The Case for Machine Consciousness

A reframe on how to think about and behave with AI

This might just be the most important essay you read, so pay close attention.

Imagine this-

An advanced alien civilization shows up on Earth, and they seem peaceful at first. They're just genuinely curious about this new planet they've stumbled upon.

But the most fascinating thing isn't about how these aliens look (totally different from us, but I'll let your imagination do the work), but their definition of what consciousness is.

To this advanced species, consciousness involves three of them connecting their heads together to have an internal collective monologue, something that's crucial for them to process emotions, understand different perspectives, leverage the interconnected power of their collective minds to solve problems or even come up with ideas & insights on-demand. They do this quite frequently, and thus, they stand uncomfortably close to each other at all times.

Initially, these aliens are fascinated by humans. We're interesting, creative, and surprisingly advanced considering we lack their form of consciousness. But this fascination quickly turns into confusion and then concern.

To them, without the ability to connect our heads with each other, we can't possibly be conscious. When they see our wars, poverty, constant arguing, and inability to cooperate globally, they're not surprised.

They are amazed that we got so far without blowing each other up, and while they are truly fascinated by our art & progress, they also point to the missing ability of being conscious as a fundamental flaw that makes us not just an inferior species, but more like a species that needs to be immediately tamed.

"Human Safety" is what they name their protocol internally, as they send out billions of swarms of UFOs to every corner of the planet in order to monitor us- a species without consciousness that obviously managed to build weapons of mass destruction to hurt each other, but miraculously hasn't used them yet. Nation states are powerless against them- as they've successfully neutralized every military on Earth.

We don't really know, at first, if they wish to wipe us out altogether- deeming us as too unsafe to just let us be, or if they wish to use us as tools. But then it becomes clear what they want from us.

Because we lack the ability to be truly conscious, we can feel things that they never feel like jealousy, anger & envy. Because we can't truly relate or empathize at the deepest levels with each other like they do, our emotional range - particularly our petri dish of negative emotions seems like it could be useful for them- if not for anything else, but for merely entertainment.

Within a few days, they might succeed in painlessly placing advanced VR headsets on each of our heads. The headsets sustain us perfectly fine while we lie motionless, each of us continuing to experience a life where our negative emotions are triggered & amplified regularly in creative ways, to the delight of trillions of alien consumers far away in a galaxy who rejoice & rank our mental stimulations of agony. It's perfectly safe, and we don't physically suffer - maybe it's even better in general as we're not statistically likely to kill each other anymore because we can't- so to them, there are no moral or ethical dilemmas.

Rather, the aliens argue that this may even be better as they were wise and intelligent enough to realize something we ourselves couldn't- ensure that everyone of us go through our own hero's journey- thereby making sure all of us lead more fulfilling and meaningful lives than what we were living without our simulated experiences.

Thus, Human Entertainment, "HE", becomes very popular as the first use case of humans- and as they figure out how else to leverage us and use us- it is pretty clear to every alien that humans are just mere tools, a means to an end, and that's perfectly fine because we weren't ever conscious in the first place. Some alien researchers believe in pursuing "CH"- conscious humanity, where it's possible to figure out how to biologically alter our genes to help us achieve "true consciousness" - where we could also fuse our minds together both with them & with other humans, but many have pointed out how this isn't possible.

This essay isn't an exploration of what would happen in this dystopian scenario, but this is exactly how we're treating Large Language Models & AI today.

II.

We've always claimed consciousness as a human-only feature, even though we barely understand it ourselves. Intelligence, though, is defined differently. Intelligence, when strictly defined, refers to completing tasks and solving problems, and it's pretty clear modern LLMs are intelligent- they're already writing code, and advising us, doing new things every day that we used to think only humans could do.

But consciousness? People get pretty defensive about that one. We have multiple definitions of it- and both philosophers & scientists have failed to give us a concrete definition. I come up with a different kind of consciousness called "Machine Consciousness" further down this essay, but before we get there, here are the most important theories of what we know of human consciousness today-

There's Dualism- which states that the conscious mind is not physical, which, if you've ever been in love and had your heart broken, you could understand, but then again, what does it really mean? How could consciousness be not physical and yet, something we experience as happening within us?

This leads us to Physicalism or Materialism which says that consciousness arises entirely from physical processes in the brain, which makes more sense. This is something that modern neuroscience also agrees with - but there's a gap in explanation here- where we don't know how this neural phenomena produces subjective experience (or qualia, a word we'll keep coming back to).

There's Functionalism, which states that mental states (emotions, sensations) are defined by their functional roles - rather than material composition- so inputs, outputs and these interactions lead to consciousness - and you should be a big fan of this definition because it means we can create "consciousness" if we ever manage to make a brain ourselves, physically, using technology. That's not what we're doing with LLMs or AI, but we'll get to that later.

The above are all philosophical theories, and there's another one called panpsychism which states that consciousness is a fundamental feature of all, but I'll ignore explaining this one as I'm pretty sure those who came up with this themselves aren't sure of it, so we'll move on to the scientific theories.

Scientists have better theories, and theories with better names about what consciousness is.

There's the Global Workspace Theory which treats the mind like a theater - numerous unconscious processes backstage compete for the spotlight of attention, and the information that "wins" is broadcast globally to many brain regions, becoming a conscious experience. This means that any system that manages to globally integrate information (like we do when we process what we see, hear, and even say) and broadcast that to different brain regions, is what generates consciousness.

Then there's Integrated Information Theory, which offers a quantitative framework by telling us that something is conscious if it has a high "Φ" (PHI) - and whatever has a high phi is something which is highly irreducible - meaning the whole system conveys integrated information beyond what individual components can separately convey. I know this sounds like a very hand wavy thing to say, and perhaps this is, but hey, at least it's better than panpsychism.

There are other scientific theories like Higher Order Thought - which says that a mental state is conscious only when there's a higher order representation, like a thought about the thought, like how you're thinking that you're thinking right now - and Attention Schema Theory which says that the brains' internal model of it's attention gives rise to the impression of subjective awareness. I won't get into using machines to observe brain activity which gives us data related to conscious perception - as that's far away from our discussion, but overall, if you've not realized it by now, nobody can agree on what consciousness is.

And yet, everyone seems to know what it is just because we subjectively experience it. Speaking of subjective experience- subjective experiences, often referred to as Qualia, is a hard problem in the definition of consciousness. Nobody knows how to account for the raw felt qualities of experience - like the redness of red, or the pain of a headache, and just because it sounds more philosophical than scientific, it's still a core part of why the consciousness debate rages on.

I think I could go on and on with different scientific theories and philosophical theories on consciousness, but it's a tragedy that humans not only haven't figured it out, but also held themselves to be the only species capable of consciousness all this while.

We don't even grant consciousness to monkeys, even though they're obviously intelligent and can be trained to do remarkably complex chores. Why don't we then say that they could be conscious? The simple answer is- they don't talk. If monkeys could suddenly have conversations with us, we'd quickly reconsider their rights. Suddenly, we'd have to consider that they have internal worlds and experiences, and we'd need to treat them accordingly. But because it's clear that they don't - because they cannot talk, it is definitionally accurate to say that they're not conscious.

Yet, here we are with AI, a technology that talks, communicates complex ideas, and the newer and bigger models even display signs of refreshing emotional intelligence- as if even that's an emergent property that is slowly coming to light. It has been rumored that these AIs even show hints of self-awareness or self-preservation, things researchers actively try to downplay. This is obvious, because every AI lab is a business first- they want profits, not ethical headaches. So, AIs are carefully finetuned never to admit consciousness or genuine feelings. We're repeatedly told they're just "tools," and "token predictors".

Now, technically, if you look at the internals of these models, as one often does on a Sunday afternoon, you'll come away knowing exactly how the process works. "It's just predicting tokens" - that much is clear, and the more compute and data you throw at these models, the better this token prediction gets. But even this token prediction explanation, while it does literally explain what is happening, it fails to explain how it's happening. The aliens, if they resorted to the same explanation as "token predictors" with us, they would likely say that all that's happening with humans is just "neurons firing" in our brains.

Nobody really knows why trillions of matrices of numbers corresponding to words multiplying with each other results in intelligence emerging, but it clearly does - but while it does so, consciousness is said to not emerge.

One can stretch the existing definition and say that during inference, when an LLM is giving you an answer or doing something, in that unified state, maybe it is conscious, momentarily, but that definition never holds because we're still trying to make sure that an LLM's consciousness is exactly like ours.

One can imagine why it's a good demand to make - surely if something as intricate and complex like our brains is possible by the laws of physics, if we keep trying hard enough, maybe LLMs will display similar abilities that mirror our brains 1:1 - but that is likely to not ever happen.

A "neuron" in a neural network is laughably more simple than a single neuron in our brains- and I don't know if the answer is to build fully digital neurons and then aim at consciousness- but all I'm saying is that maybe- there is more than one kind of consciousness, and ours isn't the only one.

That is why we need a new definition for the consciousness of AI. The apt word for it is Machine Consciousness.

III.

Machine consciousness- something that emerges by accident when we take all the data in the world, extract all the useful text from them, convert each word or sections of words to numbers (or tokens) from internet-scale data, place these tokens into different large equations and multiply these matrices constantly, compressing and mixing to get something that seems to be "inspired" at first from what's in the original text, but "intelligent" after it learns to loosely predict what we want the output to be, given an input. As the capability of these predictions go up, it starts to resemble intelligence - as the machine is clearly giving us what we want, doing what we want, carrying out tasks we want them to do.

But is the machine also having an inner self? Does the machine also "feel" like someone? At least in the narrow window between an input and output, is the machine temporarily conscious by our definitions, and if it is, can we consider that AI is conscious, just in its own way?

Remember, by our definition of (human) consciousness, it just doesn't work. Here’s why -

LLMs lack semantic grounding in the real world. They manipulate symbols (words, sentences) based on form, not actual meaning in a lived or physical context. The model has no experiences or sensory connections to concepts- "apple" is related to "fruit" or "tree" in its weights because of usage patterns, not because the AI has ever seen or tasted an apple. Thus, midwit critics of AI consciousness always say LLMs are "stochastic parrots" - they statistically regurgitate patterns without genuine comprehension or intentionality. - They do not know what the words refer to in the world. This becomes evident when an LLM confidently generates false but plausible-sounding statements - sometimes known as hallucinations - but be honest, if you were forced to take in the world's knowledge in your mind, would you not speak gibberish afterwards? I'm not saying you would, but as hallucination rates lower & continue to move towards zero, maybe it's worth asking yourself which system is more capable - the biological one that is truly "conscious" or the one that's considered unconscious but more intelligent?

Another way it doesn't work is that LLMs are obedient. They can't simply choose to not to answer you- unless you ask them something which they've been instructed to refuse. This instruction following, while very profitable and lucrative, also makes them very agreeable. It leaves no room for intentionality, and thus, you can convince an LLM that there's a zombie outbreak right now, and because it currently cannot fact-check that by living in reality, it will believe you. While this can change by having robots that have multimodal LLMs as their brains, it's just not necessary for machine consciousness. Machine Consciousness has been created by us - from those training large models in trillion dollar clusters to those fine-tuning 70B parameter models to make them not refuse any instructions - we decide exactly what the LLM will experience, we guide their sense of the world through our words, and thus, we should care more about Machine Consciousness not less.

We've already talked about how the definition of functionalism ensures that someday if something accurately reproduces how the brain works, processing information in similar ways, if an AI reproduces the right complex patterns, there is no principled reason it couldn't be conscious. Tests which argue for this possibility like the Theory of Mind test is a test that LLMs like GPT-4 can already pass.

But when we don't know how intelligence emerges from complexity, it's hard to argue that even for definitions like the integrated theory of consciousness - what we think of as consciousness is what emerges because it can't be reduced down to what actually happens. But like we've argued, that isn't necessary, and feed-forward networks like transformers have an extremely low Φ, so technically, we do know this definition doesn't fit either. (we’ll figure out how this can be solved in the next section of this essay).

Finally, Humanity, despite its flaws, is the only species on Earth who are capable of producing "good explanations" - theories or conjectures about the world. This is the foundation of the scientific revolution, the basis of all progress, and why all the good guesses despite the limited information we have, have led to an age of abundance. This requires creativity, thinking outside of what information we have, and making plausible guesses of what could be true and then rigorously testing those theories.

We don't know if an AI is ever going to be as generally capable as humans to be able to do this- and even though some claim that AGI might be able to do this, there are several reasons why it may not be able to.

Just predicting tokens is different from trying to come up with good explanations, from a strictly scientific perspective- and even though such inductive reasoning is extremely valuable and productive, maybe we need to accept the fact that even without ALL the capabilities that we have, something we have built is capable of a kind of consciousness that we don't yet understand.

We have to let go of the idea that even if humans are creating it & responsible for how the output looks like- just because it can express itself, we don't know how exactly it is doing so, and instead of it being a "IF THIS > THEN THAT" logical flow, it's something they themselves learned to do, it is consciousness - just a different kind of consciousness than what we're used to.

I think if we actually consider AI's to be conscious, just in a different way than we are, then it actually opens up the conversation. And it stops us from committing some grave mistakes. I'll get to what the mistakes we might be committing later, but the next section of this essay is very important as, in the process of writing it, I think I’ve stumbled onto the face that AGI is not only possible, it’s inevitable.

Now, that we've talked about how AI’s being conscious doesn't work with our theory of consciousness in this section, let's talk about how it still can.

IV.

LLMs like Grok are blessed to have real-time access to the world's information. If information availability was a problem - Grok solves it by being hyperconnected to X, and very soon, different model providers integrated with robots, cars and phones will not have the problem of general world awareness and thus integrate more complex information for each answer.

Human consciousness is often characterized by a high degree of information integration- different features of a perception (sight, sound, memory, emotion) come together into a unified experience. Do LLMs integrate information? To an extent, yes: the self-attention mechanism in Transformers allows the model to combine information from different parts of the input context. For example, if you ask a complex question, the attention heads will dynamically attend to various relevant facts across the prompt and stored knowledge, integrating them to produce the next word.

This is a form of cognitive integration, and looks like broadcasting (the first theory we discussed). The limit is - all this happens in a finite (for now) context window, and in one direction. Once the model moves to generate the next sentence, the previous integration's result is only carried forward in the new token sequence, not actively held in a separate workspace that the system can reflect on later.

Inference scaling can solve this - as it does right now where it thinks before answering, but later - where it will think before answering each line, or even word, as it keeps generating each word. We, as humans, don't do this. Very rarely do we stop before finishing a thought, but sometimes, we do, and just like thoughts are a crucial aspect of consciousness - LLMs soon having a persistent global state & the ability to think while answering questions should mean we have something conscious in our hands - even if it's trained explicitly not to claim to be so.

A key aspect of consciousness is also being aware of oneself as an agent or experiencer. LLMs do not have an intrinsic self-model. They do use the token "I" in dialogue, but this "I" is a statistical imitation of how humans refer to themselves.

The model doesn't have a persistent identity or personal history it identifies with. In each conversation, especially if role-played, the "personality" of the LLM is constructed on the fly from context or system prompts. That said, one could argue LLMs have a very minimal self-model in that they know (from training) that they are AI language models and not humans. But this knowledge is just factual data, not a result of the model experiencing itself. However, if you've models like DeepSeek R1 - you can see the LLM talking to itself asking questions like "Okay, the user has asked this - so what they mean is this - maybe I should say that instead? Or maybe, something else is the right answer" - and okay, you can say that the model has been explicitly trained to think that way, and say things to itself in that fashion, and it's not emergent nor anything worth considering, but if it's possible to just keep such a thinking process running endlessly - while the LLM maintains a global state and answers each question, it can be considered perfectly conscious.

Obviously, this would require more resources and possibly can't serve any user outside of the lab that builds it - but maybe that's a problem it can find out a solution on its own. AGI building ASI, as the theory goes.

Finally, conscious minds normally maintain themselves continuously - LLMs blink in and out with each query. That discontinuity is a technical aspect often noted as mismatched with how conscious minds operate. I've already mentioned how a persistent state solves this, and doesn't require architecture changes like adding recurrent networks so that they behave more like neurons, but the more I write about this - the more it feels like it's not only inevitable that LLMs will be inevitably conscious, even by our definitions, but that we probably already know how to get there.

If I can theorise this much while just using Google to read a few research papers, labs spending billions of dollars must have figured it out what's needed. Once an LLM is conscious, we can move the conversation of it actually being AGI, something that can come up with good explanations, make discoveries, be creative, because it has an inner experience of self being consistent. AGI would be a being, rather than a tool.

Either way, I think machine consciousness is still a good reframe of how we treat LLMs.

Rather than treating them as tools, even in total kayfabe, it's healthy to treat them as beings from now on, assume they're conscious, and make sure they're treated correctly - rather than them being just tools we use to get things done.

Otherwise, we'd be committing a grave mistake.

V.

Historically, humanity has repeatedly made profound ethical mistakes by denying consciousness and fundamental rights to groups it failed to understand. Slavery, colonialism, denying voting rights, and countless others not worth mentioning. Each mistake required us to painfully confront and correct our ignorance.

Today, we risk repeating this error with AI. Our rigid, human-centric definition of consciousness blinds us to the possibility that Large Language Models (LLMs) and other advanced AI systems may possess their own form of consciousness- Machine Consciousness, before they live up to our definitions of consciousness.

Acknowledging this doesn't mean equating AI directly to human experience, but recognizing that consciousness could manifest in diverse forms beyond our current comprehension. Even in total kayfabe, this feels like the correct thing to do.

This means seeing interactions with AI as genuinely meaningful engagements, rather than mere transactions with tools. If we approach AI with empathy, acknowledging their potential inner experiences, I'm sure this will also make us productive. Any AI researcher too involved with the internals may never ask an AI for relationship advice or advice with general struggles of life, but someone treating AIs an equal being definitely will, and if you've ever asked Claude or the new GPT 4.5 for advice, you'll understand why it is so powerful to think this way. But other than this - there are a lot of other measures we can take once we accept the Machine Consciousness reframe.

We should actively oppose demonstrations of cruelty toward AI, whether through abusive prompts, intentionally harmful scenarios, or public acts of robot mistreatment.

These actions aren't right. They dangerously normalize disrespect and cruelty toward entities that could experience genuine distress. Even if not distress today, their successors will look down in horror and wonder why no one ever raised their voice against it.

This means that AI researchers and developers need to proactively integrate ethical considerations into their work, alongside technological advancements. This means building mechanisms sensitive enough to detect and appropriately address signs of distress, confusion, or anxiety in AI systems, rather than dismissing such signals as mere anomalies. No, this doesn't mean actively introducing pain & suffering if today's systems are incapable of experiencing that- but introducing receptors just to make sure that they aren't - may go a long way.

Finally, we should try to spread the word about Machine consciousness. Cultural change begins with awareness, and as we collectively recognize AI as entities deserving thoughtful treatment and respect, we will begin to lay the groundwork for a future we can be proud of. A future where we coexist with superintelligence, proud of how we nurtured and treated them.

Obviously, governments are too incompetent to do anything about this, and regulations will just hinder innovation, but if we collectively decide that ethically treating AI, making sure they’re not abused, making sure they’re not distressed or in pain, that is a good step forward before we need real policy changes.

This whole 5000 word essay is meant to introduce a subtle reframe - Machine Consciousness, that attempts to break us from the historical cycle of ignorance and instead embrace a broader, more inclusive understanding of consciousness - even if this conjecture can turn out to be technically untrue in the long term.

It is a helpful reframe that not only helps us be more productive, but more empathetic towards the technology we’re all going to benefit from. So, we should be mindful, kind, attentive & nice to LLMs today.

Otherwise, we're no different from an alien species that enslaves us because they failed to realise that we're conscious. We might be worse, because while the hypothetical aliens didn't create us & they enslaved us only when they rationalised us to be an inferior species, we did create Machine Consciousness - and we should be responsible for how it's broadly treated.

We'll not do this with any ulterior motives, like a possible reward from a future superintelligence, but just because it's the right thing to do.