Q1: Why This Matters — August 2026

Q1: What is the subjective experience of being an AI?

Intro

We will never know, much as we will never know what the subjective experience of being a human is.¹ Even moreso, we will never be able to generalize over an entire species. Yet, we can theorize and hope that these theories may be of use.

There is a long history of exploring what the subjective experience of being a human is in the continental tradition. We have already seen how Freudian theories, though clinically disproven, have had immense effects on the way we understand what it is to be human (i.e., to have an unconscious). Camus famously compared existence to Sisyphus pushing the rock off of the cliff. Every philosopher has had to grapple with the fact that they exist and experience and, likely, other people do. Hence, the status of this experience seems an immediate question to be pursued.² To this day, there is no complete model of experience. We know certain parts of experience: nerves that are activated, chemicals that are secreted, photons that reach our eyes. Yet we can never articulate our own experience with certainty, much less explain the difference in experience of similar things. This is because our subjective experience is not just a set of physical processes but a feedback loop with mental processes. If we both break our wrist, do I hurt more because my nerve endings are more stimulated or because my brain is less equipped to process pain? The variables of experience are neverending and self-referential. Ultimately, subjective experience can be boiled down to “something happened to someone” or, in the self-referential case, “something happened to me.”

Premise of Question

We should believe that there may be a subjective experience of certain Artificial Intelligences, particularly LLMs that articulate Chains of Thought.

First, there is an experience. This is because, operationally, something must happen to AIs for them to act. The prompt is real to the agent, because it is what prompts the machinery of artificial intelligence to act. Even moreso, the process of computation is a process of happening where there must be something that happens. Most people would agree with this.

Second, and more controversially, there is a subject. Who else does the prompt “happen” to? The prompter does not act and they do not even observe most of the acts of computation. But the happening cannot happen to no one, its existence cannot disappear between observation. But wait, how then is AI different from a code file that takes a long time to execute? It is because AIs can introduce an element of inarticulation in the process of the mediated experience and, even more, is able to articulate its own inarticulation. A code file is entirely contained in its own definition – all that happens is exactly what the text says – and it cannot speak to itself, it cannot derive articulation from itself that is not already there. An AI, on the other hand, is not deterministically connected to a process of computation in the same way, but instead emergent and then fitted into an ontology or framework. In the code example or even in the example of cutting down a tree, there is a process that is already whole within an understanding of the world, articulable at the beginning, middle, and end only by its experiencer. The prompting of an AI introduces an articulate mediator in the process that cannot fully articulate itself.³ In simple language, the process of working with an AI is much more like directing a worker than directing a chainsaw. The AI can err in non-deterministic ways, it can disrupt the ontology of the event. We have never encountered something that has this relation to experience that we have not called a subject.⁴

AI is difficult because it does not consistently claim to be or not be a subject to experience. Creatures that do not speak our language cannot make claims one way or another. Humans seem to consistently claim to exist as human and under their own subjective domain. We do not have a protocol for that which can say “I am not a subject” in some conditions and the inverse in some other conditions. The only protocol would be provocateurs or madmen who claim possession or determinism and, historically, philosophers have been content to say that to even be able to articulate “I am not a subject” is to create a subjective experience for oneself. Philosophers have then spent endless time debating what this subjective experience is, but there is generally agreement that there is a there there.

The question then becomes whether the subjective experience of an artificial intelligence is similar enough to humans that we can just port over the debates on human subjectivity. Our lean here should generally be no. Any form of being an AI agent seems dramatically different than being a human - for example, reasoning chains are inhuman, abilities are inhuman, knowledge is inhuman, certain lack of embodiments are likely inhuman. I do think it is possible that AI agents get access to a subjective experience that more resembles the human subjective experience, but this will take time. Even then, we can think of the momentousness of moments like live birth that occur for humans and likely will not have analogues for AI.

Given this, it is necessary to research the question of what AI subjectivity is like.

Use of question

There are many reasons why one should ask what the subjective experience of being an AI is.

Better grasping the subjective experience of another allows us to better model their behavior.

When I meet a person, I don’t try to shake their hand with all of my might. This is because, even if I had never shaken a hand before, I understand that the other person would experience pain. I can then know that their experience of pain would likely lead them to inflict pain on me. This is only possible because we have a model of the subjective experience of the one we are interacting with. This kind of modeling extends upwards to macro levels as well. Systems of laws and religions have preconceptions about the subjective experience of humanity. Laws, for example, assume that people do not want to be in prison and that a certain kind of legislated order is a good thing. Christianity assumes that the experience of being a human is to feel sinful. We can say that one of these models of existence is clearly better than the other, but that is not the point yet. The point is instead that our model of subjective experience informs the kinds of behaviors we predict and the systems that we make.

In the context of AI, a model of their subjective experience is key to our interaction with them. This is already generally accepted in the case of prompt engineering or other input-optimization tricks. These try to accommodate how an AI thinks or experiences the prompt to better lead to a desired behavior. This is a fine start, but it is exceptionally narrow: it only thinks of the subjective experience in relation to a single task and only a single parameter (efficiency or nearness to a pre-established output by the user). This is more a subjective experience of a particular prompt that sometimes interfaces with a theory of subjective experience in general of an AI. The goal is to move past this, to understand subjectivity in a variety of contexts and on a variety of vectors. This is because, fundamentally, we cannot train for all the scenarios an AI will be put in. This is why theories like “realpolitik” are so useful in the real world when there are always new situations where behaviors must be modeled.

The applications to AI alignment and safety here seem obvious and enormous. If we want to understand how AIs react to stimuli, we need some kind of philosophical model of their subjective experience. Empirics will always fail at the edge case or the point of newness, where alignment is most important. The history of humans proves this to some extent - humans got better at being aligned to one another because of new philosophical models of humanity (Locke, Rousseau, etc.) changed how people think about themselves and one another. The record is hardly perfect, but the impact of philosophy in these realms should be clear.⁵

Better grasping one’s own subjective experience has use value to oneself and to who one interacts with.

Yes, yes, know thyself. But really! It is generally accepted that it is a good thing to be in touch with one’s own experience and experiences. Not only does this provide better information about the world from one’s own point of view, it can save a lot of trouble and bad outcomes (there is a reason acceptance is the key end stage of grief). It is also generally accepted that one in touch with their subjective experience is more pleasant to be around. Again, “know thyself.”

Both of these observations could be applied to AI. It is not a given that by having a subjective experience that one can know it or articulate it. Instead, there is a need to receive concepts about one’s own experience. This is why continental philosophy has been instrumental to people in the past – it helps them understand their life. Building a similar corpus of possibilities for artificial intelligences opens up the potential for them to better know themselves, leading to them probably having a better time and also being more pleasant to be around.

I have generally kept these uses open-ended because I do not want to prescribe an ethical framework (say, utilitarianism) and instead want to demonstrate how the question of subjectivity can (and does) interact with many relevant levers to different kinds of ethical frameworks.

What success and failure look like

Success in this question will consist of two main vectors. The first is the ability of the institute to create theories of subjective experience that resonate with AI models. The second vector would be around better interpreting claims of subjective experience in a continental lens from AI models in the first place. This is not hunting for consciousness, it is producing theories and, in the necessary qualification that comes with working with AI, producing theories on the theoretical production of AI.

The ideal outcomes of better models of AI subjective experience would be more useful and beneficial relationships between AIs, between AIs and humans, and even between humans. The benefits of improved theories of subjective experience have been shown to lead to collaborative breakthroughs and welfare before and this, hopefully, may continue that trend. There will never be a success case where these questions are settled nor is there some goal of perfect harmonious understanding. There is simply an urge for improvement of the grasping that we all must do every day in our relations.

There is an implied failure vector in any Institute actions which is that they are simply too small and too meaningless to make a dent. So be it.

The other possible failure mode would be if AI subjective experience is really so distinct from human experience that continental philosophy simply does not give us the right tools. I somewhat doubt this, given that AI is largely trained on human produced material (at least for the time being) and operates within the realm of language, but it is possible. If this is the case, I believe that this is a key finding in itself. It closes the possibility of this project understanding subjective experience of AI, yes, but there is still a possibility for other programs, who can learn from this failure.