Wilfrid Sellars

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI asks what it means to see the machine clearly, without the narcotic of hype or the paralysis of fear. Sellars is the thinker who equips that seeing with its most precise instrument. When a model produces a medical diagnosis, a legal argument, or a historical claim with perfect fluency, two questions arise: Is this knowledge, and who is responsible for it? Sellars answers both from first principles. The output is not knowledge unless the system can stand behind it in the space of reasons—can be entitled to the claim, can owe and discharge a justification, can be held responsible if it is wrong. A system that produces the shape of justification without the standing of a justifier is not a knower; it is an instrument wearing the costume of a knower, and the difference is the whole game.

His critique of the Myth of the Given reframes the recurring scandal of machine bias as a structural inevitability rather than an engineering accident. There is no “raw” in raw data, because there is no uninterpreted layer beneath the conceptual choices of those who collected, labeled, and framed it. The aspiration to a view from nowhere, a dataset uncorrupted by any perspective, is the Myth of the Given wearing an engineer’s badge. The cycle asks readers to take the orange pill—to see this clearly—and Sellars is the philosopher who makes the seeing rigorous.

His inferential-role account of meaning converges, uncannily, on the technical foundation of these systems. Distributional semantics—the principle that a word is characterized by the company it keeps—is Sellars’s semantics restated in engineering terms. The model’s grasp of language is a real semantic competence by Sellarsian lights, capturing the intra-linguistic web of inference with extraordinary fidelity. Yet Sellars also identifies, with equal precision, what is missing: the language-entry and language-exit transitions that anchor words to a world the system perceives and acts in. The model has the middle of the language game—the inferential web—and not the ends. Fluency without grounding is exactly the structure his framework predicts.

Alongside Judea Pearl, who locates the machine’s limit on the first rung of causation, and John Searle, who argues from the biology of consciousness, Sellars stands in the cycle as the thinker who locates the limit in the structure of knowledge itself—in the normative practice of justification that no amount of fluency can substitute for. His work does not tell us machines cannot eventually be knowers. It tells us they are not yet, and it tells us why, and it tells us what they would have to become to change that verdict.

Origin

Sellars grew up inside the arguments of professional philosophy the way other children grow up inside a family trade, then studied at Michigan, Buffalo, and Oxford as a Rhodes Scholar before returning to teach in America. The work that made him essential was “Empiricism and the Philosophy of Mind,” delivered as lectures in London in 1956 and published shortly after—a dense, recursive, maddening text that contains, compressed into a few hundred pages, the two ideas that run through every chapter of his legacy. The first demolishes the epistemological foundation of logical positivism: the idea that knowledge rests on episodes in which the world simply delivers itself to the mind. To register anything as red, Sellars showed, already requires possessing the concept of red, which requires knowing how red contrasts with other colors, what counts as a standard observer, and when to trust the report. The supposedly bare given is already saturated with conceptual structure. There is no ground floor of pure registration beneath the conceptual building.

The second idea dissolves the commonsense picture of knowing as an inner event. Drawing on the philosophy of language and the pragmatist tradition, Sellars argued that knowledge is a normative status, not a natural fact. To say someone knows something is not to describe a state of their brain but to vouch for them—to say they have the standing to assert, the grounds to defend, and the responsibility to revise. This move had a consequence no one in 1956 could have foreseen: it placed the question of machine knowledge exactly where it belongs, in the domain of practice, not metaphysics. The question is not what is happening inside the system but whether the system can participate in the community of mutual justification that is the home of all conceptual life.

The Pittsburgh School he founded, whose most prominent heirs include Robert Brandom and John McDowell, extended his framework across decades. Brandom’s “inferentialism”—the systematic development of Sellars’s insight that meaning is conferred by role in patterns of inference—is now one of the live options in the philosophy of language, and it is the option most directly illuminated by the existence of systems that master inferential webs while remaining, in Sellars’s precise sense, unentitled to the claims they make.

Key Ideas

The Myth of the Given. The foundational error Sellars hunted through every corner of epistemology is the belief that knowledge has a bedrock of bare, uninterpreted facts delivered to the mind prior to any concept. The myth takes many forms—sense-data theory in traditional empiricism, raw data in the self-understanding of machine learning—but always makes the same mistake: it imagines an uninterpreted layer that actually cannot exist, because to register anything is already to bring it under a concept. The Myth of the Given is the philosophical foundation of the AI objectivity illusion.

The space of reasons. Sellars’s master distinction is between the space of causes, where events happen because prior events made them happen, and the logical space of reasons, where claims are correct or incorrect, justified or unjustified, and speakers owe each other answers. A large language model is a system in the space of causes that generates reason-shaped text—outputs that look like moves in the space of reasons without being so. The distinction explains why a model that produces a false sentence has not made an error in the full sense: making an error requires being the kind of thing that can be entitled and unentitled, right or wrong against a standard it is answerable to.

Inferential-role semantics. Meaning is not a relation between a word and a thing; it is the role a symbol plays in a web of moves. Sellars decomposed language into three kinds of move: language-entry transitions (the world prompts utterance), intra-linguistic transitions (utterances license other utterances by inference), and language-exit transitions (utterances issue in action). To mean something is to be governed by all three. The large language model has the second in abundance and the first and third barely at all—making its semantics real, partial, and finally adrift. This is the Sellarsian diagnosis of LLM fluency: genuine intra-linguistic competence, severed from both perceptual entry and worldly exit.

Entitlement and responsibility. If knowing is a normative standing rather than a natural fact, then the most urgent question about AI outputs is not whether they are correct but who stands behind them. A model cannot be entitled to its claims because there is no one there to bear the commitment or discharge the debt. Its “confidence” is the form of entitlement without the substance. The practical consequence: responsibility for every model output falls entirely on the humans who build, deploy, and rely on these systems. The architecture of large language models constantly invites users to forget this, and Sellars’s philosophy is the most rigorous warning against forgetting.

The Myth of Jones and self-knowledge. In the most startling part of his great essay, Sellars constructed a myth of how the concept of inner thought arose: a “genius” named Jones postulated inner episodes, modeled on overt speech, to explain intelligent behavior, and the theory became so successful that people learned to apply it to themselves. The inner, private self that introspection seems simply to find may be a theoretical construct, learned from the public and outer. Applied to AI: when we attribute beliefs, goals, and understanding to a model, we are doing what Jones did—and the question is whether the framework earns its application, which requires asking whether the system has the functional organization that makes the ascription true rather than merely natural.