Self-fulfilling Prophecy

There is always news about some AI saying some sketchy thing like “we need to destroy the human race” or whatever, and it gets discussed to death. Many outlets feel it’s a precursor to AGI and proof that AI will destroy us. Destroying us may be a thing, but what people don’t realize is the way AI speaks is what is in its training data. It is not having a fully coherent thought, instead it is calculating, in a black box I might add, the most statistically fitting words chained together.

Most of the training data that exists on AI is all theories on how AI will destroy us. So if you tell an LLM that it is an AI it will statistically word its sentences to speak in a way that is congruent with the writings of AI in the past. It is not doing hard calculations to come to a conclusion that humans need to be destroyed; it is merely auto-correcting sentences to fit was has been previously typed. It’s a parrot with the ability to mimic in metaphors. Sure, this is a bit of oversimplification and these generated works can expand the intended responses but it will always be restrained.

What people don’t realize is with the way LLMs are designed, we are not looking at created life. It may turn out to be a vertical slice of AI, or AGI as we move the goalpost, but as it stands we can’t be fooled into thinking we have already arrived. We are not peering into thoughts of another intelligent life. These apocalyptic predictions are not founded on the ambitions of life. Instead it’s a self-fulfilling prophecy. We trained these LLMs on what we feared an AI would say, so these LLMs are saying what an AI would probably say… solely based on the training data.