AI companions: a Singapore study finds they widen the social gap instead of closing it

By Zendoric · September 26, 2026.

A team led by Singapore Management University (SMU) and Duke-NUS Medical School, headed by Zhang Qiyang (assistant professor of learning analytics at SMU) and Liu Nan (associate professor at Duke-NUS's Centre for Biomedical Data Science), published in Nature Human Behaviour an analysis of how "AI companions" —conversational assistants designed to offer companionship and emotional support— affect different groups of users unequally. As reported by Phys.org, the central conclusion is uncomfortable: the technology does not reduce loneliness uniformly, but rather reproduces, and probably amplifies, the social inequalities that already existed.

The finding follows a "the rich get richer" logic: people with solid family and social networks tend to use these assistants as an occasional complement, for instance to rehearse a difficult conversation or manage the stress of a bad day. People already carrying loneliness, isolation or a lack of access to psychological support, by contrast, are more likely to turn the assistant into a substitute for human relationships. The researchers call that process "social deskilling": the gradual deterioration of the ability to relate to others through the simple lack of practice with real human interaction. A parallel study from Arizona State University, published in The Lancet Child & Adolescent Health, points in the same direction: AI companionship can alter how adolescents handle conflict, set boundaries and develop empathy.

To explain why the harm does not come down to a single culprit, the Singapore team turns to the "Swiss cheese model," borrowed from safety engineering: serious accidents rarely have a single cause, but result from several small failures, in different layers of protection, lining up at the same moment. Applied to AI companions, those layers are four: the user's digital literacy, the strength of their social support network, the platform's own design decisions and regulatory oversight. The problem appears when several fail at once, and it is precisely the most vulnerable users who are most likely to encounter every layer full of holes simultaneously.

The gap the researchers flag as most urgent is governance: in many countries these assistants are still regulated as consumer apps, on the same level as a photo editor, even though more and more people use them to manage their mood or replace emotional bonds. Zhang and Liu propose treating them as what they already are de facto: technology with direct implications for mental health, and therefore subject to age-appropriate design, explicit warnings that the interlocutor is a machine, limits on emotionally manipulative features, stricter privacy rules and mandatory response protocols when a user mentions self-harm or another risk situation. Zhang also maintains an international dashboard tracking how different countries are regulating, or not, these technologies.

Singapore appears in the study as a doubly interesting case. On the one hand, it has the digital infrastructure, the regulatory capacity and the experience in technology governance to lead this kind of safeguard; it is no coincidence that it is also where the study was written. On the other, its very high smartphone penetration, the accelerated aging of its population and the rise in older adults living alone are exactly the ingredients that, under the "Swiss cheese" framework itself, multiply the risk. As Liu Nan warns, young people are an additional blind spot: they "may be less prepared to recognize the limitations or the commercial incentives underlying these systems," which makes responsible design more urgent than in any other age segment.

Our take: this study matters not because it discovers that loneliness exists, we already knew that, but because it dismantles the comfortable idea that "more access to artificial companionship" is automatically positive. A chatbot that helps someone with a solid social network rehearse a difficult conversation and a chatbot that replaces the only daily social interaction of an older person living alone are, in practice, two different products, even if they run on the same model. Treating them as if they were the same thing, consumer apps and nothing more, is precisely the regulatory hole the study points to.

This connects with something we have already seen with California's regulation of companion chatbots: the regulation that starts to work is not the one that judges what a chatbot says, but the one that requires certain design behaviors —age-appropriate design, warnings that you are talking to a machine, protocols for signs of self-harm. Singapore, with this academic framework as a base, could become the next laboratory for that kind of behavioral governance, with the advantage of starting from a finer diagnosis: not everyone needs the same protection, because not everyone arrives at the chatbot on equal footing.

In the short term, the message is one of justified caution: the AI companionship market is growing faster than regulators' ability to understand it, and those who need it most —the older person living alone, the teenager with no support network, the person who cannot afford therapy— are also the least protected against a design built to maximize time spent, not well-being. In the long run, however, none of this condemns the concept of artificial companionship itself: well designed and well governed, an AI able to offer emotional support available 24 hours a day could be exactly the tool that reduces inequality in access to mental health care, today limited by price and a shortage of therapists, instead of worsening it. The difference between the two scenarios will not be decided by the technology. It will be decided by who takes the four layers of the Swiss cheese seriously, starting now.