Large language models (LLMs) are often described as mirrors of human language and culture. This paper argues that what they reflect is not neutral linguistic content, but socially structured forms of ignorance. I develop this claim through the concept of automated ignorance. AI-generated language appears meaningful and informative despite being produced without understanding, intention, or participation in social practices (Bender and Koller 2020; Pavlick 2023). This apparent meaningfulness is best explained by externalist theories of meaning, according to which linguistic content depends on relations to shared social practices and the external world rather than solely on speaker intention (Putnam 1975; Burge 1979; Kripke 1980). Because LLMs are trained on human-generated text, their outputs can be interpreted within these existing linguistic frameworks. However, they do not themselves participate in the causal–historical and epistemic practices that sustain meaning.
This limitation has important consequences. Human linguistic practices are shaped by historically structured forms of epistemic injustice, including conceptual gaps and systematic distortions of meaning (Fricker 2007; Dotson 2014; Medina 2013). When these patterns are encoded in AI systems, they are not merely reproduced but transformed. I introduce the concept of automated ignorance to describe this transformation: the algorithmic reproduction and stabilisation of socially structured patterns of not-knowing, whereby epistemic distortions are generated, scaled, and presented as neutral outputs. In this sense, LLMs do not simply reflect human discourse but reorganise the epistemic conditions under which meaning becomes intelligible.
To illustrate this, I analyse cases of AI-generated gendered hate speech. Such outputs arise not from intention but from interactional dynamics between users and systems, where prompts and responses jointly reproduce socially available interpretive frameworks. Even under moderation, users can iteratively reframe prompts to elicit outputs that reinforce stereotypes or distort social reality. LLMs therefore do not merely mirror human discourse but transform the epistemic conditions under which meaning becomes intelligible, stabilising and amplifying existing structures of ignorance. Automated ignorance names this shift: the transformation of socially situated ignorance into an infrastructural feature of digital media and knowledge production.
Author: Shaoyu Han