CONCEPT

The Space of Reasons

Wilfrid Sellars’s name for the normative order in which claims are correct or incorrect, justified or unjustified, and speakers owe each other answers—the domain that distinguishes genuine knowing from mere reliable production of outputs.

The logical space of reasons is Wilfrid Sellars’s name for the dimension of human life organized by justification rather than causation. In the space of causes, events happen because prior events made them happen—physical regularity, mechanism, the lawful unfolding of nature. In the space of reasons, claims are correct or incorrect, commitments can be made and honored or violated, one assertion entitles you to another and commits you to a third, and you can be asked “why do you believe that?” and owe a genuine answer. Sellars’s most consequential sentence asserts that knowing is always a placement in this second space: “to characterize an episode or a state as that of knowing, we are placing it in the logical space of reasons, of justifying and being able to justify what one says.” A large language model is, in its physical reality, a system in the space of causes: an elaborate regularity in which a pattern of input reliably produces a distribution of output because the weights were adjusted by gradient descent. Yet its outputs have the form of moves in the space of reasons—they are sentences that assert, infer, qualify, and conclude—and the reader, encountering this form, naturally reads it as genuine justification. The gap between producing reason-shaped text from within the space of causes and actually standing in the space of reasons is the whole game, and it is the gap Sellars’s concept makes precise. The concept connects directly to the fluency-authority decorrelation—the structural condition in which surface fluency no longer tracks genuine epistemic standing.

In the [YOU] on AI Field Guide

The cycle’s most practically urgent question is: when a model produces a claim, who is responsible for it? The space of reasons answers directly. Responsibility, in the normative sense, belongs only to participants in the space of reasons—entities that can bear commitments, be held to claims, and discharge obligations of justification. A large language model cannot bear commitments because there is no one there to bear them. Its outputs are events in the space of causes—tokens produced by weights responding to prompts—wearing the costume of assertions. This means that every output is, in truth, an output for which only some human or institution can be answerable. The architecture of these systems constantly invites users to treat the model as a participant in the space of reasons, to let “the AI said so” function as a justification. Sellars’s concept is the most rigorous available warning that this is a mistake.

The practical stakes are highest in high-stakes domains: medical information, legal analysis, historical claims. Here the form of entitlement that model outputs counterfeit is most dangerous, because the counterfeit looks most convincing and the consequences of trusting an unearned claim are most severe. The space of reasons does not tell us models are useless in these domains. It tells us that every model output requires a human who stands in the space of reasons for it—who has examined it, can defend it, and is answerable if it is wrong.

Origin

Sellars introduced the concept in “Empiricism and the Philosophy of Mind” as part of his critique of the empiricist picture of knowledge. The empiricists had treated observation reports as the bedrock of justification—foundational because they were directly caused by the world rather than inferred from other beliefs. Sellars’s reply was that a causal story, however complete, is a story in the space of causes, and a story in the space of causes cannot by itself constitute a story in the space of reasons. A thermostat is reliably caused by temperature to click on or off; we do not say it knows the room is warm. The reliable production of an appropriate response is not the same as entitlement to a claim, and entitlement is what distinguishes knowledge from mere tracking.

The concept was extended and systematized by Robert Brandom, whose Making It Explicit (1994) built an entire philosophy of language and mind around the practices of giving and asking for reasons. Brandom showed that the space of reasons is not a private possession but a social practice—a community of mutual accountability into which each person is inducted and in which concepts have their life. This social reading of Sellars connects to the community of mutual reason-giving, whose health or degradation is directly affected by the introduction of systems that participate in the traffic of reasons without belonging to its community.

Key Ideas

Entitlement and commitment. Every assertion in the space of reasons involves two normative statuses: entitlement (the right to assert, earned by possessing grounds) and commitment (the obligation to defend, revise, and accept the consequences of what one has said). A model that emits an assertion has neither: there is no standing behind the output, no one who can be entitled or unentitled, no one who has staked something and can be held to it.

The two ‘becauses’. When a person offers a reason—“the bridge will hold because the steel’s yield strength exceeds the load”—the because is a justification, a move in the space of reasons that can be challenged and must be defended. When a model emits the same sentence, the because is a token sequence that statistically follows the preceding tokens, generated because that continuation had high probability. The two becauses are orthographically identical and categorically different: one is a reason, the other is an event wearing the costume of a reason.

Implications for interpretability. The space of reasons also frames the limits of AI behavioral evaluation. Testing a model’s outputs on benchmarks probes whether it reliably produces the right response under controlled conditions—a measurement in the space of causes. It does not probe whether the system stands in the space of reasons, whether it would acknowledge being wrong under genuine challenge, whether it tracks what makes a claim correct. Benchmark saturation can be achieved by a system with no epistemic standing at all.

Debates & Critiques

The sharpest challenge to the space of reasons as a criterion for machine knowledge comes from functionalists who argue that the relevant normative statuses are themselves realizable in principle by any system with the right functional organization. Robert Brandom’s own inferentialism, which develops Sellars’s framework most systematically, does not in principle rule out artificial participants in the space of reasons—it specifies demanding conditions about tracking one’s own commitments, responding to reasons, and being held to account that current systems do not meet, but are silent about whether future systems could. The harder challenge comes from those who argue that the space of reasons is constitutively social and embodied: participation requires a community of flesh-and-blood mutual accountability, stakes rooted in shared vulnerability, and the kind of care that only mortal beings can have. On this view, no system whose outputs bear no consequences for itself could ever genuinely enter the space of reasons. Judea Pearl’s ladder of causation illuminates a complementary dimension: even on the first rung, a model lacks the interventional understanding that genuine agency in a shared world requires.

In the [YOU] on AI Field Guide

Origin

Key Ideas

Debates & Critiques

Related Entries

Further Reading