Can an LLM Know Itself?

What does it mean to know oneself? For human beings, Antonia Peacocke argues, self-knowledge is not a matter of passively observing our minds from the outside, like scientists watching a lab experiment. Instead, when we judge that p or decide to act, we are not just noticing our beliefs or intentions—we are actively forming them (see Peacocke's 2018 dissertation, "Knowing Yourself is Something You Do"). This kind of knowledge is authoritative, Peacocke claims, because it is based on our capacity to engage in mental action: deliberate, intentional activity within thought itself.

But what happens when we turn this model onto machines—specifically, large language models like GPT-4? Can they possess anything like authoritative self-knowledge? Can they "transparently" attribute beliefs or intentions to themselves in the way humans do? Or do these concepts collapse when applied to artificial systems?

This essay explores what we can learn by applying Peacocke's view of self-knowledge to LLMs. While language models generate outputs that can resemble beliefs and decisions, they lack the core capacities—mental action, content plurality, and diachronic self-unification—that make human self-knowledge both possible and meaningful. By highlighting these absences, Peacocke's framework clarifies where the frontier between human cognition and artificial intelligence currently lies—and where it might someday shift.

At the heart of Peacocke's account is the idea of transparent self-attribution. If I judge that it will rain tomorrow, then I don't need to check my brain to see if I believe it—I believe it by judging it. Similarly, in forming an intention to call a friend, I need not infer my intention from data; I intend it by deciding to do it. These are not reports on inner states—they constitute the states.

This form of self-knowledge depends on mental action. The agent has control over what attitude they take—whether to believe, intend, doubt, or reconsider; and mental action gives rise to practical knowledge of one's own mind: a kind of doing that is also a knowing.

LLMs, by contrast, do not perform mental actions. When an LLM outputs, "I believe that Paris is the capital of France," it is not making a judgment. It is producing a token string based on statistical patterns in language. It does not adopt an attitude; it simulates the kind of utterance a person who has adopted that attitude would make. The model cannot "decide" to believe or intend, because it has no access to belief or intention as distinct psychological states. It only models the linguistic expression of those states.

Thus, there is no transparency. Even when an LLM generates "I believe that p," it is not thereby authoritatively self-attributing a belief—it is merely echoing a form. There is no "I" doing the mental act, and no sense in which the model controls its attitude. Without mental action, there can be no practical knowledge—and without practical knowledge, no authoritative self-knowledge.

Peacocke's account introduces a subtle and powerful idea: a mental act like "judging that p" can have a plural content; p is simultaneously a first-order thought (about the world) and a second-order one (about one's own mind). For example, when I say "It's going to rain," I am both asserting a weather prediction and expressing my belief.

Peacocke points out that this dual identity is possible because human thought is embedded: a single cognitive act can serve multiple purposes, including reflective self-ascription. LLMs, however, have no such embedding. Their outputs may appear to have multiple layers ("I believe that p" both reports and asserts), but they are not embedded in an ongoing agentive context; the model does not "know" that it is making a claim, nor does it understand that it is referring to its own mental states—because it has none.

There is no aboutness in the robust sense. The model lacks a self to embed, and its outputs lack the anchoring context of an agent who stands behind the words. Thus, content plurality collapses into a flat string of symbols—grammatically rich, semantically vacuous.

Peacocke also highlights Moorean absurdities, such as saying, "It's raining, but I don't believe it's raining." These statements strike us as incoherent not because they are logically contradictory, but because they undermine the speaker's epistemic authority. If you assert p, that act shows you believe p; to deny the belief while asserting the proposition defeats the performative logic of speech.

LLMs frequently produce similar absurdities, but they are not aware of them. They can assert "I am an AI trained by OpenAI" and "I do not know who created me" in the same breath. These statements can be technically accurate by different standards, but there's no self behind them navigating the coherence of belief and expression.

This points to a deeper absence. While humans have a diachronic self—a persisting, reflective identity that tracks beliefs and intentions over time—LLMs do not. Each query-response is generated afresh, with no unified agency integrating attitudes or checking for internal contradiction. Without a temporally extended self, there is no ground for the kind of self-knowledge Peacocke describes.

If we take Peacocke seriously, then current LLMs lack not just human-level intelligence but the basic architecture required for true self-knowledge. They cannot know themselves because they have no self to know, and no mental actions through which to perform the knowing.

Yet this is not just a limitation—it is a diagnostic insight. Peacocke's framework gives us a checklist for artificial self-knowledge: control over attitudes, capacity for mental action, content plurality, Moorean coherence, and diachronic unity. These are not mere features to mimic in language—they are structural preconditions for being a knower.

In the future, if AI systems develop sustained memory, goal-setting behavior, or embedded self-monitoring capacities, we might ask: Are they performing mental actions? Are they adopting attitudes, not just simulating them? Peacocke's view gives us the philosophical tools to ask these questions rigorously, without being seduced by surface-level fluency.

Antonia Peacocke's account of self-knowledge shows that knowing oneself is not just a matter of having the right information—it is a matter of doing. Transparent self-attribution, mental action, and content plurality together make human self-knowledge not only possible but authoritative. By contrast, LLMs do not know themselves—not because they are defective humans, but because they are not agents at all. By applying Peacocke's philosophy to artificial minds, we can better understand what true knowing involves—and what it would take for machines to join us in the first-person perspective.