Epistemic Injustice

Occurs when an AI system assigns a deflated level of credibility to a speaker's input based on prejudicial associations with their identity. This is the direct algorithmic translation of philosopher Miranda Fricker's core concept. In practice, this manifests when a model trained on historically biased data systematically undervalues information provided by certain demographic groups. Key mechanisms include:

• Voice analysis systems that assign lower accuracy confidence to non-native speakers
• Automated fraud detection that flags legitimate transactions from specific postal codes as higher risk
• Content moderation AI that disproportionately removes content from marginalized creators as 'toxic'
• Clinical decision support that discounts pain reports from certain patient populations The harm is not merely an error in output but a wrong done to the speaker in their capacity as a knower, degrading their epistemic agency.

Arises when an AI system's ontological framework—its categories, labels, and taxonomies—renders certain lived experiences unintelligible or impossible to express. This is a structural injustice embedded in the design of the model's world representation. When a classification schema lacks the conceptual resources to capture a group's social experience, those individuals cannot make their reality understood within that system. Examples include:

• Gender classification limited to binary male/female options, erasing non-binary identities
• Medical coding systems that lack diagnostic categories for conditions predominantly affecting marginalized groups
• Sentiment analysis trained only on Western emotional expression, misclassifying culturally distinct affect
• Content recommendation taxonomies that flatten diverse cultural production into dominant-culture genres The system creates a lacuna in collective understanding where certain injustices become structurally invisible.

The systematic exclusion or silencing of certain groups' knowledge contributions to the datasets that form the foundation of AI systems. Coined by Kristie Dotson, contributory injustice occurs when marginalized communities are denied the opportunity to contribute their epistemic resources to the collective pool of understanding. In AI, this manifests as:

• Underrepresentation in training datasets leading to model failure on minority populations
• Annotation labor where crowd workers from the Global South label data but their cultural interpretations are overridden by Western quality control
• Participatory design exclusion where affected communities are not consulted during problem formulation
• Language model training that privileges high-resource languages, treating low-resource languages as noise The result is a self-reinforcing cycle where excluded knowledge remains absent from the systems that increasingly govern social and economic life.

Occurs when an AI system treats individuals not as sources of knowledge but merely as objects of knowledge—data points to be classified, predicted, and acted upon without reciprocal epistemic engagement. This concept, drawn from Fricker's work, describes the reduction of a knower to a passive information source. In AI systems, this manifests when:

• Predictive policing treats community members solely as generators of risk scores rather than as informants with contextual knowledge
• Hiring algorithms reduce candidates to feature vectors without acknowledging their self-knowledge about capabilities
• Credit scoring substitutes algorithmic inference for an individual's testimony about their financial circumstances
• Social media ranking treats users as engagement-optimization targets rather than epistemic agents with communicative intent The individual is denied the status of an informant and instead becomes raw material for the system's knowledge production.

The unequal distribution of epistemic resources—including access to AI systems themselves, the ability to interrogate their outputs, and the capacity to contest their decisions. This structural dimension concerns who gets to benefit from and participate in AI-mediated knowledge practices. Manifestations include:

• Explainability gaps where only technically sophisticated users can understand and challenge automated decisions
• Asymmetric transparency where institutions deploy opaque models on populations while protecting their own operations from similar scrutiny
• Resource barriers preventing marginalized communities from developing or deploying counter-models
• Language monopolies where English-dominant models exclude non-English speakers from knowledge production This creates a two-tier epistemic order where the already powerful gain enhanced knowledge capabilities while the marginalized are rendered increasingly transparent to institutional surveillance.

A form of injustice occurring at the design stage when system architects preemptively discount or fail to seek the testimony of certain groups about their needs, risks, and contexts. Before any data is collected or any model is trained, decisions about what problems to solve and whose knowledge matters shape the entire pipeline. This manifests as:

• Problem formulation that prioritizes the concerns of paying customers over affected non-users
• Risk assessment frameworks that consult industry stakeholders but not civil society organizations
• Performance metric selection that optimizes for aggregate accuracy while ignoring distributional harms
• Deployment decisions made without consulting communities where systems will operate The injustice lies in the anticipatory dismissal of certain voices as irrelevant to the knowledge-production process, embedding exclusion before the system ever runs.

The specific form of epistemic injustice where prejudice causes a hearer to assign deflated credibility to a speaker's testimony. In AI contexts, this manifests when:

• User feedback from marginalized groups is systematically undervalued in model retraining
• Bug reports from non-technical users are dismissed as user error
• Domain experts from underrepresented regions are excluded from training data curation

Coined by philosopher Miranda Fricker in her 2007 work Epistemic Injustice: Power and the Ethics of Knowing.

A structural form of epistemic injustice where a gap in collective interpretive resources prevents someone from making sense of their social experience. In AI systems, this occurs when:

• Model taxonomies lack categories for experiences of marginalized groups
• Classification systems render certain harms illegible
• Automated content moderation fails to recognize culturally specific forms of harassment

This concept explains why some algorithmic harms remain systematically invisible to auditing frameworks.

The study and practice of designing machine learning systems that make decisions without unjustified discrimination against individuals or groups based on protected attributes. Key approaches include:

• Statistical parity: Equal positive prediction rates across groups
• Equalized odds: Equal true positive and false positive rates
• Counterfactual fairness: Decisions unchanged under counterfactual demographic shifts

While fairness metrics address distributive outcomes, epistemic injustice focuses on credibility and interpretive harms that fairness metrics alone cannot capture.

A structured transparency document that reports the intended use, evaluation results, and ethical considerations of a trained machine learning model. Model cards address epistemic injustice by:

• Disclosing disaggregated performance across demographic groups
• Documenting known limitations and failure modes
• Providing context about training data provenance

Standardized by Google researchers in 2019, model cards create an institutional record that validates the experiences of users who encounter model failures.

The ability to provide a negatively impacted individual with actionable, feasible steps to reverse an unfavorable automated decision. Recourse mechanisms combat epistemic injustice by:

• Treating affected individuals as knowers capable of self-advocacy
• Providing transparent explanations rather than opaque denials
• Enabling contestability of algorithmic decisions

Without recourse, automated systems commit a form of testimonial quieting, where the subject's voice is structurally excluded from the decision process.

A framework for evaluating algorithmic bias that examines how overlapping social identities—such as race and gender—combine to create unique, compounded experiences of discrimination. This directly addresses epistemic injustice by:

• Recognizing that single-axis analysis erases the experiences of multiply marginalized individuals
• Preventing hermeneutical marginalization where compound harms lack interpretive frameworks
• Requiring disaggregated evaluation beyond binary demographic categories

Pioneered by Kimberlé Crenshaw's legal scholarship and extended to AI by researchers like Joy Buolamwini in the Gender Shades project.