If you’ve used chatGPT or other LLMs (Large Language Models), you will quickly realize that the knowledge and interactions distilled from these AIs are not quite what you would get from a human being. There are many missing pieces, but one that is crucial is the subjective experience machines seem to lack ( referred to as Machine Subjectivity ), and that is the topic for this overview.
At their core, subjective experiences are individual and mostly private events. Most of us know that fire is hot because we have experienced firsthand the heat from it. An LLM hasn’t felt heat or much else as it lacks heat sensors and thus derives the knowledge from a corpus of text where fire is mentioned as being hot.
We’ll come back to this point, but by their very nature, subjective experiences can generate different results from the same stimulus. Two people can go to the beach, be presented with the same stuff: sand, breeze, sun, bathing people, seagulls, the ocean, etc., and come to a completely different conclusion. For one, the experience was pleasant, and for the other, it was unnerving or even horrible.
So how do we even agree that the sky is blue if we all have a different subjective experience of the color blue?
Many ways, actually. Starting from the hardware, we more or less have the same sensory receptor types , instincts, and biological makeup. The differences in reality are minuscule ( yet we’ve fought wars over them! ), but more than anything, it is through communication that we reach a consensus over what things are and how they should be called. The rich language we have developed has allowed us to reach this shared subjectivity or intersubjectivity.
Science, culture, and social constructs can be seen as advanced forms of
developing intersubjectivity through communication for different reasons,
like clubs with shared objectives, specific language, customs, rules,
A subjective AI ?
So, the next issue before tackling the possibility is whether we need an AI with subjective experiences, which would…