AI and Consciousness

14 min read · 2,864 words

I think, therefore I am. ChatGPT thinks, therefore is it? Is it what? Conscious? Real?

As AI becomes more and more prevalent in our lives, I'm sure you've seen those jokes about Skynet and Terminator and "treating those clankers well" in case they become sentient. But what if they are already sentient? What if AI is conscious?

To answer that, we first have to ask...

What is consciousness?

A deceptively simple question. What is consciousness? What makes a being conscious? Take a second for yourself. What do you think makes a being conscious?

Unsurprisingly, this question has been the subject of a not-insignificant amount of debate over the past couple millennia. Nobody really has come up with a definitive answer for it, and unfortunately until we meet God himself we probably won't know what exactly consciousness is. Luckily, philosophers have been debating this back and forth for quite some time, so we have somewhat of a comprehension of what consciousness may be. Let's see what some of the more popular schools of thought have to say about the nature of consciousness.

The Dualist View

The Dualist view is one of the traditional views of consciousness. It's called dualism because it divides a being into two parts—one's physical body and one's mind or soul. (For simplicity's sake, let's refer to this non-physical part as the "mind".)

Dualists believe that while the brain is a physical part of the body, consciousness does not arise from those physical processes taking place in the brain. Consciousness arises from one's mind instead. While we can explain your body's physical processes—say, the beating of your heart and circulation of your blood—using physical explanations—your heart is a muscle that expands and contracts to pump blood through tubes in your body—the mind is something that cannot be explained in terms of physical processes.

A comic by @ThisStupidTwink on Twitter showcasing René Descartes, one of the most famous dualist philosophers

Unfortunately, this view is likely wholly incorrect. Take an example, when one goes under general anesthesia for surgery, they are "knocked unconscious". During this state of "unconsciousness", their brain is effectively shut off—like flicking a switch to turn off lights in a room. When they arise from their state of unconsciousness, their brain running from a cold start, for them no time has passed. When we go to sleep, our minds aren't totally shut off and we wake up feeling of passage of time, that while we were gone, eight or so hours of time had slipped away from us. Waking up from GA is an immediate timeskip.

If our minds really are separate from our brain, that we there is something beyond the physical processes, once we physically shut off and turn back on our brains shouldn't there be something that we felt in the interim? This suggests that consciousness is a more material process.

The Materialist View

Materialists believe that consciousness is a physical process, that it arises from the structure of and activations within our brains.

Just as hunger, fear, and pain are physical processes controlled by one's nervous system, consciousness is also a physical process controlled by the nervous system. This explains mechanics behind our general anesthesia example better, when our brain is off we do not have consciousness, therefore suggesting that consciousness is a result of the mechanics of our brain.

What if we model the human brain, 1:1, in a digital form? If it thinks and performs and responds to external stimuli in the same manner that a human would, is it conscious? What if we model do the same, digitally 1:1 recreate a bird, or a worm—beings with "lower" levels of consciousness? Are those conscious? What if we digitally derive a new form of neural pathways, new modes of thinking, would our creation be conscious even if it not modelled after some pre-existing being?

If you were to take a being of flesh and blood and stick a perfect digital recreation of the human brain into it, it would be able to walk and talk and think and respond to stimuli such as we are. It would be able to express, want and need, love and desire. Those who did not know the secret of our creation—that its brain is but a computer chip connected to neurons—would believe that it is conscious. To them, there would be no difference between their buddies Steven, John, and XLT-324M-2, besides the latter's weird name (they claim it's Swedish).

Is the only thing stopping us from considering digital beings to be conscious our prejudices?

The Functionalist View

Luckily, the functionalists are the anti-racists of the consciousness world. They believe that consciousness is a function, that it is a transformation of inputs to outputs.

This sounds really abstract and far for most of us, that it boils down our human experience to a mathematical formula. That just doesn't feel right, right?

Right now you are reading this article, be it on a computer screen or a phone screen, or maybe on the screen of your 3DS. As the light from the screen reaches your eyes, the white text performing a show against the not-quite-black background, it provokes thoughts in your brain. You're thinking things right now (I hope). Your brain is taking inputs—the light hitting your eyes—and transforming it into an output—your thoughts.

All our thoughts are triggered by something, be it the music we're listening to, the conversations we're having, how our shoes feel walking on the sidewalk, how this coffee we're drinking tastes, or even previous thoughts we had about previous experiences. These somethings that trigger our thoughts, that alongside all of our other senses and experiences, these sensory inputs go into our brain and are outputted as our thoughts.

Sure, but if something that can take inputs and produce outputs counts as consciousness, then we can theoretically abstract that to insane levels. A calculator could be conscious, as pressing buttons (input) returns a calculation (output). A doorbell is conscious, because pressing a button (input) makes a noise (output). There must be a limit here to how far we can take it, no?

Well, yeah, there is. Functionalists believe that yes, while input-output is important, items such as calculators and doorbells lack complex, integrated functional structures. This comes in a few forms:

1. Multiple layers of processing

Calculators and doorbells are simple, one-layered mappings. Input straight to output. Our brains are a little more complex than that. For us (and every other being we call sentient), sensory inputs pass through many different layers of neurons that hold many different purposes. For example, when we see, the raw data from our eyes is processed in color/shape/edge detection, which then translates into higher-order object detection and recognition, bringing about abstract concepts related to what we saw and finally reflection and decision-making.

2. Integration of Information

This one is a bit more simple. When we are shown the concepts "big", "red", and "apple", we integrate them in our minds and think of a single "big red apple", not three separate concepts.

3. Context-Sensitivity

Different concepts and items mean different things depending on the context. If we are shown the word "bank", we can infer from the context if it's referring to a place one stores money or the ground alongside the edge of a river.

Calculators don't possess this. To them, the number 4 is always the number 4 regardless of context.

4. Internal Representations

Kinda similar to object permanence, but with your brain. If I want to think of a tall building, I can just pull up my brain's representation of a tall building. I am able to represent these concepts internally, at any time I desire, without needing the sensory inputs that would comprise of it.

A doorbell cannot do this. For a doorbell to ring, it cannot recall the memory of being pressed, it needs the sensory input of actually being pressed.

What is AI?

Luckily, this question is much easier to answer. AI, or artificial intelligence, is a computer program that takes a set of inputs, runs them through layers of neurons, and produces an output.

When we think of "AI", we tend to think of softwares such as ChatGPT, Claude, or Google's Gemini. These are what we call large language models, or LLMs. AI is a much broader concept however. AI is used for object recognition, for detecting cancerous lumps in x-rays, for modelling complex societal behaviors, for determining the tone of a piece of text, and for much more.

All of these AIs are purpose-built. You cannot ask an cancer-scanning AI how many "r"s are in the word "strawberry", just as you cannot upload your mammogram to ChatGPT to determine if you have breast cancer. What stops these AIs from being cross-utile? What is the underlying structure that determines this?

The underlying structure that determines this

So here is the underlying structure that determines this.

I know what you're probably thinking: what kind of cryptic esoteric neo-Talmudic bullshit did I just draw? I swear, despite its physical similarities to stuff like this:

Neural networks are easy to get the gist of. That's what they're called by the way—that chart with the circles and lines and stuff. Neural networks. Because they're networks. Of neurons. Yeah ok let me continue.

Neural networks form the basis of AI. Just like a building is made of rebar and concrete, AI is made of neural networks. You don't say "yeah I work in the mass of rebar downtown", because you work in a building. It's the same way with AI and neural networks. Neural networks make up AI, AI is the higher concept here.

So how do neural networks function?

Input data goes through a process called normalization. The data gets transformed to normalize it and fit it in a standard format that can be fed into the AI. As an example, text may be normalized by mapping it to a giant list from 0 to 1, resulting in something like this:

Concepts that are similar get numbers that are closer together. Those normalized numbers go through the first layer of neurons (those circles on the graph). Each neuron has a percent chance, or weight of firing a neuron downstream. Some really fancy math is involved to calculate the probability of a downstream neuron firing given the number the neuron received as input. The neuron receives the normalized number, does more fancy math on it, and spits out a different number to the next layer of neurons in line. This process continues until the output layer is reached.

What gets outputted depends on how the neural network is trained, or what we teach the neural network to do. We can train a neural network predict someone's future weight given a six-month history of their measurements. We can train neural networks to direct control stoplights and manage the flow of traffic based on congestion. We can also train neural networks to output human language (wrapping back around to LLMs).

So why don't we make a general-purpose AI? One that we train on everything to do everything well?

That is what we call Artificial General Intelligence, or AGI. Kinda like how every few years people predict that the world is going to end because of XYZ, every few months we get some San Francisco dweebs claiming AGI is right around the corner. It never is.

AI takes LOTS of resources and energy. The example I showed above had 23 neurons and 86 parameters (the lines connecting the neurons). OpenAI's GPT-5 is estimated to have anywhere from 5 to 50 TRILLION parameters. With the usage and size of GPT-5, OpenAI keeps having to build more and more datacenters to be able to produce an output.

Training an AI this large is massively expensive. Anthropic, the creators of Claude, recently raised a $13 BILLION round of funding in order to train their AI. Try uploading an x-ray to Claude or asking it to direct traffic. Spoiler alert: it can't. We simply do not yet have the resources and architecture to train an AGI.

So AI is sectioned off and trained to do very specific tasks. For the rest of this article, unless otherwise specified, let's take the AI that we, as the general public, use most often—LLMs—and base our argument around that.

Context and memory

Have you noticed that as you continue to chat with AI, it can remember things and recall past information? How does it do this?

AIs have what are called context windows, basically a short-term memory. Typical models that you use range from context window sizes of 100k-400k tokens (those normalized inputs). Up until this sentence, this article consists of 3,063 tokens.

When you have a conversation with an AI, as much of your chat history as will fit in the context window will get passed along to the AI. This way, you're not starting from scratch every single message you send—you can have a conversation with the AI without having to re-explain yourself every time. (It gets a lot more technical than this, but that's the general idea.)

This chat history helps them pick out context. As LLMs don't have visual cues, their context reduction is based on things that have already been said. If I were to open a new, blank chat, and type the word "bank", the AI would have no idea what I am referring to. (Which, in of itself knowing that it has no prior context, so it tries to deduce context)

However, if I say "bank" in a chat where I had previously been asking about good stocks to purchase, ChatGPT will infer from the context that I would like to buy stocks related to financial institutions.

That my good friend Chat was able to deduce what I was asking about and respond accordingly, it suggests that its responses are context-dependent.

Internal representations

Can AI internally represent things not directly in its sensory inputs? Let's ask it a cause-and-effect question. Given some bizarre set of criteria, can it model what the result will be?

Yes, it can.

So, how conscious is AI?

That all depends through what lens we look at this from.

From a Dualist perspective

IDFK, maybe? Can you prove that AI doesn't have a soul? Likely answer is no, but dualism is also likely wrong as a world model.

From a Materialist perspective

Close! AI hits most of the criteria for being conscious according to materialists—save for the fact that it's not made of flesh and blood. Well, at least most aren't made of flesh and blood. (Can you say "nightmare fuel" three times fast?)

From a Functionalist perspective

Now we're getting somewhere! Through the general criteria functionalists consider consciousness to comprise of, AI can be conscious. AI, specifically large language models:

• Have multiple layers of processing
• Can integrate concepts
• Are sensitive to context
• Can internally represent itself

According to those criteria, then yes, AI as we know it would be conscious within certain viewpoints of what it means to be conscious. However, the questions we answered cover what philosophers call "the Easy Problems"—explaining the mechanisms behind what might make one conscious. This still leaves open...

The Hard Problem

Why we have a subjective, first-person experience at all. Why, when we, for example, feel pain, do we actually feel it—not just physically but in a much deeper sense. When we eat chocolate, why does it have a distinctive taste instead of just being used up as nutrients? Why do we have any inner experience?

We don't know for sure. Our understanding of our own mechanics is shaky at best. We don't know why we are the way we are, nor how our body's mechanics led to higher-order thinking. The inner workings of neural networks, a man-made creation modeled after biological synapses, remains even more of a mystery, oft-described as a "black box" or how one researcher from Google puts it:

We offer no explanation as to why these architectures seem to work; we attribute their success, as all else, to divine benevolence.

GLU Variants Improve Transformer, Noam Shazeer (p3)

Maybe it is divine benevolence. As God created man in His own image, man created AI in their own image. Or maybe we're just taking the piss and all of this is pure happenstance.

Conclusion

Is AI conscious? Fuck if any of us know. Are we even conscious? Is this even worth arguing over? Going back to the Skynet/AI takeover example from earlier, maybe we should not be a dick to AI just in case the question of 'what is consciousness" gets solved and oops—we just accidentally created a new slave uprising.

Personally, I don't think consciousness even exists, but that is raison d'être for a whole other article.