Language Models, Experts, and Epistemic Trust

Recent philosophical debates have often denied epistemic trust to current language models on three grounds. These models lack the understanding, epistemic virtues, communicative intentions, and responsibility required of good testimonial sources. They are opaque. Their outputs are unreliable because of hallucinations, harmful bias, and sycophancy. This paper argues that these objections primarily target current general-purpose chatbots and do not show that all language models are unworthy of epistemic trust. It identifies a developing trend that I call the expertization of models. Expertized models are restricted to specific domains, rely on training and retrieval materials selected by human experts, are post-trained for truth rather than user satisfaction, and are embedded in emerging mechanisms through which developers can bear partial responsibility for outputs. The paper first reconstructs the main arguments against epistemic trust in language models. It then shows how expert models can answer these arguments through developer proxy, data and source transparency, domain limitation, truth-directed training, and institutional responsibility. Existing expert models remain inadequate because their foundations are often opaque, responsibility mechanisms are weak, competence boundaries are insufficiently marked, and accuracy remains insufficient in some domains. These shortcomings, however, are not principled barriers. If expert models become sufficiently reliable, traceable, and accountable within bounded domains, there is no categorical reason to deny them epistemic trust.