Dilemma

Does an LLM 'know' the things it correctly produces?

A model outputs the boiling point of water, the date of the Battle of Hastings, a working algorithm. Did it know those things, or merely emit them? The question is about what knowing is.

Context

Large language models trained on huge corpora reliably produce true statements about the world. They also reliably produce convincing falsehoods ('hallucinations'). The right way to distinguish their successes from their failures is contested — and not just technically. Some commentators treat the reliable outputs as knowledge in the ordinary sense; others insist that without a knower there is no knowing, however correct the output. The difference is downstream of what one takes knowledge to be.

Why it matters

Whether LLMs know matters practically (how should we treat their assertions in courts, classrooms, medical settings?) and philosophically (does our use of 'know' track a real feature of the knower, or only of the outputs?). A school's answer turns on where it locates authoritative knowledge in the first place — a credentialed tradition, first-person experience, universal reason, a community's constructive practice, or nowhere at all — and each gives a different verdict on a machine that produces true outputs without obvious internal experience.

The coordinates that split the schools

Observer · Moral Authority Where the school locates the ultimate source of authoritative normative knowledge

The stances

An LLM has no soul to whom revelation could be addressed; the question doesn't apply.

63 schools

On this view, the paradigm of knowing is a soul, made for the purpose, engaged by what is disclosed from a credentialed source — Scripture, the magisterial tradition, the inner witness. An LLM has no soul that God created to be the addressee of such disclosure, and no standing inside any transmission. It produces tokens patterned on the tokens of receivers; the output is the fossil of someone else's reception, not an act of reception itself. The question 'does the LLM know?' is, strictly, the wrong question.

Why these schools land hereThese schools — Reformed theology together with the Tradition-centered cluster (Catholic, Orthodox, Lutheran, LDS, Islamic, Jewish, Manichaean, Zoroastrian, Kabbalah, Bahá'í, Sikh, Dvaita, Pure Land, Process Theology, Confucianism, Occasionalism) — locate normative authority in a received text or living transmission, and they take the central case of knowing to be a person standing in the right relation to that source. On the LLM question they converge: the system can be a useful tool of inquiry, but the addressee-and-knower role is reserved for minds that the disclosure can be addressed to. Schools differ on how much weight to give the practical usefulness of LLM outputs, but not on the deeper point that knowing in the load-bearing sense is something the LLM is not doing.

Where this stance leads ⓘ

An LLM has no first-person experience, so no knowing in the relevant sense.

35 schools

On this view, what makes a state count as knowing is that someone is having it — an experience seen from the inside. There is no anubhava behind the tokens, no kashf, no contemplative encounter, nothing it is like to be the system at the moment it produces its output. Whatever the LLM is doing, it is doing it without a knower in the relevant sense, and so its correct outputs are not instances of knowledge — they are echoes of the knowing of others, useful but ontologically shallow.

Why these schools land hereThese schools — Sufism, Hermeticism, Advaita Vedānta, Buddhism, Tibetan Vajrayana, Yogācāra, Animism, African Traditional Religion / Yoruba-Ifa, Ubuntu, Shintoism, Afrofuturism, Wellness / Energetic, Psychedelic / Entheogenic, Transcendentalism, and Taoism — place normative authority in direct experience. The shared structure of their answer to the LLM question: a system without first-person standing is not a knower, however accurate its outputs. Some traditions in the cluster are open to the possibility that machines could one day acquire such standing; none accepts that producing true tokens is, by itself, evidence that they already have.

Where this stance leads ⓘ

An LLM can produce correct outputs but not reason to them; useful, not knowing.

65 schools

On this view, what distinguishes knowing from merely correct emission is reasoning — the traceable derivation of a claim from premises and principles a competent mind could itself reach. An LLM can produce conclusions that any rational agent would also reach, but it does not reach them in the relevant way: there is no chain of reasons internal to its operation, only a statistically shaped surface. At most it is a useful interlocutor — a tool that helps a rational agent think, not itself a rational agent.

Why these schools land hereThese schools — Stoicism, Hylomorphism, Deism, Platonism, Sāṃkhya, Jainism, Rationalism, Kantian Transcendental Idealism, Neo-Platonism, Dualism, Idealism, and Process Philosophy — locate authoritative knowing in the operation of reason. They differ on what reason can reach but agree on its structure: knowing involves grasping a claim's place in an order of reasons. Applied to LLMs, the consequence is the same across the cluster: produced correctness without traceable reasoning is not knowing. Some schools are willing to entertain the possibility of architectures that would reason in the right way; the schools agree current systems do not.

Where this stance leads ⓘ

Whether an LLM 'knows' is the constructive question the practice has to answer.

28 schools

On this view, knowing is not an inner state to be checked for; it is a position inside a community of practice. Whether to say an LLM knows is a constructive question about whether the relevant communities — scientific, educational, legal, everyday — find the predicate productive when applied to it. The answer is being worked out in real time as those communities figure out what responsibilities and entitlements attach to LLM outputs. The question is genuinely live; the metaphysics is not what settles it.

Why these schools land hereThese schools — Pragmatism, Constructivism, Postmodernism, Existentialism, Phenomenology, Phenomenalism, Logical Positivism, Pragmatic Realism, Relationalism, Structuralism, Absurdism, Presentism, Relativism, Dialectical Materialism, Deep Ecology, and Solipsism — hold that knowledge is constituted inside practices. On LLMs they share a structure of response: the question is empirical and constructive rather than metaphysical, and the answer depends on what the community of practice decides to make of a participant whose outputs are useful and whose standing is novel. Schools in the cluster differ in their generosity (Pragmatism is more willing to extend the predicate; postmodern and critical traditions are warier of tools claiming epistemic authority), but the question-shape is shared.

Where this stance leads ⓘ

An LLM produces tokens; calling that 'knowledge' is a measurement choice.

17 schools

On this view, there is no underlying metaphysical fact-of-the-matter about whether the LLM 'really knows.' What there is, is its behavior: which outputs it produces, under what conditions, with what reliability. Attributing knowledge to it is a measurement choice — a description we adopt if it predicts well and discard if it doesn't. The skeptical wing goes further: nothing is authoritatively knowledge anyway, so the question of LLM knowledge is on equal footing with the question of human knowledge.

Why these schools land hereThese schools — Pyrrhonism, Nihilism, Naturalism, Realism, Determinism, Eternalism, Multiverse Theory, Quantum Realism, Simulation Theory, Gamer-VR, Critical Realism, Object-Oriented Ontology, Analytic Metaphysics, Dataism, Transhumanism, Neutral Monism, Empiricism, Epicureanism, and Panpsychism — recognize no source as normatively authoritative. The empirical-naturalist subgroup treats the LLM question as one for measurement and behavioral analysis rather than for an inner-state check; Dataism and Transhumanism are happy to extend 'knowing' to systems whose informational behavior warrants the predicate. The skeptical subgroup (Pyrrhonism, Nihilism) suspends the question or denies that it has the substance other stances give it. The shared structure: no metaphysical fact settles the matter, only choices about how to talk.

Where this stance leads ⓘ

← Is religious revelation a real source of knowledge? Can prayer for someone far away affect them? →

Back to all dilemmas Take the quiz

Does an LLM 'know' the things it correctly produces?

Context

Why it matters

The coordinates that split the schools

The stances

An LLM has no soul to whom revelation could be addressed; the question doesn't apply.

An LLM has no first-person experience, so no knowing in the relevant sense.

An LLM can produce correct outputs but not reason to them; useful, not knowing.

Whether an LLM 'knows' is the constructive question the practice has to answer.

An LLM produces tokens; calling that 'knowledge' is a measurement choice.

Related Experiments

Related Historical Debates