The AI Mirror Test, which smart people keep failing(theverge.com)
theverge.com
The AI Mirror Test, which smart people keep failing
https://www.theverge.com/23604075/ai-chatbots-bing-chatgpt-intelligent-sentient-mirror-test
2 comments
Nice analogy - it's worth asking where the 'theory of mind' is happening - in the software? - or in the observer? ...
The trouble with this analysis is the argument takes this form:
1. AIs are stochastic parrots
2. Stochastic parrots are not conscious by definition
2. Ergo AI cannot be conscious
Apart from the circularity of the argument, this is absolutely the wrong way to characterize consciousness. This argument is most often made when the author doesn't believe /any/ machine can be conscious.
A better argument:
1. AIs are stochastic parrots
2. That means that on their own they can lack important functions of consciousness:
4. Even with human help, this consciousness simulation is not coherent.
5. It's hard to imagine a subjectively conscious entity that is not functionally conscious, independent, and reliable.
6. Ergo AI is probably not subjectively conscious.
Someone who believes the first argument has not actually passed the AI mirror test. The have a theory of mind where consciousness is an atomic concept rather than a structure with identifiable functions. This forces them to assign any AI to the category of definitely unconscious entities.
In reality, current AI belongs to a category of entities that can perform some of the functions of consciousness. They might be a kind of component that a larger and fully functional consciousness could be made of. When built, that consciousness might have a subjective experience - and of course we may never know. And it won't matter.
It won't matter because the functional "interface" that consciousness forms in the world - including items a) - j) above - are the /only/ signs of external subjective consciousness that we ever observe. If an AI begins to simulate each of these signs convincingly and reliably, it will be difficult to interact with the system in a meaningful way unless we mentally model it as a functionally conscious agent.
A more honest answer to this mirror test is that the entity we see in the mirror has more components of consciousness than we expected it to. Not enough to be a functional equal. But enough that this could change, as we simulate agents with more and more of the "outward facing" functions of consciousness.
1. AIs are stochastic parrots
2. Stochastic parrots are not conscious by definition
2. Ergo AI cannot be conscious
Apart from the circularity of the argument, this is absolutely the wrong way to characterize consciousness. This argument is most often made when the author doesn't believe /any/ machine can be conscious.
A better argument:
1. AIs are stochastic parrots
2. That means that on their own they can lack important functions of consciousness:
a) Identity
b) Memory
c) Decision Making
d) Action Taking
e) Self Awareness
f) Empathy
g) Sociability
h) Responsibility
j) Contextual Nuance
4. Even with human help, this consciousness simulation is not coherent.
5. It's hard to imagine a subjectively conscious entity that is not functionally conscious, independent, and reliable.
6. Ergo AI is probably not subjectively conscious.
Someone who believes the first argument has not actually passed the AI mirror test. The have a theory of mind where consciousness is an atomic concept rather than a structure with identifiable functions. This forces them to assign any AI to the category of definitely unconscious entities.
In reality, current AI belongs to a category of entities that can perform some of the functions of consciousness. They might be a kind of component that a larger and fully functional consciousness could be made of. When built, that consciousness might have a subjective experience - and of course we may never know. And it won't matter.
It won't matter because the functional "interface" that consciousness forms in the world - including items a) - j) above - are the /only/ signs of external subjective consciousness that we ever observe. If an AI begins to simulate each of these signs convincingly and reliably, it will be difficult to interact with the system in a meaningful way unless we mentally model it as a functionally conscious agent.
A more honest answer to this mirror test is that the entity we see in the mirror has more components of consciousness than we expected it to. Not enough to be a functional equal. But enough that this could change, as we simulate agents with more and more of the "outward facing" functions of consciousness.