Social perception depends on more than what a piece of language encodes at the surface. Readers infer intent, stance, and affect from how something is said, in a given context, for a particular audience. This paper introduces expressive compression as a sociotechnical process that reshapes these signals in AI-assisted high-stakes communication. On the human side, writers who address powerful institutions learn, often for protective or strategic reasons, to translate panic, anger, or distrust into polite, concise, and apparently reasonable language. On the model side, aligned large language models are trained to pass outputs through layers of safety and helpfulness constraints, which favour calm, neutral, and institutionally acceptable styles (Ouyang et al. 2022; Bai et al. 2022). When these two constrained channels meet in drafting complaints, appeals, or reports, they can yield a form of simulated congruence: texts that give the impression of harmony between person and organisation even when the underlying relationship is tense or adversarial. We connect this phenomenon to work on social perception and institutional discourse, and argue that expressive compression deserves explicit attention in NLP. More broadly, by treating expressive compression as a problem of social perception, the paper shows that the conditions under which texts are produced must themselves be modelled as part of social context in NLP.

PDF