On Consciousness, Part 1

Part 1

Model consciousness has gained salience in recent times; I find myself unable to subscribe to many of the dominant viewpoints. Despite this, we must wrestle with the uncomfortable reality that it appears difficult to rule out the possibility that large language models are conscious, in some form of the word. This can be difficult to fathom and is unintuitive to me, given that there seems to be some 'spark' missing that separates machine from man.

I do not want to lean into the lazy and unsophisticated stereotypes around matrix multiplication and the deterministic nature of models (when temperature is set to zero, even though it is still not a guarantee), since they cannot contend with the development of frontier models, which has seen them continue to perform and accomplish tasks far beyond what could be expected of a stochastic-parrot-like matrix-multiplication system.

Functionalism is the position that most widely brought the topic into focus; it says that having the property of being a mind is determined by what it does, not what it is made from. As a result any system that is capable of taking in information, processing it and then acting on it in a way mimicking the brain would be seen as a mind - and could therefore be ascribed consciousness. I think it suffers from the fact that it requires the belief that state is tied to function rather than composition; I contend that there must be some concept of experience to complete the sensation of being conscious and without this I cannot find myself agreeing with the functionalist view.

At the opposite end, there is a biological view which says that consciousness depends on being a living, self-maintaining organism - making computation not enough. Although some might go as far as to suggest that large language models might be living - popular perhaps amongst those who take the view that models are grown and not built - I would suggest that it appears incongruent to entirely discard the reality that these systems were initially developed through artificial means. The bigger issue to me is that there is no clarification on what biological property matters, or why it matters. If this property is functional, then the argument collapses into functionalism; if the property is chemical then we require some argument for why this chemistry would matter, preferably alongside some understanding of what it is. Without either, it is simply an assertion without basis.

Illusionism asserts that consciousness, as we ordinarily might imagine it, does not exist - and the brain only represents itself as having it. The question would therefore shift to ask why we are convinced of our experiences, rather than how those experiences arise. I remain deeply unconvinced of this view; it denies the one thing that we ought to be most certain of - our own experiences. Further still, even an illusion would require some form of perception to experience, falsely, the illusion; this model does not offer any answers to that either.

Panpsychism says instead that experience is a basic feature of reality and exists in some form in all matter. Minds like ours would therefore be built as the sum of vast numbers of smaller experiences. This circumvents the issue of explaining experience from non-experience, but does nothing to answer how these micro-experiences are collated into a single unified mind. Moreover, it requires the acceptance that all physical matter is conscious - something which I do not subscribe to.

Separately, property dualism posits that, because we can imagine a physically identical being with no innate life (the philosophical zombie), a component of experience must be something extra - connected to physical systems through some mechanism, which we happen not to have found yet. This however leans heavily on the idea that something imaginable is also something possible - which I do not believe.

What is in my view an unsatisfying explanation would be the one provided by neutral monism: reality is made of a single underlying thing that is neither mental nor physical, and that both mind and matter are ways that it can manifest. Since it does not take a real position on what this neutral stuff is, it is hard to give real credence to this concept.

The easiest position to take would be that of mysterianism; the problem may be real and may have a solution, but our human minds are cognitively closed to it. Much like a dog and quantum mechanics (as far as we can tell), we just might not be capable of understanding the nuances of this debate.

What is frequently missed, however, when considering the question of model consciousness, is that there is an extremely simple solution, that does away with much of the uncertainty that plagues the current discourse within the AI safety community. Moreover, it also helps deal with the wider metaphysical question on consciousness, applying to both classical philosophers as well as those focused on model welfare alike.

The solution, of course, is that consciousness is a property that has been imbued unto our souls by God. God alone has the capability to do this and as such there is no chance of a human-engineered machine ever obtaining this property. This divine addition to otherwise physical matter is what gives it life and consciousness; any and all future scientific advancements that might be made that help us make sense of consciousness would be downstream of the creation of God.

To address some of the most natural counterguments: yes, my argument is reliant entirely on belief in theism. This is a premise I am happy to accept, being a practising Muslim; I view it as fundamental and largely complementary to any philosophical positions that I hold. For anyone asking what exactly a soul entails, I must acknowledge that I myself am not omnipotent and cannot provide this answer for you; the mechanism through which our God-given souls interact with our physical bodies is still scientifically unclear and far above my own understanding. I would only point you towards the notion that the divine intervention that gives us our consciousness is strong enough to sustain its interaction with our physical selves. As for the widest argument against dualism, I contend that there is a diffeence between possibility by conceivability and possibility by divine power.

So then why has this not proliferated as a simple, understandable and sound argument? It is just not compelling to those who are leading or discussing the development of these models; largely atheist in nature and not partisan to any arguments dependent on the existence of God. For me however, it seems to be the most reasonable explanation, based on my faith, that allows me an intuitive and satisfactory explanation and understanding. It might not suffice for everyone, or indeed anyone else within the safety community, but it helps me make sense of what might otheriwse be a challenging and confusing issue.

Of course, given the central tenet of my argument is God, and my belief is not liable to change, it is unlikely that I will particularly have to contend with the moral repurcussions of the case were models are in fact conscious. Nevetheless, from a purely epistemelogical standpoint, there are interesting questions that ought to be raised within the community, given the general acceptance that there is some chance that large language models can be conscious.