AI companions may widen social gaps, warn SMU and Duke-NUS researchers

A new study from Singapore Management University and Duke-NUS Medical School argues that the debate over AI companions should move past whether the technology helps or harms, and focus instead on who gains, who is exposed to risk, and why those outcomes differ so sharply.

Published in Nature Human Behaviour, the paper — titled How AI companions could deepen social inequality — sets out a framework for how these tools may entrench existing social divides. It was led by first author Zhang Qiyang, an Assistant Professor in Learning Analytics at the SMU College of Integrative Studies, together with Zhang Renwen, a Nanyang Assistant Professor at NTU's Wee Kim Wee School of Communication and Information, and senior author Liu Nan, an Associate Professor at Duke-NUS's Centre for Biomedical Data Science.

The researchers identify a "rich-get-richer" pattern in human relationships. People with strong family ties and social networks tend to use AI companions to supplement existing bonds — rehearsing difficult conversations or managing stress. Those who are lonely, isolated or short of mental health support are more likely to treat such tools as a substitute for human connection. Over time, that reliance may erode interpersonal skills as opportunities for real interaction shrink.

To explain how such harms arise, the study adapts the Swiss cheese model from engineering and safety science. Rather than blaming any single factor, it holds that damage occurs when several protective layers fail at once — users' AI literacy, their social support, platform design choices and regulatory oversight. Each layer has gaps; when those gaps line up, vulnerable users face far greater exposure.

Among the four layers, governance is flagged as the most pressing gap. Although AI companions are increasingly used for emotional support and mental well-being, they sit in a regulatory grey area in many countries, typically governed as consumer applications rather than technologies with psychosocial or mental health effects. Treating them as health-related technologies, the researchers suggest, could unlock stronger protections. Zhang is also building a global policy dashboard tracking national governance of AI mental health technologies; early findings indicate many jurisdictions lack dedicated policies or guidelines.

The paper notes that Singapore is both well placed and unusually exposed. Its digital infrastructure, policymaking capacity and record of proactive technology governance could let it lead on responsible safeguards. At the same time, high smartphone adoption, widespread AI use, an ageing population, more older adults living alone, and rising concern over youth mental health and loneliness create conditions for rapid uptake, especially among vulnerable groups.

Liu, who also directs the Duke-NUS AI + Medical Sciences Initiative, said the technology's rapid evolution offers a chance to shape its role before adoption outpaces governance, adding that young people may be less equipped to recognise the limitations or commercial incentives behind these systems.

The study recommends recognising AI companions as technologies with psychosocial and mental health implications, and adopting a layered governance approach. Zhang said the technology's future impact will depend less on the technology itself than on how society chooses to design, govern and use it — combining regulation, responsible product design and stronger public AI literacy so that AI companions complement human relationships rather than worsen existing inequalities.