This paper proposes a novel measure for operationalizing authoritarianism: the narrowing of semantic dispersion. This paper defines semantic dispersion as the directional dispersion of a set of document’s semantic embedding vectors as obtained with a language model. Compared to democracy and press freedom indices, which are infrequently updated and rely on subjective evaluations by experts, semantic dispersion can be directly measured and is, therefore, immediately available for any time range (data permitting) and is not biased by direct human judgment (having been based on a language model’s self-supervised training). This paper argues that semantic dispersion is a promising metric for understanding authoritarianism, and illustrates this point by analyzing the semantic dispersion of GDELT internet news article embedding data from Russia and Myanmar from 2020-2022.