A taxonomy of epistemic injustice in the context of AI and the case for generative hermeneutical erasure

AI systems can perpetuate multiple forms of epistemic injustice — the paper identifies diverse sources including opacity in machine learning models, discriminatory automation of biases, AI hallucinations that distort human beliefs, exclusion of the Global South from AI governance decisions, and algorithmic systems that enable bureaucratic violence.

Large language models may erase non-Western ways of knowing — the paper introduces "generative hermeneutical erasure" as a new concern, arguing that LLMs' "view from nowhere" approach systematically undervalues and suppresses non-Western epistemologies and conceptual frameworks, gradually eroding these alternative knowledge systems.

AI opacity creates accountability gaps — epistemic opacity in machine learning models is identified as a source of injustice, meaning people cannot understand how AI systems reach conclusions, making it difficult to identify and challenge unfair outcomes.

AI governance decisions exclude crucial voices — the exclusion of the Global South from global AI governance represents an epistemic injustice, meaning decisions about AI development and deployment are made without input from affected populations with different knowledge and perspectives.

A taxonomy can help identify and address AI-related injustices — the paper proposes a framework for mapping different types of epistemic injustice in AI contexts, which could help policymakers, technologists, and advocates recognize and prevent these harms systematically.