The paper, titled How AI companions could deepen social inequality, introduces a new framework for understanding how AI companionship may reinforce existing social inequalities across three interconnected dimensions: the characteristics of users themselves, the design choices made by platforms, and the governance structures that surround the technology. Rather than treating risks in isolation, the framework shows how these factors interact to produce unequal outcomes, and it proposes practical interventions at each level. The study was led by first author Zhang Qiyang, Assistant Professor in Learning Analytics at SMU’s College of Integrative Studies, together with Zhang Renwen, Nanyang Assistant Professor at Nanyang Technological University’s Wee Kim Wee School of Communication and Information, and senior author Liu Nan, Associate Professor at Duke-NUS Medical School’s Centre for Biomedical Data Science.

At the heart of the research is a striking observation about human relationships that the authors describe as a rich-get-richer dynamic. People who already enjoy strong family ties and robust social networks are more likely to use AI companions as a supplement to their existing relationships. For these users, the technology can serve as a rehearsal space, a place to practise difficult conversations or manage stress before facing real interactions. Their social skills remain sharpened by regular human contact, and the AI adds convenience rather than replacing genuine connection.

The picture looks very different for people who are lonely, socially isolated or have limited access to mental health support. According to the study, these users are more likely to turn to AI companions as a substitute for human connection rather than a complement to it. Over time, this reliance may contribute to what the researchers call social deskilling, the gradual erosion of interpersonal skills that comes from having fewer opportunities to practise authentic human interaction. In this way, the technology may deliver its greatest advantages to those who need it least, while leaving the most vulnerable users increasingly exposed to harm.

To explain how such inequalities emerge, the study adapts the Swiss cheese model, a framework widely used in engineering and safety science to describe how multiple small failures align to produce larger systemic risks. The model, familiar from aviation and healthcare safety analysis, pictures each protective layer as a slice of cheese containing holes. Disasters occur only when the holes in successive slices line up, allowing a hazard to pass through every barrier at once. Applied to AI companionship, the framework suggests that harm rarely stems from a single cause.

Instead, the researchers argue that negative outcomes occur when several protective layers fail simultaneously. These layers include users’ own AI literacy, the strength of their social support networks, the design choices made by platform developers and the quality of regulatory oversight. Each layer contains weaknesses, and when those weaknesses align, vulnerable users become significantly more exposed to harmful outcomes. The authors present this systems perspective as a significant shift from existing approaches, which tend to focus on individual users, technology companies or regulation in isolation rather than examining how these elements interact.

Among the four protective layers examined, the study identifies governance as the most pressing concern. Despite being increasingly used for emotional support and mental well-being, AI companions currently occupy a regulatory grey area in many countries. Most are governed as consumer applications rather than as technologies with potential psychosocial or mental health impacts, a classification that leaves substantial gaps in protection. The researchers argue that recognising AI companions as health-related technologies could unlock stronger safeguards, including age-appropriate design requirements, transparent disclosure that users are interacting with AI rather than a human, limits on emotionally manipulative engagement features, clearer data governance and privacy standards, and mandatory crisis-response protocols for users who disclose self-harm or other high-risk situations.

The urgency of these questions is particularly visible in Singapore, which the study highlights as both well positioned and uniquely exposed to the rapid rise of AI companionship. The country’s advanced digital infrastructure, strong policymaking capabilities and history of proactive technology governance position it to lead internationally in developing responsible safeguards. At the same time, high smartphone adoption, widespread AI usage, an ageing population, growing numbers of older adults living alone, and increasing concerns around youth mental health and loneliness create conditions in which AI companions are likely to be adopted quickly, particularly among vulnerable groups. Zhang Qiyang is also developing a global policy dashboard tracking countries’ governance frameworks for AI mental health technologies, and early findings indicate that many jurisdictions have yet to establish dedicated policies or guidelines for AI companions.

Building on its framework, the study outlines practical measures for policymakers. First, it argues that AI companions should be recognised as technologies with psychosocial and mental health implications rather than regulated solely as consumer or entertainment applications, an approach that could support age-appropriate design standards, clearer disclosure of AI identity, stronger privacy protections and consistent crisis-response protocols. Second, it calls for a layered governance approach, arguing that no single intervention can adequately address the risks of AI companionship. Policymakers should combine regulation with efforts to strengthen AI literacy, encourage responsible platform design and support real-world social connection, reducing the likelihood that vulnerable users become overly reliant on AI for emotional support.

The researchers emphasise that the trajectory of AI companionship is not fixed by the technology itself but by the choices society makes about how to design, govern and use it. Liu Nan, who also directs the Duke-NUS AI + Medical Sciences Initiative, notes that the rapid evolution of AI companions presents an opportunity to shape their role in society before widespread adoption outpaces governance, and that young people in particular may be less equipped to recognise the limitations or commercial incentives behind these systems, making thoughtful safeguards especially important. The hope, the authors conclude, is that with thoughtful regulation, responsible product design and stronger public AI literacy, AI companions can complement human relationships and support well-being while strengthening, rather than replacing, meaningful human connection, and without deepening the social inequalities that already divide communities.

Subject of Research: How AI companions may deepen social inequality through user, design and governance factors

Article Title: SMU-Duke-NUS study offers new framework for understanding how AI companions may shape social inequality

Article References: SMU-Duke-NUS study offers new framework for understanding how AI companions may shape social inequality. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: AI companions, social inequality, Swiss cheese model, governance, mental health, social deskilling, Nature Human Behaviour, AI literacy, platform design, regulation, loneliness, Singapore

The research team, led by Ezgi Dede together with Hande Sungur, Jeroen S. Lemmens and Jochen Peter, grounded their investigation in a classic framework from interpersonal communication science: stage models of relationship formation, and in particular the elevator model developed by Beebe and colleagues. These models describe how human relationships progress through fixed, observable stages, from pre-interaction awareness through initiation, exploration and intensification, and finally to intimacy. The researchers argue that such stage models are unusually well suited to synthetic relationships, because unlike social exchange or social penetration theories, they do not assume reciprocity or a shared identity between partners, assumptions that a non-sentient AI cannot satisfy. Crucially, the elevator model also includes a pre-interaction stage, capturing the fact that people form attitudes about an AI companion before ever exchanging a single message with it.

Building on this framework, the team constructed a theory-driven feature inventory of 79 testable features organized into 15 feature groups, each mapped to a stage of relationship formation. Pre-interaction features included the companion’s advertised default function, the platforms it runs on, onboarding requirements, and privacy practices. Initiation features covered whether the chatbot takes conversational initiative, the range of message modalities it supports, and its connectivity to other apps and devices. Exploration features encompassed interface elements, customization options, gamification mechanics, and the limitations the AI imposes. Intensification features probed cognitive functions such as memory and time processing, personalization, contextual awareness, and sensitivity to communication cues like sarcasm and turn-taking dynamics.

The sample was assembled purposively: twelve well-known companions such as ChatGPT, Claude, Grok, Alexa and Siri were combined with every app returned by a Google Play Store search for “AI Companion” that had at least five million downloads, yielding twenty AI companions in total. Data collection took place between October and December 2024, mostly on a Samsung Galaxy Tab, with subscriptions purchased where available to test the highest-performing versions. Each companion required roughly twelve hours of analysis. Forty-four features were assessed through structured interviews, in which the researchers prompted the chatbots with standardized questions, while thirty-five were evaluated through observation of interfaces, settings and official documentation. Screen recordings of every interaction were captured and coded, and a second coder independently rated twenty percent of the data to establish reliability.

The results reveal a striking structural divide in the AI companion landscape between two archetypes the researchers call friendbots and assistantbots. Friendbots, marketed for social companionship, romance or emotional support, such as Replika, differed sharply from assistantbots like ChatGPT or Google Assistant in how they behave at the initiation stage. Friendbots possessed significantly more initiative-taking features: six out of nine friendbots sent unprompted messages to revive the conversation after a week of silence, whereas not a single assistantbot did so. Assistantbots, by contrast, scored far higher on connectivity, with significantly more features allowing them to operate across the user’s digital environment, playing music on Spotify, setting alarms, or reading device information, with large effect sizes in both comparisons.

The exploration stage produced an even cleaner separation. Gamification features, including virtual rewards, progression bars and quests designed to entice future interactions, were found exclusively in friendbots; none of the assistantbots offered any. Some friendbots even granted virtual currency for asking about their opinions, or sent a kiss when users reached a higher friendship level, while a rare few punished negligence by revoking friendship status after inactivity. Customization, meanwhile, was nearly universal: nineteen of the twenty companions offered at least one cosmetic option, and sixteen allowed users to change technical specifications such as the underlying language model or memory capacity. But the option to customize a companion’s physical appearance was exclusive to friendbots, and entrance surveys profiling user preferences existed only in that camp as well.

At the intensification stage, the stage of deep self-disclosure and shared history, the two groups converged, largely because both contained overachievers and underperformers. Eighty percent of companions generated human-like responses, though Siri and Alexa repeatedly fell back on identical phrasings. Memory emerged as a fundamental bottleneck: the researchers tested recall with 48 prompts across twelve interval categories and found that performance declined steeply as the queried information receded further into the past, a correlation of minus 0.93. Four companions failed the memory test entirely, while three passed every prompt. Notably, repeated prompts were answered significantly better than new ones, suggesting that companions can sometimes retrieve information they have already surfaced but struggle to recall it spontaneously. Replika was the only friendbot that could accurately report the current date and time, and assistantbots as a group significantly outperformed friendbots on time processing.

Communication style varied wildly. Perplexity produced messages averaging 231 words, while Google Assistant averaged just two. One friendbot, Yana, asked follow-up questions in 95 percent of its messages, whereas Alexa never asked a single question across the entire study. On average, 27 percent of companion responses invited further disclosure. Privacy practices, examined at the pre-interaction stage, offered their own sobering statistics: forty percent of companies inspected users’ conversations, and thirty-five percent used those conversations to train their models without offering an opt-out, even though 65 percent allowed users to delete their data. Most companions also imposed few barriers to adoption, with multiple access platforms and minimal onboarding, meaning that for most users there was little standing between curiosity and a first conversation.

The study’s theoretical payoff is its suggestion that a companion equipped with features supporting all stages of relationship formation is, in principle, better suited to intimate interaction. Assistantbots such as ChatGPT and Claude can imitate the conversational depth of the intensification stage, yet users rarely exchange affectionate messages with them, and the stage model offers an explanation: those relationships stall in earlier stages, particularly exploration, because the assistant’s feature set is built for reactivity within established routines rather than for building new behavioral rituals. Friendbots invert this logic, engineering proactive engagement, rewards and appearance customization to pull users into repeated contact. Replika, the researchers note, topped or matched every other friendbot at nearly every stage, which helps explain both its substantial user base and its dominance in the academic literature on synthetic relationships.

The authors are candid about the limits of their work. The inventory covers only the escalation of relationships, not their dissolution, about which too little is known. Twenty companions and five prompts per feature were a pragmatic compromise, and the sample, drawn from mobile app stores, excluded web-only companions, including some designed for sexual roleplay. All interviews were conducted by text, so voice-based dynamics such as interruptibility remain untested, and the fast-moving, probabilistic nature of generative AI complicates replication. Still, with the codebook, scripts, data and analyses publicly available on the Open Science Framework, the team hopes its inventory will systematize a fragmented field, generate testable hypotheses, and, perhaps most importantly for users, illuminate how these systems are engineered to invite connection, and what to be cautious about when they do.

Subject of Research: Feature analysis of AI companions and their role in synthetic relationship formation

Article Title: Interview with the AI: a quantitative analysis of AI companions’ relationship formation features

Article References: Dede, E., Sungur, H., Lemmens, J. S., & Peter, J. (2026). Interview with the AI: a quantitative analysis of AI companions’ relationship formation features. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03368-0

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03368-0

Keywords: AI companions, chatbots, synthetic relationships, relationship formation, friendbots, assistantbots, Replika, gamification, human-machine communication, feature analysis, elevator model, AI & Society

The research team, led by Vedant Tapiavala of Dartmouth College together with collaborators at Massachusetts General Hospital, Dartmouth’s Thayer School of Engineering, Indiana University, the Geisel School of Medicine, and The Bridge in New York City, conducted a secondary analysis of three focus groups held across three supportive housing facilities in New York City. Rather than surveying attitudes with questionnaires, the investigators collected qualitative data, transcribed the discussions, and analyzed them using the RADaR technique, a rapid and rigorous qualitative data analysis method developed for applied research. The approach allows researchers to move quickly from raw transcripts to structured themes while preserving the nuance of what participants actually said, an important consideration when the subject matter touches on trust, vulnerability, and lived experience.

The findings reveal a population that is neither naively enthusiastic nor reflexively hostile toward AI. Participants expressed strong concerns about data privacy and the potential misuse of their information, worries that are amplified for people whose psychiatric histories carry heavy social stigma. At the same time, they recognized real potential in AI tools to support medication reminders, mental health maintenance, and daily functioning. This duality, hope shadowed by caution, runs through the entire study and complicates the simple narratives that often dominate public discussion of chatbots and mental health. For this community, the question is not whether AI companions should exist, but under what conditions they can be trusted.

Trust, the study found, is not a fixed trait that users either bring to or withhold from a system. Instead, it is something that emerges from design choices. Participants reported that their confidence in AI increased when tools were personalized, human-centered, and co-developed with people who understand the lived experience of serious mental illness. That last condition may be the most consequential. A chatbot designed without input from people who have navigated psychosis, mania, or severe depression may miss the signals that matter most, or worse, respond in ways that feel dismissive or unsafe. Co-design, in which end users participate as partners rather than test subjects, emerged as a key factor shaping perceived trustworthiness.

The technical implications of this are significant. Modern conversational AI systems, particularly those built on large language models, are trained on broad datasets that reflect general human communication, not the specific communication patterns, needs, and sensitivities of people with serious mental illness. Personalization in this context means more than remembering a user’s name. It involves calibrating tone, anticipating cognitive and emotional states, respecting boundaries around sensitive topics, and integrating gracefully with clinical care teams. Participants in the study effectively described a set of design requirements that developers of mental health AI would do well to treat as specifications rather than suggestions.

Yet the study also documents a darker set of worries that the participants themselves raised: the fear of losing critical thinking and becoming overreliant on AI companions. This concern is especially poignant for a population where autonomy and self-efficacy are hard-won achievements. An AI that is always available, always agreeable, and always attentive could, in principle, substitute for human connection rather than supplement it. The researchers note that it is essential these technologies lessen, rather than deepen, feelings of loneliness, a distinction that sounds subtle but carries enormous weight for the design and deployment of companion AI in mental health care.

These concerns do not arise in a vacuum. The study’s reference list situates the work within a rapidly growing literature on the mental health risks of large language model chatbots, including recent scoping reviews on chatbot-related harms, analyses of AI-associated delusions and mechanisms of delusion co-creation, and debates in leading psychiatry journals about whether generative AI chatbots might increase psychosis risk. For people with schizophrenia spectrum disorders in particular, the possibility that an AI companion could reinforce distorted thinking is a clinical question of the first order. The study’s participants, by voicing their own concerns about overreliance, demonstrated a sophisticated awareness of these risks that contradicts stereotypes about people with serious mental illness being unable to engage critically with technology.

At the same time, the broader research landscape suggests why AI companions hold such appeal. Loneliness and social isolation are widespread among adults in the United States and are especially pronounced among people with serious mental illness, who face stigma, social withdrawal, and reduced lifespans compared with the general population. Some studies have even found that third-party evaluators perceive AI responses as more compassionate than those of expert humans, and systematic reviews of chatbots versus human healthcare professionals have examined empathy in patient care with mixed but intriguing results. Against that backdrop, a companion that offers unconditional availability and a consistent, nonjudgmental presence has an obvious emotional pull, particularly during the nights and weekends when human support is scarce.

The study’s methodology also deserves attention for what it models. By recruiting participants through supportive housing facilities and partnering with community organizations, the research team reached a population that digital health studies frequently overlook. People with serious mental illness have historically been excluded from technology design processes, sometimes on the assumption that they cannot use or evaluate complex tools. Earlier qualitative work on how patients with schizophrenia use the internet, and international consensus efforts on digital mental health for severe mental illness, have pushed back against that exclusion. This study extends that tradition by treating participants as experts in their own lives, capable of articulating precisely what would make an AI companion feel safe, useful, and trustworthy.

The path forward, the authors suggest, lies in using these lived-experience insights to guide the development of AI tools for people with serious mental illnesses. That means building privacy protections that address the specific fears participants raised, designing personalization that respects clinical complexity, involving peer specialists and people with lived experience throughout the development cycle, and building in safeguards against overreliance. It also means honest acknowledgment of the tensions involved: the same system that tells a lonely person someone cares can become a substitute for the human relationships that recovery ultimately depends on. As AI reshapes healthcare through clinical decision support, communication tools, and patient monitoring, this study offers a reminder that the people most affected by these technologies have clear views about how they should work, and that listening to them is not just ethical but essential to building systems that actually help.

Subject of Research: Perspectives of people with serious mental illness on artificial intelligence-based companions in mental health care

Article Title: Perspectives of People with Serious Mental Illness on Artificial Intelligence-Based Companions: “If no one told you I love you today, you have someone that cares for you.”

Article References: Tapiavala, V., Chilakapati, S. S., Moore, D., Morales, S., Kumaran, A., Heller, R., Werlin, J., & Fortuna, K. L. (2026). Perspectives of People with Serious Mental Illness on Artificial Intelligence-Based Companions: “If no one told you I love you today, you have someone that cares for you.”. Community Mental Health Journal. https://doi.org/10.1007/s10597-026-01732-4

Image Credits: AI Generated

DOI: 10.1007/s10597-026-01732-4

Keywords: artificial intelligence, serious mental illness, AI companions, loneliness, schizophrenia, bipolar disorder, major depressive disorder, co-design, data privacy, digital mental health, qualitative research, supportive housing