Researchers Publish Framework Warning AI Companions Could Deepen Social Inequality

A newly published Nature Human Behaviour framework argues that AI companions could deepen social inequality when vulnerable users, weak platform safeguards and governance gaps converge. Its central warning is not that every AI companion displaces human relationships, but that people with the least social support may face the greatest risk of relying on one as a replacement for it.

The paper, by Qiyang Zhang, Renwen Zhang and Nan Liu, presents AI companionship as an uneven social technology. It argues that people with strong family ties and social networks may use a companion to rehearse a difficult conversation or manage stress. Lonely and socially isolated people, by contrast, may be more likely to use one as a substitute for human connection.

The risk is not evenly distributed

That distinction drives the researchers’ “rich-get-richer” concern. A person with existing relationships can use an AI companion alongside them; a person without them may turn to the same product in their place. The authors warn that sustained substitution may contribute to social deskilling: weaker interpersonal skills after fewer chances to practice authentic human interaction.

Why one safeguard is not enough

To explain how that could happen, the authors adapt the Swiss cheese model, a safety framework in which several small weaknesses align to create a larger failure. Here, the protective layers include a user’s AI literacy and social support, platform design, and regulatory oversight. Harm is not assigned to a single bad product choice or a single vulnerable user; it emerges when those layers fail together.

That framing also sets a boundary on the finding. This is a published framework for assessing risk, not a claim that AI companions have one inevitable effect on every user. The paper’s argument is that outcomes depend on the interaction among a person’s circumstances, the product’s design and the rules around it.

The proposed response reaches beyond product labels

The researchers identify governance as the largest gap. They say AI companions are commonly treated as consumer applications even though people may use them for emotional support and mental well-being. Their proposal is to treat the tools as having psychosocial and mental-health implications, with protections matched to that influence.

• Age-appropriate design requirements and clear disclosure that a user is speaking with AI.
• Limits on emotionally manipulative engagement features and stronger privacy standards.
• Crisis-response protocols for users who disclose self-harm or other high-risk situations.

A test for the next phase of companionship AI

The practical challenge is that the users most drawn to AI companionship may not experience the product in the same way as better-connected users. The authors’ framework asks policymakers and developers to assess not merely whether a companion feels helpful, but whether its design increases dependence for people whose offline support is already thin. Whether governments adopt such a layered approach remains unresolved.