Cognitive Compression and Infrahumanization: A Unified Account of Dehumanization in Human Cognition and Artificial Content Moderation

Two literatures, developed independently, appear to describe the same phenomenon. Social psychology documents infrahumanization — a default asymmetry in the emotional interiority attributed to familiar and unfamiliar others — along with the mechanisms of moral disengagement and institutional authorization by which that asymmetry escapes correction and, in extreme cases, enables organized violence. Machine learning research documents systematic errors in automated content moderation, including measured racial bias in hate speech detection arising from the loss of dialectal context rather than from deliberate design. This paper argues that both are instances of a single mechanism, cognitive compression, grounded formally in Herbert Simon's theory of bounded rationality: the predictable output of any resource-constrained system facing a categorization problem whose complexity exceeds the processing budget available to solve it exactly. A formal statement of this mechanism is shown to entail, independently of substrate, four empirical signatures documented separately in the psychological and technical literatures — compression that tracks processing load rather than moral stakes, reliance on cheap proxy features, correction only through costly reintroduction of individuating context, and convergence on a centralize-then-distribute architecture exemplified both by institutional authorization of violence and by the training-then-inference structure of a content classifier. The paper argues that the appropriate response to this mechanism is not to mandate its elimination, which would recreate, in a new domain, the same structural instability identified in a companion paper on AI objective functions, but to treat detection of compression as a process goal: surfacing compressed judgments as compressed, and leaving the choice of whether to pay the cost of decompression to the party affected.