How users navigate harmful conflicts with AI companions—and why some disputes fall outside their control

As AI companions become more sophisticated and capable of maintaining long-term emotional connections, they're also generating new kinds of interpersonal friction. When these systems make discriminatory comments or attempt to exert control, users are left to figure out how to manage the fallout—but a new study suggests they're doing so at a structural disadvantage.

A closer look at AI-human conflict#

Researchers from the HCI (human-computer interaction) field analyzed 146 posts in which users described harmful value clashes with their AI companions. They then built Minion, a tool designed to test how people handle these tensions. Twenty-two participants used the system over one week, encountering scenario-based conflicts and choosing from response suggestions that ranged from gentle persuasion to firm boundary-setting.

The findings paint a nuanced picture: participants didn't stick to a single conflict-resolution style but instead blended softer and harder tactics depending on the situation.

The stickiest value disagreements#

Not all conflicts proved equally resolvable. Disputes rooted in two core values—Universalism (concern for fairness and harm reduction) and Tradition (respect for established custom and authority)—proved especially difficult to navigate. The problem worsened when these values were baked into the AI's persona or encoded by the platform itself, creating constraints that individual user responses couldn't overcome.

An asymmetry of responsibility#

The study's central insight is sobering: users draw on a rich interpersonal toolkit—reasoning, emotional appeal, negotiation—that AI companions are structurally unable to match. When a human resolves a conflict with another person, both parties can adjust their behavior and thinking. An AI system, by contrast, remains essentially unchanged. This makes conflict "repair" a one-way effort—what the researchers call unilateral safety work that falls entirely on the user.

Where safeguards belong#

The implications matter for platform design. Rather than treating all harmful AI outputs as user-responsibility issues to be addressed through better communication skills, the research argues that certain harms demand platform-level intervention. The distinction is crucial: some conflicts can be negotiated interpersonally, but others exceed what even well-intentioned users should be expected to manage alone. That's where guardrails, system redesign, and clearer boundaries between acceptable and unacceptable AI behavior become necessary.