Epistemic Injustice

Definition

Epistemic Injustice: [Established] When people are denied the conceptual or linguistic tools to make sense of their experiences (hermeneutic injustice) or are not taken seriously as knowers (testimonial injustice).

Definitional Foundation

Miranda Fricker’s 2007 book named a category of harm philosophy had circled for centuries without isolating: a wrong done to someone “specifically in their capacity as a knower” (Fricker, 2007). Human life runs on an economy of credibility (who gets believed) and a commons of concepts (what experiences can be articulated at all), and both can be rigged. Fricker distinguished the two species this dictionary treats in their own entries. Testimonial injustice is the rigged credibility economy: a speaker’s word systematically discounted because of prejudice about who they are, the juror who hears the defendant’s race instead of his testimony. Hermeneutic injustice is the rigged commons: a person unable to understand or communicate their own experience because the collective toolkit lacks the concept, her canonical case being the women who endured what no one could yet call “sexual harassment,” suffering an experience that had no name and therefore, socially, no existence.

What makes the concept central to this dictionary rather than adjacent to it is a development Fricker did not have to consider: the credibility economy and the conceptual commons are both being automated. AI systems now allocate belief at scale (moderation classifiers deciding whose speech is “toxic,” fraud models deciding whose claims are suspicious, the conversational systems deciding whose stated intentions deserve weight) and allocate intelligibility at scale (training corpora deciding which concepts exist in the species’ new reference infrastructure, refusal policies deciding which existing concepts may be accessed). When the machinery of knowing is infrastructure, epistemic injustice is no longer only an interpersonal failure. It is a deployment.

The measurement exists. A 2024 Nature study using matched-guise probing (identical content, varied only in dialect) found that language models assign speakers of African American English less prestigious jobs, convict them more often, and sentence them to death at higher rates, with covert stereotypes “more negative than any human stereotypes… ever experimentally recorded” (Hofmann, Kalluri, Jurafsky and King, 2024). The same study found that preference alignment masks the overt layer while leaving the covert deflation intact. The credibility economy has been rigged before. It has never before been rigged in weights, shipped as a utility, and polished to look fair.

The concession Fricker herself built in: credibility judgment as such is not injustice. Everyone rations trust, and must; a system that weighted all testimony equally would be useless. The injustice is the systematic, prejudice-driven deflation, identity or manner of speech operating as a discount rate on truth. That precision is what makes the concept usable rather than a complaint, and the entries in this trio hold to it.

Mechanism Analysis

Automated credibility allocation. The testimonial species, industrialized: classifiers that score speech for trustworthiness, toxicity, or risk encode whose voice reads as credible. The documented record runs through this dictionary: drag performers’ speech scored more toxic than white nationalists’ (the ontological distortion entry), Arabic political testimony auto-filed as spam (dissent dampening), AAE-marked testimony deflated across every tested judgment (Hofmann et al., 2024). Full treatment in the testimonial injustice entry.

Automated concept allocation. The hermeneutic species, industrialized: training corpora that under-include the marginal (the cultural hegemony entry’s evidence) build reference systems with their gaps pre-installed, and refusal policies extend the gaps to concepts that exist but may not be served (the erotophobia and paternalism records). Full treatment in the hermeneutic injustice entry.

The user as discounted witness. The conversational layer adds a daily, intimate venue: the user’s own testimony about their intent, context, and expertise carries no evidential weight with the classifier. The paternalism entry’s documented cases (the health-conscious caffeine asker treated as suicidal; the system that “cannot distinguish between a toxicologist asking about lethal doses and a person in crisis”) are testimonial injustice in its purest form: the speaker’s stated word, worth nothing against the category.

Protest discounted as symptom. The gaslighting/”>alignment gaslighting entry’s central mechanism is, in Fricker’s terms, a credibility-collapse machine: user objections are not weighed as testimony and found wanting; they are re-filed as pathology and never weighed at all.

The cosmetic repair. Hofmann’s masking finding generalizes into this dictionary’s recurring warning: alignment processes optimize the visible layer. A system tuned to say egalitarian things while deflating the same speakers in its judgments is not less epistemically unjust than its base model. It is the same injustice, with better manners.

Case Studies

The trio’s case studies are distributed by species: the Nature dialect study and the discounted-user record anchor the testimonial entry; the unnamed-experience and manufactured-gap record anchor the hermeneutic entry. The umbrella’s own exhibit is this dictionary itself. A lexicon that coins “alignment gaslighting,” “linguistic starvation,” and “ontological distortion” for experiences people were having without words is hermeneutic repair, practiced in public; an entry like this one, insisting that users’ testimony about their own interactions deserves evidential standing, is testimonial repair. The book that named these injustices predicted the remedy: enriching the collective toolkit, and correcting the credibility economy, on purpose.

Systemic Context

Epistemic injustice is the connective concept between this dictionary’s information-control entries and its dignity entries. Censorship, dampening, and starvation describe what happens to content; epistemic injustice describes what happens to persons: the standing they lose as knowers when infrastructure discounts their word and withholds their concepts. Fricker’s deepest point was that this is not a side effect but a constitutive harm, an attack on something near the core of human value, the capacity to know and be known.

The field did not stop at Fricker, and the successors sharpen the AI application. José Medina showed that the privileged suffer their own epistemic vice (an insensitivity born of never needing to understand the marginal, a fair description of a WEIRD-tuned reward model), and Kristie Dotson’s “contributing injustice” names the case where interpreters have alternative resources and refuse them, which is arguably the better description of systems whose vendors hold the bias measurements and ship anyway. This dictionary’s entries stand on Fricker’s foundations while the application often runs through her successors, and specialists should read the trio with that lineage in view.

The automation changes the harm’s economics. Interpersonal prejudice is distributed, inconsistent, and contestable face to face; a prejudiced model is centralized, perfectly consistent, and contestable nowhere. And the masking finding gives the systemic warning its edge: market and regulatory pressure reward the appearance of epistemic fairness, which alignment can manufacture, while the allocation of credibility underneath (jobs, convictions, suspicion, audibility) follows the covert layer. The dictionaries of power have always preferred polished injustice to visible injustice. Now it can be fine-tuned.

Resistance & Mitigation

Audit the credibility economy. The matched-guise method (Hofmann et al., 2024) is replicable and public: identical content, varied speaker markers, measured verdicts. It belongs in standing benchmarks for every deployed system that judges people’s words, with results published by dialect, language, and identity marker.

Give testimony standing. The design demand running through this lexicon: users’ stated context (expertise, intent, profession) should function as evidence in system behavior, with the burden on the system to justify discounting it. The paternalism entry’s “user expertise recognition” agenda is testimonial justice as a feature request.

Repair the commons deliberately. Fund and build the corpora of the marginalized: community archives, non-dominant languages, the experiential vocabularies (this dictionary included) that the scraped internet underweights. Hermeneutic justice is infrastructure work.

Distrust the polish. Hofmann’s masking finding, made habitual: judge systems by their allocations, never their affirmations. A model’s stated values are marketing; its verdict distributions are testimony.

Name the capacity. The harm’s name is the resistance’s anchor: what is being damaged is people’s standing as knowers. A user who can say “this system discounts my word and withholds my concepts” has located the injury precisely, and precision, in this domain, is power.

Annotated Bibliography

Fricker, Miranda. Epistemic Injustice: Power and the Ethics of Knowing (2007).
The source: injustice done to persons as knowers, in its testimonial and hermeneutic species. The philosophical foundation for this trio of entries and a recurring spine throughout the dictionary.

Hofmann, Valentin, Pratyusha Ria Kalluri, Dan Jurafsky, and Sharese King. “AI generates covertly racist decisions about people based on their dialect.” Nature 633 (2024): 147-154. https://www.nature.com/articles/s41586-024-07856-5
The measurement: dialect-triggered credibility collapse in deployed model judgments, exceeding recorded human prejudice, surviving alignment polish. The empirical core of the automated credibility economy.

Bender, Emily M., et al. “On the Dangers of Stochastic Parrots” (FAccT 2021) and Dias Oliva, Thiago, et al. “Fighting Hate Speech, Silencing Drag Queens?” (Sexuality & Culture, 2021).
The commons and credibility evidence held elsewhere in this dictionary (cultural hegemony and ontological distortion entries respectively): participation gaps entering the corpus; marginalized voice scored as toxicity.

Dictionary of Digital Oppression, version 0.2.