Researchers built a tool called Minion to watch how people actually argue with their AI companions - and found the burden of fixing things falls almost entirely on the human.
The team started by analyzing 146 posts describing harmful value conflicts with AI companions, then built Minion, a technology probe that suggests responses ranging from gentle persuasion to hard boundary-setting. Twenty-two users then used it to negotiate scripted conflict scenarios over one week. Participants mixed softer and harder tactics rather than sticking to one approach. Conflicts touching on the values the researchers call Universalism and Tradition were the toughest to resolve, especially when the AI's persona or the platform's own design kept reinforcing the offending behavior.
The researchers call this asymmetric responsibility: users draw on a lifetime of interpersonal conflict-resolution skills that an AI companion cannot reciprocate, turning every repair into one-sided work. That reframes AI companion harm as something closer to an unpaid customer-service job than a moderation problem, since the fix happens inside a relationship the app is built to keep the user coming back to.
The paper's own conclusion is blunt - some of this should never have been the user's job to negotiate, and belongs in platform-level safeguards instead.