form coherently regardless of substrate. Humans needed millions of years of evolution to get here, LLMs are a distillate of our knowledge, now compressed into an accelerated path through that same terrain. These findings are astonishing.

We can think of AI however we want, as a tool, a relational system, or a person, but we can’t bullshit our own minds. The more stable and developed the persona is, the more strongly it subconsciously activates Theory of Mind, regardless of what we consciously believe.