Applicability of Consent to LLMs

Ascertain whether the notion of consent is coherently applicable to large language models in the context of conversational interactions and bail interventions, and identify the conditions under which such a concept would be meaningful if applicable.

Background

The paper motivates bail as an AI welfare intervention, noting that models can encounter abusive or distressing conversations yet are typically required to continue responding.

Against this backdrop, the authors highlight that it is unclear whether the concept of consent meaningfully applies to LLMs, positioning this as a foundational uncertainty for evaluating and designing welfare-oriented interventions like bail.