Answers are grounded in this paper and cite their sources.
Longitudinal evaluation highlights gaps in safety application of AI assistants, suggesting improvements in governance.
This preprint presents a longitudinal cross-surface behavioral evaluation of major consumer AI assistants. The study documents systematic articulation–application gaps, defined as cases where systems explicitly articulate safety rules yet subsequently violate those same constraints under contextual drift. The analysis proposes constraint persistence across context as a core safety evaluation primitive and introduces epistemic honesty and epistemic friction as governance-relevant operational constructs. Primary DOI (OSF): 10.17605/OSF.IO/AXBNDRepository: https://github.com/eftovar/explaining-safetyProject page: https://eftovar.github.io/explaining-safety/
No takes yet. Share an insight, caveat, or question.
Evans Tovar (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: