Epistemic Injustice in Generative AI: When Algorithms Become Gatekeepers of Knowledge

This article explores how large language models (LLMs) cause systemic epistemic injustice through probabilistic generation mechanisms, including testimonial injustice and hermeneutical injustice. It analyzes their inherent structural mechanisms (such as capacity erosion and credibility inflation) and their real-world impacts in high-risk fields like healthcare and law. It calls for multi-stakeholder collaboration among technology, ethics, and policy sectors to safeguard epistemic justice in the algorithmic age.

Epistemic injustice was proposed by philosopher Miranda Fricker. It refers to the systemic deprivation of credibility or understanding ability of certain groups due to their identity in knowledge transmission, divided into testimonial injustice (dismissing statements due to bias) and hermeneutical injustice (lack of shared concepts to express experiences). In the AI field, researchers have proposed the 'AI-mediated Testimonial Injustice (AITI)' framework, which describes how LLMs, as knowledge intermediaries, amplify social biases and create new injustices.

LLMs are essentially probabilistic machines that generate 'most likely' content by learning statistical correlations in text. However, this leads to four mechanisms of epistemic injustice:

1. Capacity Erosion: Over-reliance on AI weakens humans' knowledge acquisition and critical thinking abilities;
2. Credibility Inflation: Fluent text creates false authority, leading to excessive user trust;
3. Exacerbated Marginalization: Scarce voices of minority groups in training data further compress their communication space by the model;
4. Diffused Responsibility: AI intervention obscures responsibility attribution, making it difficult to hold accountable for injustices.

AI epistemic injustice is embodied in fields like healthcare and law:

• Healthcare: If AI-assisted diagnostic systems are trained primarily on data from mainstream groups, they may provide misleading information about rare disease symptoms in minority groups, delaying diagnosis;
• Law: When lawyers use LLMs to search for precedents, the model's biased understanding of minority group cases affects the fairness of legal arguments, exacerbating the double marginalization of marginalized groups.

The complexity of LLMs makes their decision-making processes difficult to fully explain, forming a 'transparency paradox'—requiring an unexplainable system to explain itself. However, researchers call for a new governance framework: integrating epistemic justice into the core of AI design, including proactively incorporating diverse training data, developing bias detection tools, establishing human supervision mechanisms, and fostering user media literacy.

Generative AI was supposed to promote knowledge democratization, but epistemic injustice is a product of the interaction between its probabilistic nature and social inequality. Addressing this requires collaboration across technology (optimizing data and models), ethics (considering epistemic justice), and policy (clarifying responsibilities). We need to rethink the definition of knowledge and the right to truth in the algorithmic age to ensure that technology serves epistemic liberation rather than oppression.