Inspired by the East Asian notions of “face”, Erving Goffman conceptualized face theory as the management of one’s social self-image (“face”) during interactions goffmanFaceWorkAnalysisRitual1955. Formally, “face” refers to the positive social value a person claims for themselves through their behavior and how others perceive it. People engage in face-work—verbal and nonverbal actions—to maintain their own face and support others’ faces, avoiding embarrassment or conflict.
Brown and Levinson expanded on the idea of face in their politeness theory, distinguishing between positive face—the desire to be liked, approved of, and included, and negative face—the desire for freedom of action and to not be imposed upon. Using this framework, they defined politeness as the management of face through linguistic choices calibrated to protect both parties’ positive and negative face wants brownPolitenessUniversalsLanguage1987.
In studying refusal through the lens of politeness theory, Johnson et al. found that refusal can induce threat to both positive and negative face johnsonPolitenessTheoryRefusals2004. More specifically, they found refusal to threaten:
- Requester’s negative face: Refusals inherently threaten this the most. The requester’s autonomy is constrained—they must either persist, give up, or find another way to meet their goal.
- Refuser’s negative face: Less threatened overall, but increases when the refuser has the ability to comply yet focuses the refusal on the requester (e.g., “You should handle it yourself”), which limits future relational autonomy.
- Requester’s positive face: Threatened when the refusal conveys both low ability and low willingness—implying the requester made a poor judgment in asking.
- Refuser’s positive face: Threatened when unwillingness is expressed, especially if the refusal centers on the requester. Showing willingness or focusing the refusal away from the requester (e.g., citing external obstacles) protects the refuser’s image of competence and goodwill.
In AI alignment research, in particular those investigating how to train large language models (LLMs) to refuse queries jiAIAlignmentComprehensive2025, refusal is often viewed as a binary mechanism of safety—refusal of queries that are inappropriate, and non-refusal to those that are not. However, very little research has explored how different deliveries of refusal affect human-machine dialogue. Situating this line of research in the theory of face, I aim to understand whether the known threats to the requester’s negative and positive face hold in the context of human-LLM interaction.