Could AI Companions Deepen Social Inequality? The Next AI Debate Is About Vulnerability

AI companionship is often discussed as if every person encounters the technology under equal conditions.
They do not.
A new article in Nature Human Behaviour, published on 21 September, argues that AI companions could deepen existing social inequalities when user vulnerabilities combine with platform-control weaknesses and gaps in governance.
That framing introduces an important dimension into the AI-companion debate.
An AI companion may offer conversation, emotional support, information, and a sense of availability at very low marginal cost.
For someone with a strong offline support network, the AI may function primarily as an additional tool.
For someone who is lonely, socially isolated, financially constrained, or experiencing other vulnerabilities, the same system may occupy a much larger psychological role.
That difference matters.
Platform decisions — subscription pricing, model changes, memory limits, advertising, recommendation systems, and access restrictions — may therefore have very different consequences depending on the person using the technology.
This is where psychology and technology policy begin to collide.
The important question is not simply: Are AI companions good or bad?
A more useful question is: For whom, under what circumstances, with what protections, and who controls the relationship environment?
Researchers examining AI companionship increasingly point toward a complicated picture.
Separate research involving 1,131 Character.AI users has examined associations between companion use, relationship experiences and psychological wellbeing using surveys and hundreds of thousands of messages.
The emerging field therefore needs to resist simplistic conclusions.
AI companionship may create genuine benefits for some people while creating dependencies or vulnerabilities for others.
That is normal for psychologically powerful technologies.
The responsibility now is to understand the conditions under which different outcomes emerge.
The future of socially intelligent AI will not be determined only by how convincingly machines can communicate.
It will also be shaped by whether those systems are designed around the psychological realities of the humans communicating with them.