Classical epistemology treats propositions as sharply true or false, and probability theory assigns credences accordingly. But natural language is riddled with predicates that resist sharp boundaries. When does a person become tall? At what precise grain count does a collection become a heap? Vagueness is not epistemic laziness—it is a structural feature of how predicates carve up continuous domains. The question that concerns us here is what happens when the formal machinery of Bayesian epistemology collides with the semantics of vague language.

This collision is not merely academic. Probabilistic reasoning underpins decision theory, machine learning, and scientific inference. If the propositions over which we define probability measures are themselves indeterminate in truth value, then the foundations of credence assignment require reexamination. A credence of 0.7 in the proposition "Jones is tall" conflates two fundamentally different sources of uncertainty: uncertainty about Jones's height, and the semantic indeterminacy of tall itself. Disentangling these is non-trivial.

Three major frameworks in the philosophy of vagueness each interact with probability theory in distinct and sometimes surprising ways. Supervaluationism preserves classical logic at the cost of truth-value gaps. Degree-theoretic approaches replace bivalence with a continuum of truth values. And higher-order vagueness threatens any clean resolution with an infinite regress of borderline cases. Each framework forces different formal commitments upon the epistemologist, and each reveals something important about the architecture of rational belief under semantic indeterminacy.

Supervaluationism, developed most thoroughly by Kit Fine, handles vagueness by quantifying over admissible precisifications—ways of making a vague predicate precise that are consistent with its usage. A sentence is super-true if true on all admissible precisifications, super-false if false on all, and indeterminate otherwise. Classical tautologies remain super-true. The law of excluded middle is preserved: "Jones is tall or Jones is not tall" holds on every precisification, even if neither disjunct does.

Now consider defining a probability measure over a supervaluationist semantics. The natural approach is to introduce a two-dimensional probability space: one dimension ranges over possible worlds (capturing empirical uncertainty), the other over admissible precisifications (capturing semantic indeterminacy). A credence in a vague proposition becomes an expectation over both dimensions. Formally, let W be a set of worlds, S a set of admissible precisifications, and P a probability measure over W × S. The credence in proposition φ is then P({(w,s) : φ is true at w under s}).

This framework yields a surprising result. If the agent's credence function marginalizes cleanly over precisifications—treating semantic and empirical uncertainty as independent—then borderline cases receive intermediate credences even when the agent has complete empirical information. Suppose the agent knows Jones is exactly 180 cm. If 60% of admissible precisifications classify 180 cm as tall, the credence in "Jones is tall" is 0.6. But this credence does not reflect uncertainty about the world. It reflects uncertainty about meaning.

This is philosophically significant because it challenges the standard interpretation of credences as degrees of belief about how the world is. Under supervaluationist probability, some credences are better understood as degrees of semantic commitment—measures of how determinately a predicate applies. The Bayesian updating machinery still functions: new empirical evidence shifts the world-dimension of the probability space while leaving the precisification-dimension intact. But the interpretation of the resulting credence is fundamentally hybrid.

A further complication arises with conditionalization. If an agent conditionalizes on a vague proposition—say, receiving testimony that "Jones is tall"—the update must propagate through both dimensions simultaneously. The posterior will depend on the agent's prior distribution over precisifications, which introduces a form of semantic prior that has no obvious empirical constraint. This is where supervaluationist probability departs most sharply from orthodox Bayesianism: some priors are about language, not the world, and no amount of evidence can determine them.

Takeaway

When propositions are vague, credences become hybrid objects—part empirical uncertainty, part semantic indeterminacy. Recognizing which dimension of uncertainty a credence tracks is essential before any Bayesian update can be properly interpreted.

Degree-theoretic (or fuzzy) approaches to vagueness reject bivalence outright. Instead of a proposition being true or false, it takes a truth value in the interval [0,1]. The predicate tall becomes a function from heights to truth values, typically modeled as a smooth sigmoid or similar curve. A person of 190 cm might have truth value 0.95 for tall; a person of 170 cm, 0.4. The sorites paradox dissolves because each small increment changes the truth value only slightly—there is no sharp boundary to locate.

The formal question is immediate: if truth values live in [0,1] and credences live in [0,1], how do they interact? Suppose you know with certainty that Jones is 175 cm, and your semantic function assigns truth value 0.6 to "a 175 cm person is tall." Is your credence in "Jones is tall" then 0.6? If so, the credence is not expressing uncertainty at all—it is mirroring the truth value. This identification of credence with truth value has been defended by some theorists, but it collapses two conceptually distinct notions.

A more careful treatment separates the layers. Let v(φ,w) denote the degree of truth of φ at world w, and let P be a probability measure over worlds. The agent's expected truth value of φ is then ∫ v(φ,w) dP(w). Under empirical certainty, this reduces to the degree of truth at the actual world. Under genuine uncertainty, it integrates over possible truth values weighted by their probability. This expected truth value is a natural candidate for the agent's credence in a vague proposition, and it preserves the distinction between the two sources of gradation.

The algebraic consequences, however, are non-trivial. Standard probability theory requires that P(φ ∨ ¬φ) = 1. But in many-valued logics, the truth value of φ ∨ ¬φ under a standard fuzzy disjunction (max operator) is max(v, 1−v), which is less than 1 for borderline cases. If credences track expected truth values, then P(φ ∨ ¬φ) < 1 becomes possible. This is a direct violation of a Kolmogorov axiom. The formal epistemologist must then choose: abandon additivity, restrict the domain to precise propositions, or reinterpret the axioms within a non-classical framework.

Several proposals navigate this. Field (2003) develops a conditional-credence framework that replaces standard conditionalization with a primitive conditional probability suited to degree-theoretic truth. Smith (2008) argues for a dual measure approach: one measure for empirical uncertainty, one for degree of truth, combined only at the point of decision. Each proposal preserves some classical structure at the cost of complicating others. The philosophical lesson is clear: once truth admits degrees, the probability calculus cannot remain unchanged. Something in the classical package must give.

Takeaway

When truth itself comes in degrees, credences can no longer be pure measures of uncertainty—they become entangled with semantic gradation. Any formal epistemology of vague language must decide which axioms of probability to preserve and which to revise.

Both supervaluationism and degree theory face a deeper challenge: higher-order vagueness. The predicate tall is vague, but so is the predicate borderline tall. There is no sharp boundary between the clearly tall and the borderline tall, nor between the borderline tall and the clearly not tall. Any attempt to draw these boundaries generates new borderline cases at a higher order. The regress is structurally similar to the original sorites, and it infects every formal framework that attempts to model vagueness with precise mathematical tools.

For supervaluationism, higher-order vagueness means that the set of admissible precisifications is itself vaguely bounded. Which precisifications count as admissible? If this is a vague matter, then the two-dimensional probability space described earlier inherits indeterminacy in its second dimension. The measure over precisifications becomes ill-defined unless we introduce a meta-level set of admissible sets of precisifications—and the regress begins. Williamson (1994) uses this regress to argue that supervaluationism ultimately collapses into epistemicism, but the formal epistemological consequences are independently severe: the semantic priors required for conditionalization on vague propositions may themselves be indeterminate.

For degree theory, higher-order vagueness implies that the semantic function v mapping heights to truth values is not a precise function. If the assignment of truth value 0.6 to 175 cm is itself a vague matter—perhaps 0.6 is as good as 0.58 or 0.62—then the expected truth value integral ∫ v(φ,w) dP(w) becomes a set-valued or interval-valued computation. Interval-valued probabilities have been studied by Walley, Weichselberger, and others under the heading of imprecise probability, and this connection is not accidental. Higher-order vagueness naturally pushes degree-theoretic epistemology toward imprecise credence frameworks.

The convergence is striking. Whether one starts from supervaluationism or degree theory, higher-order vagueness drives the formal apparatus toward sets of probability functions rather than single measures. This resonates with the imprecise probability tradition independently motivated by concerns about prior elicitation, ambiguity aversion, and the limits of rational belief. The suggestion, then, is that vagueness is not merely an annoyance for formal epistemology—it is a fundamental reason to abandon sharp credences in favor of credal sets or interval-valued probabilities.

Yet even imprecise probabilities face the regress at the next level. The boundaries of the credal set—which probability functions are included?—may themselves be vague. One response is to embrace the regress as a structural feature of rational belief under vagueness, accepting that precision is unattainable at every level. Another, following Sainsbury's notion of boundaryless concepts, is to develop topological rather than measure-theoretic models of belief, where the relevant mathematical structure is open sets rather than σ-algebras. These remain active research programs, but the philosophical takeaway is already clear: the formal epistemology of vague propositions is irreducibly more complex than its classical counterpart, and the complexity is not a defect of our models but a faithful reflection of the phenomena.

Takeaway

Higher-order vagueness pushes formal epistemology toward imprecise credences and away from single probability functions—not as a concession to human limitation, but as a structural demand of the subject matter itself.

The intersection of probability theory and the epistemology of vagueness reveals that our standard formal tools are calibrated for a sharper world than language delivers. Supervaluationism splits credences into empirical and semantic components. Degree theory entangles truth value with belief. Higher-order vagueness destabilizes both frameworks and points toward imprecise probability as the natural habitat for vague propositions.

These are not merely technical complications. They expose a foundational question: what is a credence about when the proposition it targets lacks determinate truth conditions? The answer reshapes how we understand evidence, updating, and rational belief under the realistic conditions of natural language.

Formal epistemology gains, rather than loses, by confronting vagueness directly. The resulting models are richer, more honest about the limits of precision, and better suited to the messy semantic landscape in which actual reasoning takes place. The sharpness of our mathematics need not pretend that the world—or our words—are equally sharp.