[weinzierl]

"This suggests that the CASA Theory may only work for emergent technology[..]"

I think we can see this unfold in real time with large language model based chatbots like ChatGPT. First they seem almost human and the initial reaction is to treat them like that. Always say "Thank you!" and potentially get angry at them for being wrong. It doesn't take long though, to realize that the bot is a bot and even if it speaks human language it behaves significantly different from a real human. Then the human behavior starts to change as well and the bot is treated differently.

[pixl97]

Except we see to see behaviors in asking questions to LLMs where we get different/better responses if we ask "please".

[chihuahua]

And also, we get better answers from LLMs for solving captchas if we claim that the hard-to-decipher letters were written on our dear grandmother's Christmas ornament. I find this quite amusing.

[l33t7332273]

This makes sense if you think about chatGPT as a ML model and not as a sentient AI. Its training data would show that asking nicely elicits better answers.

[pixl97]

I mean should I think of my mother as a ML human and not a sentient one? Why does asking nicely work on either?

[l33t7332273]

If I set up a linux box with the root password “please” I don’t think you would describe that as sentient.