Miranda Fricker's seminal work on epistemic injustice identified a category of harm long overlooked by mainstream philosophy: the wrong done to someone specifically in their capacity as a knower. Testimonial injustice occurs when prejudice causes a hearer to assign deflated credibility to a speaker's word. Hermeneutical injustice occurs when gaps in collective interpretive resources place someone at an unfair disadvantage in making sense of their social experiences.

These concepts have proven philosophically generative, but they have largely remained at the level of qualitative description. This presents an opportunity. Formal epistemology, with its apparatus of probability functions, likelihood ratios, and evidence updating, offers tools for transforming these intuitive notions into precisely measurable quantities.

What follows is an attempt to render epistemic injustice mathematically tractable. We will model credibility deficits as systematic distortions in Bayesian likelihood assignments, examine how seemingly rational updating can entrench prejudicial priors, and explore what these formal results imply for normative correction. The aim is not to reduce injustice to arithmetic, but to demonstrate that mathematical precision can sharpen, rather than flatten, our grasp of how knowledge practices go wrong.

Consider a hearer H evaluating testimony from speaker S regarding proposition p. In standard Bayesian terms, H updates her credence Pr(p) using the likelihood ratio Pr(testimony | p) / Pr(testimony | ¬p). This ratio encodes how reliably S's testimony tracks the truth—what we might call S's epistemic credibility from H's perspective.

Testimonial injustice can be formalized as a systematic deflation of this likelihood ratio when S belongs to a marginalized group g. Define the credibility deficit δ(g) as the difference between H's assigned likelihood ratio for members of g and the likelihood ratio she would assign to a comparably situated speaker from the dominant group. When δ(g) > 0 in the absence of any evidential basis, we have a measurable instance of testimonial injustice.

This formalization permits empirical inquiry. Audit studies, courtroom transcripts, and peer-review records can be analyzed for systematic likelihood deflation. Crucially, the framework distinguishes warranted credibility differentials, which track genuine reliability differences, from prejudicial ones, which do not. The mathematical structure makes this distinction crisp where ordinary discourse leaves it muddled.

The formalism also reveals something subtle: small per-encounter deficits compound. If a speaker faces δ(g) = 0.1 across thousands of testimonial exchanges, the cumulative effect on her epistemic standing—measured as expected influence on community credences—can be enormous, even when each individual instance appears trivially unjust.

Hermeneutical injustice admits a parallel treatment via information theory. The interpretive resources available to a community can be modeled as a code with finite expressive capacity. When experiences common to group g lack adequate encoding, members of g face higher description-length costs and lower mutual information between their experiences and their communicable utterances.

Takeaway

Injustice that resists qualitative dismissal becomes undeniable once quantified; the cumulative arithmetic of small deflations is rarely small.

A troubling result emerges when we examine how rational updating interacts with biased priors. Suppose H begins with a prior belief that members of group g are less reliable testifiers than average. She then encounters testimony from g-members and updates accordingly. Under what conditions will her posterior credence converge to the truth?

Standard convergence theorems require that the evidence be diagnostic—that observations have different probabilities under competing hypotheses. But when H's prejudicial prior causes her to discount g-testimony from the outset, she generates fewer opportunities to observe its accuracy. The evidence she gathers is filtered through her prior, producing what statisticians call selection bias.

Formally, if the probability that H attends carefully to g-testimony depends on her prior credibility assignment, then the sequence of observations she collects is not independent of her belief state. Convergence to truth fails. Her posterior can remain stably miscalibrated indefinitely, even as she scrupulously applies Bayes' rule at every step.

This is the formal signature of what feminist epistemologists call epistemic closure: a self-sealing belief system that appears internally rational while systematically excluding correcting evidence. The mathematical lesson is sobering—local rationality does not guarantee global accuracy when evidence-gathering is itself belief-dependent.

The result generalizes. Any agent whose attention, retention, or interpretation of testimony correlates with prior credibility assignments will exhibit path-dependent belief trajectories. Two agents with slightly different priors can diverge permanently, both updating impeccably, both settling on incompatible posteriors.

Takeaway

Rationality at each step does not entail rationality across time; an agent can update flawlessly toward a false conclusion when her evidence stream is shaped by her own biases.

If biased priors generate self-confirming evidence, individual rationality cannot be the corrective. Formal models suggest several structural interventions whose effects can be precisely characterized.

The first is credibility inflation: deliberately raising one's likelihood assignment for marginalized speakers above one's intuitive default. This appears irrational from a single-agent perspective, but in expectation it counteracts known prior distortion. The mathematics resembles regularization in statistical learning—introducing controlled bias to reduce variance and improve out-of-sample performance.

The second is forced exposure: institutional mechanisms that ensure attention to testimony regardless of credibility prior. Blind review, structured interviews, and mandated consultation interrupt the selection bias that sustains epistemic closure. Formally, these mechanisms decouple evidence-gathering from prior credence, restoring the conditions under which Bayesian convergence theorems apply.

The third concerns hermeneutical resources. If interpretive codes have insufficient expressive capacity for marginalized experiences, the remedy is concept-creation: extending the code. Terms like sexual harassment, gaslighting, and microaggression represent measurable reductions in description-length cost for previously inarticulable experiences.

These interventions are not merely ethical impositions on epistemology. They are formally derivable from the goal of long-run accuracy under realistic conditions of bounded attention and prior uncertainty. Justice and truth-tracking, often portrayed as rivals, turn out to be coupled when we model knowledge practices honestly.

Takeaway

Sometimes the most rational response to your own rationality is to constrain it; structural correctives are not concessions to fairness but prerequisites for accuracy.

Formal models do not dissolve epistemic injustice into equations, but they do something valuable: they make the phenomenon measurable, the dynamics predictable, and the corrections derivable. What philosophical description identified as wrong, mathematics shows to be quantifiably costly—both to those harmed and to the community's collective accuracy.

The deeper result is that local Bayesian rationality, long treated as the gold standard of epistemic virtue, is insufficient. When evidence-gathering depends on prior beliefs, rational agents can converge on falsehoods and remain stably wrong. Justice in our knowledge practices is not a luxury added atop epistemology; it is structurally necessary for epistemology to deliver on its own promises.

The formal turn in social epistemology is still young. Its instruments are sharp, but its applications have only begun. The question is no longer whether epistemic injustice exists, but how precisely we are willing to measure it.