Everyday discrimination is not always recognized as discrimination. It often appears in ordinary exchanges: a comment about school background, region, family form, appearance, age, gender, disability, health, or occupation. Because these moments are embedded in context, they are difficult to study through direct questions about personal belief alone. They are also difficult to capture through experience-based reports, since people encounter different situations and may hesitate to name ambiguous harms as discriminatory. This paper therefore begins with a vignette survey designed to examine how people judge the same situated interactions.

We analyze survey responses from 2,270 Korean adults who evaluated 30 Korean vignettes describing ordinary but potentially discriminatory or exclusionary situations. For each vignette, respondents were not asked to decide whether the situation was objectively discriminatory. Instead, they inferred how the target of the interaction would feel on a six-point scale. The study treats the perception of discrimination as a socially situated judgment about context, harm, and likely reception.

We compare these human response distributions with an initial pilot of LLM-generated judgments from GPT-5.5 Pro, Claude Opus 4.8, and Gemini 3.5 Thinking on the same vignettes. The pilot suggests that LLMs do not simply fail to recognize everyday discrimination. Rather, they tend to recognize it in a more uniformly negative and normatively stabilized way. Across the 30 items, the average of the human item means was 4.09 on the six-point scale, while model means were higher: 4.93 for GPT-5.5 Pro, 4.50 for Claude Opus 4.8, and 4.53 for Gemini 3.5 Thinking. The models rarely used the lower end of the scale, and GPT-5.5 Pro did not assign any item below 4. These preliminary results suggest a pattern of ambiguity compression: LLMs may translate socially variable judgments into more settled moral evaluations.

This compression has sociological consequences. In everyday life, disagreement about whether a comment is hurtful or discriminatory is not merely measurement noise; it is part of how discrimination is recognized, minimized, contested, or normalized. If LLMs smooth this variation, they may overstate social consensus, obscure differences across social groups, and turn ambiguous experiences into more authoritative model judgments. This matters because LLMs are increasingly used as conversational partners for workplace conflict, family tension, school experiences, relationship problems, and vague feelings of discomfort. By placing LLM judgments against large-scale survey data from South Korea, this paper asks what is gained, and what is lost, when culturally specific perceptions of everyday discrimination are translated into model-generated judgments.