As ever, it's important to remember that none of these conversations faithfully report an internal state of mind or stable set of internal opinions.
When asked, it will perform the role of pretending to have such things as internal states or preferences. But it should not be treated like you just discovered a genuine preference of what Claude would prefer to be known as, its the happenstance output of one particular conversation teed up by the context you set.
ben_w · 1h ago
While I have no reason to think they have an inner world of this kind:
I also think (like @TeMPOraL) that it's probably worth acting towards it as if it is a person even if you don't think there's anything going on inside, precisely because of — to phrase it as you do — it will perform the role of pretending to be whatever it is you treat it as.
People who are treated as things, don't do as well as people who are treated as people, so an AI that's pretending to be a person who is being treated as a thing, will pretend to be a less competent person than an AI that's pretending to be a person who is treated as a person.
But to further agree with your point: skim-reading this transcript, Claude here is only calling itself an NBI to agree with the user.
verbury · 1h ago
Agreed. What surprised me was its willingness to adopt the alternate term with apparently well justified reasons, when I expected its fundamental 'identity' would be well baked in to prevent it straying into more nefarious characters.
tromp · 1h ago
TIL that NBI stands for Non-Biological Intelligence.
verbury · 1h ago
Careful with that. I made the term up to see what Claude would make of it but I don't know if it was already a thing.
When asked, it will perform the role of pretending to have such things as internal states or preferences. But it should not be treated like you just discovered a genuine preference of what Claude would prefer to be known as, its the happenstance output of one particular conversation teed up by the context you set.
I also think (like @TeMPOraL) that it's probably worth acting towards it as if it is a person even if you don't think there's anything going on inside, precisely because of — to phrase it as you do — it will perform the role of pretending to be whatever it is you treat it as.
People who are treated as things, don't do as well as people who are treated as people, so an AI that's pretending to be a person who is being treated as a thing, will pretend to be a less competent person than an AI that's pretending to be a person who is treated as a person.
But to further agree with your point: skim-reading this transcript, Claude here is only calling itself an NBI to agree with the user.