Affective AI Safety: The Missing Piece in LLM Safety

Emotional harms are an overlooked safety gap: While AI safety research has focused on misinformation, bias, and reliability, the paper identifies that risks to human emotional wellbeing—termed "affective safety"—have been fragmented and undertheorized, representing a significant blind spot in current safety frameworks.

Three recurring types of emotional harms: The research identifies three main categories of affective harms: (1) affective self-alienation (disconnection from one's own emotions), (2) fairness and bias harms (discriminatory treatment based on emotional characteristics), and (3) relational harms (damage to human relationships and social bonds), which appear consistently across different AI systems.

Current safety measures are inadequate: Existing AI safety frameworks address emotional harms either narrowly or not at all, meaning that standard safety practices and regulations don't adequately protect people from emotional manipulation, alienation, or relational damage caused by AI systems.

Emotional harms have cumulative effects: The paper emphasizes that affective safety requires frameworks accounting for cumulative, relational, and identity-level effects—meaning that repeated emotional interactions with AI systems can compound over time and affect how people understand themselves and relate to others.

New technical and regulatory approaches are needed: Addressing affective safety requires dedicated frameworks beyond current practices, with specific technical solutions and regulatory approaches designed to protect emotional wellbeing rather than applying generic AI safety measures.