Argues that welfare evaluations need to specify which persona they measure, since a model hosts several persona-like regions (the Assistant, an ‘evil’ persona, ‘Aura’) that may express different preferences.
Author list not yet verified for v1. Title, year and link were checked.