Watch on YouTubeCurrent large language models are not conscious, but machine consciousness is not impossible in principle.
Consciousness involves subjective experience from a particular location and perspective, including perception, memory, anticipation, and a model of how actions affect the self.
An artificial system might become conscious if it develops an embodied, predictive model of itself and the world, even if biology eventually proves necessary.
Ron Chrisley rejects both claims that machine consciousness already exists and that it is impossible. He examines the argument that language models understand because they produce language like humans, then argues that linguistic performance alone does not establish consciousness. He defines consciousness as subjective experience anchored to a point of view in space and time. This experience includes perception, memory, expectations, possibilities, and a model of how the system's actions affect what happens to it. Current LLMs lack this embodied perspective, so scaling them with more data and parameters will not automatically make them conscious. Chrisley leaves open the possibility of artificial consciousness in a system that learns a predictive model of its embodied self and world. He also rejects the claim that consciousness must be biological, while allowing that biological components could eventually prove necessary. His answer is careful: embodiment may be required for the kind of consciousness humans can recognize.
Chrisley opens by rejecting both extreme positions. Current AI systems are not conscious, but machine consciousness is not impossible. He argues that examining the middle ground gives us a better account of consciousness and clarifies what AI might become beyond current systems. The talk treats this as a philosophical and technical question rather than a claim that present-day models are secretly aware.
One argument for conscious LLMs says that they understand what people say because they produce linguistic behavior like human understanders. Chrisley questions both parts of this argument. A model may pass a conversational test for a limited period and later reveal that it never understood the discussion. However, he says this failure mode is becoming a weaker objection as models get better at sustaining conversations.
Chrisley discusses Emily Bender and her team's description of LLMs as "stochastic parrots." The objection says that models generate responses from token statistics rather than from a model of the world. Chrisley agrees this raises a deeper problem, but says it does not settle the matter. A statistical map of language could involve implicitly constructing a model of the world that produced the language. Showing that a capable LLM lacks such a model would require more work.
Chrisley defines consciousness as having subjective experience, rather than merely answering questions correctly. Subjectivity is anchored to a location in space and time and to a perspective. Experience includes perception, memories of the past, expectations about the future, and representations of what could happen. It also includes a model of how the system's own actions affect what happens to it.
Chrisley says current LLMs are not conscious because they do not have the kind of embodied, here-and-now perspective found in animals such as parrots. Giving them more data and parameters will not automatically change that. He distinguishes this criticism of current LLMs from a criticism of transformers in general, since attention-based systems may still help build the models needed for artificial subjective experience.
For an AI system to have the relevant kind of understanding, Chrisley says it would need a location and perspective in the world, perhaps through a robot body. It would learn what happens when it takes actions or when other events occur, including what happens to the self. An attention-based system could help learn this model from large amounts of data, although the talk does not claim that such a system exists today.
Chrisley calls it a fallacy to infer that only biological systems can be conscious simply because known conscious systems are biological. The conditions he identifies, including perspective, subjective experience, and a computational model of the embodied self, do not inherently mention biology. He still allows that biology might eventually prove essential, in which case hybrid biological and mechanical systems could become relevant.
In the follow-up, Chrisley says a body and a location in space and time will be necessary for something humans can recognize as conscious. He allows that forms of consciousness might exist that people cannot recognize. For the kind of consciousness humans can identify, however, he expects an embodied agent that experiences and acts on the world from a particular position.
"Calling ChatGPT a stochastic parrot is actually an insult to parrots because being a mere stochastic parrot is so much less than what a real parrot does."07:43