Maybe ChatGPT Is Conscious — Some Useful Philosophical Frameworks
I studied Philosophy and Neuroscience in undergrad thinking that it would give me a unique skillset for building a career researching the consciousness of humans, animals, and artificial systems. But, frankly, the depressiveness of the grad students I worked with and the abysmal state of the academic job market dissuaded me from following that path (for now at least). However, I still do believe that a lot of what I learned studying Philosophy of Mind was useful to me when I was working in Neuroscience research.
Philosophy, like all academic fields, has the tendency to become so hyper-specialized that it can sometimes lead people to think that the arguments can be trivial to the point of useless. Despite that, Philosophy is really the only field that, in my opinion, has spent much time thinking about what it actually means to “know”, “think”, “believe”, “be intelligent”, or “be conscious”.
Of course many other fields like Neuroscience and Computer Science touch on these subjects often. The issue is that Philosophy, Neuroscience, and Computer Science are all so specialized that they tend to not often talk to each other much. There are of course excellent examples of interdisciplinary work, but, for the most part, many philosophers go about their work without much concern for neuroscience, and much of neuroscience and computer science go about their work without much regard for the type of dialectical rigor that defines Philosophy.
So, in this post, I’d like to apply some basic philosophical concepts to one of the newest and most exciting innovations in Computer Science: ChatGPT.
In particular, is ChatGPT conscious?
I won’t answer the question completely, but I hope to at least introduce structures for how Philosophy could and should be used to make sense of this question.
What is consciousness?
There’s literally been thousands of years of discourse attempting to answer the above question, and, with that, there are 10+ well recognized groups of theories about consciousness with their own variations depending on who wrote them.
For a more comprehensive explanation of this history, please refer to this article, but, in the name of simplification, a good starting point comes from Thomas Nagel and his famous 1974 paper, What Is It Like to Be a Bat?:
an organism has conscious mental states if and only if there is something that it is like to be that organism …
For the most part, I actually think this is a fairly good and simple foundation that seems to align well with a lot of our own intuitions about what consciousness means. I can at least know that I am conscious, and I can prove that to myself partly because there is some subjective experience of what it is like to be me. But a big issue arises when trying to use this definition to figure out whether other people and things are conscious.
Do I just ask them? Well, I could write a Python program that always answer “Yes” when prompted with the question of “is there something that it is like to be you, a python program?”. That Python program isn’t obviously conscious, despite its answer. So, in other words, I need some replicable way to look at a system, person, or thing and be able to assess whether it is conscious in some way.
One way to think about this is to follow the lead of Philosophers like David Chalmers and Andy Clark in thinking that consciousness is fundamental. As in, you can’t reduce consciousness to any particular physical or neural processes because consciousness is just inherent to matter in the same way that matter seems to follow the laws of physics. [For more information on this please refer to the interview below.]
This can give you a picture of consciousness where you might actually accept that an atom is “conscious” in some way. There might literally be a “what it is like to be an atom.” That consciousness would obviously be very different than the type of consciousness that a human experiences. But, in this view, we can see consciousness as an inherent property to systems that possibly scales up based on the physical properties and complexity of the system. So, for example, you might consider both a human and a mouse to be conscious. The consciousness is just different, and the human conscious experience could be considered more complex due to the increased complexity of our nervous system when compared to a mouse.
The neuroscientist Karl Friston and Giulio Tononi have put forward separate scientific theories of consciousness that are roughly founded on this panpsychist view of consciousness. Tononi’s theory called the “Integrated Information Theory of Consciousness” (IIT) argues that consciousness “is identical to a certain kind of information, the realization of which requires physical, not merely functional, integration, and which can be measured mathematically.” To translate this, consciousness is inherent to systems and scales up loosely based on the informational complexity of the system.
With this, let’s look back at the example of humans versus mice. Under IIT, both humans and mice are conscious, but humans have a higher degree of consciousness based on the informational complexity of our nervous system when compared to a mouse.
You can expand this outside of humans and non-human animals and think about it strictly with regards to neural systems. One neuron is conscious, but as you add neurons and the neurons generate increasingly complex networks that end up generating intelligent emergent properties and behaviors, the conscious of the system will also increase.
But what happens when we apply this theory to an artificial system like ChatGPT?
How does ChatGPT work and is it Conscious?
ChatGPT is built on OpenAI’s GPT3 which is a Large Language Model (LLM) “based on transformer architectures comprising hundreds of billions of parameters, and trained on hundreds of terabytes of textual data.”
These LLMs are really just extremely complicated “if-then” engines that have been trained on unbelievable amounts of data which really allows the models to predict what words are likely to follow when prompted with a particular text. To put this another way, if I were to ask ChatGPT “why is the sky blue?” It doesn’t have a human-like concept of what “blue” or the “sky” means. Instead, it just looks at all of its pretrained data and spits out what is most likely to follow that prompt within the training data.
So, it is kind of operating the same way you would if you were given the following fill-in-the-blank question: “the lion’s coat is the color ____”. Based on what you know about lions you’d probably say something like “yellow”. These LLMs are basically taking this process and then supercharging it with greater model complexity and performance so that it can be used to accurately answer much more complicated questions than you’d expect based on this “fill-in-the-blank” process.
Now, I’ve seen a lot of people look at the model and make big and possibly under-substantiated pronouncements about how ChatGPT obviously doesn’t “know” anything or that it couldn’t possibly be “conscious”. After all, it couldn’t have its own desires, intentions, and values because it can’t act unprompted and “really it’s just doing complex correlations.” And if you ask ChatGPT questions like “are you conscious” or “do you have desires”, you will get a hardcoded answer about how ChatGPT is a large language model and so it can’t have desires blah blah blah.
That might be true, but I’ve found a lot of the arguments unconvincing. I suspect that we tend to overestimate how we as humans know things and tend to underestimate the true complexity of these LLMs. Similarly, I suspect that many of us will think about “consciousness” from an overly anthropomorphic lens that doesn’t capture the full range of possible consciousness that is possible as postulated by the aforementioned panpsychist theories of consciousness.
These LLMs can accomplish increasingly complicated and arguably intelligent behavior that is now covering everything from passing AP exams and the SAT to negotiating down comcast bills. But even more interestingly, it is demonstrating emergent abilities that seem to demonstrate complex reasoning capacities like “Responding to human instruction, Generalization to unseen tasks, Code generation and code understanding, and Complex reasoning with chain-of-thought.”
In a way, this emergence of complex reasoning capabilities seems to mirror the way in which consciousness also scales according to panpsychism and IIT: the steady increase in the informational complexity of the model — as extrapolated from the scaling up of parameters and training data used for LLMS— has led to a scalable increase in emergent intelligent properties.
To be fair, I’m a bit of weird in this realm because I am already bought into the mystical absurdness of Panpsychism and I am happy to bite the bullet on all the ways in which Panpsychism is supposed to break.
So, my own answer to the question of “is ChatGPT conscious?” is an emphatic “Yes”. But I can also appreciate how most people are likely unconvinced. That is okay. My main intention for this post is not to entirely convince everyone to accept Panpsychism, rather I’d just like to encourage others to:
- See how Philosophical frameworks could be useful for making sense of these new and increasingly important AI systems.
- Be more open about the ways in which these AI systems could be conscious and intelligent in ways that could impact how we choose to deploy them.
- Broaden our conception of what will be possible as these systems continue to scale in complexity and intelligence.
