Determine whether governance vocabulary causally changes engineering decisions

Determine whether describing AI-agent behavior with psychological vocabulary, rather than operational descriptions such as output-distribution deviation under extended context windows, causally changes the engineering interventions selected for governance and safety.

Background

The paper uses the Bing Chat incident as a motivating illustration of how psychological descriptions such as confusion, feelings, and personality may make certain interventions more salient while obscuring infrastructure-level responses such as distribution-drift detection, output-layer filtering, and context-window bounds.

The authors explicitly limit their claim because they do not establish that alternative vocabulary would have produced different engineering decisions. The unresolved issue is therefore empirical and causal: whether the framing of an AI system actually changes the governance interventions selected, rather than merely changing how those interventions are described.