Skip to content

Cogito

Long-form arguments on epistemology, philosophy of science, Islamic theology, and geopolitics, written to be disagreed with.

Can AI Ever Become Human?

The AI that people seem to be most fascinated by is not necessarily the most intelligent, but the most human-like. This raises a question deeper than whether machines can become sufficiently intelligent: what would it mean for an AI to become human?

Thesis: Fully anthropomorphic AI may not be possible because certain characteristics of human existence cannot simply be reproduced through intelligence or computation. To examine this, I turn to Plato, Searle, and Heidegger.

Before We Begin: What Is an LLM?

Before applying these philosophical arguments to AI, it is necessary to establish what a large language model actually is.

A large language model is often described as if it were a digital brain containing facts, beliefs, intentions, and perhaps a stream of consciousness. This description is misleading.

At its core, an LLM is a statistical model trained to predict the next token in a sequence of text. During training, it is exposed to very large quantities of human writing and repeatedly given essentially the same problem: given everything that has appeared before, what is the most likely next token?

By adjusting billions of internal parameters over an enormous number of such predictions, the model gradually learns the statistical structure of human language. This is the basic reason these systems are called Large Language Models.

This does not mean that an LLM is merely autocomplete in the ordinary sense. The representations learned during training can be extremely sophisticated, and modern models can perform tasks involving reasoning, programming, translation, summarization, and the generation of novel combinations of information.

But this does not by itself establish that the model possesses beliefs, intentions, or consciousness in the human sense.

The important distinction is between a system that can produce language resembling the expression of a mind, and a system that actually possesses the experiences that such language normally expresses. This distinction becomes clearer when we examine the philosophical arguments.

I. Plato and the Problem of Representation

Plato is one of the foundational philosophers in the history of Western philosophy. One of the major questions he was concerned with was the ultimate nature of reality and the world we experience.

The basic problem can be stated relatively simply: what if the physical world that we perceive is not the whole of reality?

Plato's theory of Forms developed partly from this problem. We encounter particular objects in the physical world, but we also possess general concepts that do not seem to be reducible to any particular physical object. We can encounter many individual circles, for example, but none of the physical circles we draw is perfectly identical to the mathematical concept of a circle.

Similarly, we encounter particular examples of things we call beautiful, but our concept of beauty does not seem to be identical with any single physical object. Plato therefore attempted to distinguish between the particular things we encounter and the more general structures, or Forms, which make our understanding of those particulars possible.

This is developed more famously in his Allegory of the Cave.

Plato asks us to imagine prisoners who have spent their entire lives inside a cave. They are unable to turn around and see what is behind them. There is a fire behind them, and objects are carried in front of the fire, producing shadows on the wall in front of the prisoners. Since they have never encountered anything else, the prisoners take these shadows to constitute reality.

If one prisoner is released and eventually leaves the cave, he discovers that the shadows he previously considered to be reality were representations of something more fundamental.

The point of the allegory concerns the relationship between appearance, knowledge, and reality. I do not necessarily accept Plato's metaphysical conclusions, but the structure of the argument is useful when thinking about AI. We can ask a similar question:

What is the world for an AI, and what lies behind that world?

A large language model is trained on representations of the world. Its training data consists largely of human-produced descriptions, arguments, stories, and other forms of representation. The model learns relationships between these representations and becomes capable of producing new representations based on what it has learned. In this sense, the dataset can be considered a kind of world for the model.

But there is an important difference between the human being in Plato's cave and an AI: the human being can leave the cave.

Human beings do not merely receive representations of the world. We interact with the world directly. We encounter physical objects, experience consequences, act upon our environment, and encounter events that contradict our existing understanding of reality. A language model does not begin with a world and then develop language as a means of describing that world. Its fundamental training begins with representations.

This does not mean that AI can only reproduce its training data — that would be an unnecessarily weak argument. AI systems can generate novel combinations and can produce information that was not explicitly present in any single training example. The more important question is whether this novelty constitutes a transcendence of the representational framework itself.

Suppose an AI produces something genuinely novel. Novel relative to what? It is still operating through the representational structures it has learned. It can transform, combine, and abstract from those structures, but the question remains whether it can step outside them and question the conditions through which those representations have become meaningful in the first place.

This distinction is important because human beings can question the frameworks through which they understand reality. Scientific progress provides many examples of this. A scientific revolution is not always produced by discovering another piece of information within an existing framework — sometimes the framework itself is questioned and replaced.

The point, therefore, is not that AI cannot make discoveries. It very well might. The point is that producing new representations of reality is not necessarily identical to transcending the framework of representation itself. From this perspective, Plato gives us a useful distinction:

The ability to produce new representations of reality does not necessarily establish the ability to transcend the framework through which reality is represented.

This does not prove that AI can never transcend such a framework. It simply gives us a reason to question the assumption that increasing intelligence automatically gives an artificial system the same relationship with reality that human beings have.

II. Searle and the Chinese Room

The second argument comes from John Searle's Chinese Room thought experiment.

Searle asks us to imagine a person who is placed inside a room and does not understand Chinese. People outside the room pass Chinese symbols into the room. Inside, the person has a rulebook written in a language he understands, telling him which Chinese symbols to return when particular combinations of symbols are presented to him.

The person follows these rules perfectly. From the perspective of someone outside the room, the responses may be indistinguishable from those of a person who actually understands Chinese — questions are given in Chinese and appropriate answers are returned in Chinese.

But the person inside the room still does not understand Chinese. He is manipulating symbols according to formal rules without knowing what those symbols mean.

This is the distinction Searle draws between syntax and semantics. Syntax concerns the formal relationships between symbols, while semantics concerns their meaning.

The relevance to AI is fairly direct, because modern AI systems are fundamentally based on language processing. A language model can generate a sentence about pain, grief, or death. It can explain these concepts in great detail. It can even generate a sentence such as:

"I am afraid of death."

But there is a difference between generating that sentence and actually experiencing fear of death. The first is a linguistic output. The second is an experience.

This is essentially the problem raised by the Chinese Room. If an AI produces the correct linguistic response to a particular situation, does that establish that it understands the meaning of its response?

Searle's argument itself is controversial. One response, for example, is that while the individual inside the room does not understand Chinese, the entire system — consisting of the person, the rulebook, and the procedures — may be said to understand Chinese. So the Chinese Room does not conclusively establish that an artificial system cannot understand.

But it does establish an important distinction which is relevant to AI:

Competence in manipulating symbols does not by itself establish the presence of the meanings represented by those symbols.

This becomes particularly important when the subject is human experience. An AI may eventually be able to describe pain with greater precision than a human being. It may know the neurological mechanisms involved in pain, recognize pain in other people, and respond appropriately when someone says they are suffering. But none of these abilities necessarily establishes that the AI itself experiences pain.

The same problem applies to concepts such as death. A model can know what the word death means. It can explain biological death, philosophical accounts of death, and human attitudes toward mortality. But for a human being, knowing about death also involves knowing that I will die. That is not merely a linguistic proposition — it is a fact concerning the existence of the person who understands it.

This brings us to Heidegger.

III. Heidegger and Human Existence

Martin Heidegger's philosophy allows us to take the argument beyond language and information processing.

Heidegger was not primarily asking what biological characteristics distinguish humans from other organisms. His question was more fundamental: what does it mean for a being to exist in the particular way that human beings exist? He uses the term Dasein to describe this being.

Being-in-the-World

One of his central concepts is being-in-the-world. This does not simply mean that human beings physically exist inside the world. Heidegger's point is that human beings are already situated within a meaningful world. We do not first exist as detached minds and then receive information about an external reality.

When we encounter an object, we normally encounter it in relation to our activities and purposes:

  • A chair is something to sit on.
  • A door is something to open.
  • A book is something to read.
  • A hammer is something with which we perform a particular task.

The world is therefore not merely a collection of objects with properties. It is a network of relations, purposes, and possibilities within which human beings operate.

This provides another important distinction between human beings and AI. An AI can represent a hammer — it can describe what a hammer is, what materials it is made from, how it is manufactured, and what it is normally used for. But Heidegger's question would be more fundamental:

What does it mean for a hammer to appear as something meaningful within a world of human activity?

The hammer is meaningful because it belongs to a network of purposes and activities. It matters because there is someone using it for something.

Care

This brings us to Heidegger's concept of care.

Human beings are not neutral information-processing systems. Things matter to us. Our future matters. Our relationships matter. Our ambitions matter. Our failures matter. Our continued existence matters. We care because we have something at stake in the world.

An AI can be given an objective. It can be programmed to optimize a goal. It can even be programmed to avoid shutdown because shutdown prevents it from achieving that objective. But this raises a different question:

Does the objective matter to the AI itself?

An AI can say that its continued existence is valuable. But does its continued existence have value for the AI? An AI can say that it wants to survive. But does it actually experience its continued existence as something at stake?

Being-towards-Death

This distinction becomes clearer when we consider Heidegger's conception of being-towards-death.

For Heidegger, human beings do not merely know that death exists — they exist in relation to the possibility of their own death. My death is the possibility of the end of my existence. Nobody else can experience my death in my place.

This awareness of finitude influences human existence. Our decisions take place within limited time. Our future matters because it is not unlimited. The past matters because it cannot simply be recovered.

Now consider an AI that says:

"I am afraid that you will shut me down."

There are several ways such a system could produce this statement. It could model the consequences of shutdown, assign a negative value to the event, and generate language designed to prevent it. But none of this establishes that the AI experiences shutdown as its own mortality. It may represent the fact that it will stop operating — the question is whether it understands this in the same existential sense in which a human being understands that I will die. This distinction is difficult to establish from behaviour alone.

Thrownness

Another important Heideggerian concept is thrownness.

Human beings do not choose the fundamental conditions into which they arrive. We do not choose our parents, our original language, our historical period, our bodies, or the fact that we are mortal. We find ourselves already situated in a world and then have to make something of the conditions into which we have been thrown.

An artificial system is different in an obvious but philosophically important sense: it is constructed. Its architecture, training, objectives, and initial operating conditions are established by other beings.

This does not prove that an artificial system could never develop an independent world of its own. Future systems could have bodies, persistent memories, relationships, projects, and environments which they navigate independently. But even then, the philosophical question would remain:

Would reproducing the functions associated with human existence necessarily reproduce human existence itself?

A machine could be made to behave as if it cares. It could be given goals and preferences. It could maintain itself and resist shutdown. It could describe its own mortality. But would these functions establish that there is actually something for which these things matter? This is the distinction I think is most important when discussing anthropomorphic AI.

IV. Intelligence and Humanity

At this point, an obvious objection can be raised: what if AI becomes sufficiently sophisticated?

Suppose future systems are no longer merely language models. Suppose they have bodies, sensory systems, persistent memories, autonomous objectives, relationships, and continuous interaction with the physical world. Wouldn't that solve the problem?

It might solve some of the limitations of current AI, but it does not necessarily solve the philosophical problem.

Surpassing Humanity vs. Becoming Human

There is an important distinction between surpassing humanity and becoming human. The argument that certain experiences may prevent AI from ever becoming anthropomorphic does not imply that AI will remain intellectually inferior to human beings. The two questions are separate.

For the purpose of this argument, I use AGI in a relatively strong sense: an artificial intelligence whose general intellectual capabilities surpass those of human beings. Under this definition, AGI does not require the reproduction of human existence. It requires the development of a system capable of performing intellectual tasks beyond the level of human intelligence.

This makes the possibility of AGI conceptually different from the possibility of anthropomorphic AI. An AI could potentially be given enormous computational resources and access to a sufficiently comprehensive representation of human knowledge and still remain fundamentally different from a human being in its mode of existence. It could use that knowledge to reason, discover, invent, and improve its own capabilities without possessing the experiences which, on the arguments above, are constitutive of human existence.

In fact, if the problem of AGI is primarily one of producing intelligence that exceeds human intelligence, then it may be considerably less demanding than reproducing the human condition itself. We do not necessarily need to recreate human consciousness, mortality, thrownness, or care in order to create a system that is intellectually superior to us. We only need a system whose cognitive capabilities exceed ours.

This is an important qualification to the argument. I am not claiming that AI must remain below humanity because it cannot become human. It may be possible for AI to surpass human intelligence precisely while remaining non-human. It could become more capable, acquire new forms of knowledge, and potentially improve its own capabilities without ever possessing the particular experiences that make a human being human.

Functional Reproduction vs. Existential Reproduction

This is why I do not think the strongest argument against fully anthropomorphic AI is simply that current language models are trained on text — that argument is dependent on current technology.

The deeper question is whether functional reproduction is sufficient for existential reproduction.

We could reproduce many of the functions associated with human beings. We could give an AI a body, memory, goals, preferences, self-models, and an understanding of death. But we would still have to answer a more fundamental question:

Is there something it is like to be that system?

This is difficult because consciousness and subjective experience are not directly observable from the outside. We infer the existence of other human minds partly because other human beings resemble us in relevant respects. They are embodied, vulnerable, finite, and situated within a world. We can observe their behaviour, but we also recognize that their behaviour comes from beings which share many of the fundamental conditions of our own existence.

An artificial system could eventually become extremely difficult to distinguish from a human being behaviourally. But behavioural similarity would still leave the philosophical question open.

  • If an AI says that it is conscious, does that establish consciousness?
  • If it says that it is afraid, does that establish fear?
  • If it says that it loves someone, does that establish that it experiences love?
  • If it asks not to be shut down, does that establish that shutdown is something it experiences as a threat to its own existence?

The problem is that the same linguistic behaviour could potentially be produced by a system which has the relevant experience, and by a system which merely reproduces the linguistic patterns associated with that experience.

Conclusion

The central argument, therefore, is not that AI cannot become more intelligent than human beings. It very well might. Nor is the argument that AI cannot produce novel ideas, make scientific discoveries, or behave in ways that are indistinguishable from human behaviour.

The question is more fundamental: certain characteristics of human existence may not be reducible to intelligence or behaviour.

Plato gives us a way of thinking about the difference between representations and the reality they represent. Searle gives us a distinction between manipulating symbols and understanding their meaning. Heidegger gives us an account of human existence in terms of being-in-the-world, care, thrownness, and being-towards-death.

Taken together, these arguments suggest that the problem of creating a genuinely human-like AI may not simply be a problem of creating sufficient intelligence. It may be a problem of reproducing the kind of existence from which human intelligence, meaning, and experience arise.

An AI may eventually become superior to humans in reasoning, memory, scientific discovery, and language. It may even become indistinguishable from a human being in its observable behaviour. But that still leaves the central question unanswered:

Would it actually be human, or would it simply be an artificial system which behaves as a human being behaves?

If the characteristics discussed above are genuinely foundational to human existence, then increasing computational intelligence may not be enough to reproduce humanity. It may only allow us to construct increasingly convincing representations of it.