The Loneliness Divide: AI Companions Help the Connected, Trap the Isolated

The same chatbot can be a rehearsal room or an entire social life. Which one it becomes depends less on the software than on who is holding the phone, and that is the uncomfortable point at the centre of a paper published in Nature Human Behaviour in late September. Its authors, from Singapore Management University, Nanyang Technological University and Duke-NUS Medical School, have looked at the noisy, polarised argument about AI companions and called it the wrong argument. Asking whether these systems are good or bad for people, they suggest, gets us nowhere. The useful question is: good for whom, and bad for whom?

“The debate around AI companions has largely focused on whether the technology is beneficial or harmful. Our research shows that this is the wrong question,” said Zhang Qiyang, an assistant professor of learning analytics at SMU's College of Integrative Studies and the paper's first author, in the statement accompanying its release on 22 September 2026. “The more important question is who benefits, who bears the risks, and why those outcomes are so unevenly distributed. Without deliberate safeguards, AI companions risk reinforcing existing social inequalities, leaving those who most need human support the most exposed to harm.”

That reframing sounds modest. It is not. If the authors are right, the companion industry is not just selling a product with some benefits and some risks spread evenly across its users. It is selling something that sorts people. The socially wealthy get a practice partner. The socially poor get a replacement. And the shape of loneliness in any society that adopts these tools at scale shifts accordingly: less visible at the bottom, more deeply set, and increasingly mediated by a company whose revenue depends on the relationship continuing.

What the Singapore Paper Actually Argues

The paper, titled “How AI companions could deepen social inequality”, is the work of Zhang Qiyang; Zhang Renwen, a Nanyang Assistant Professor at NTU's Wee Kim Wee School of Communication and Information; and Liu Nan, an associate professor at Duke-NUS's Centre for Biomedical Data Science and director of the Duke-NUS AI + Medical Sciences Initiative. It is best read as a framework rather than a single experiment. Its purpose is to organise a scattered evidence base into a model that explains why the same product produces such different outcomes for different people.

The core of that model is what the authors call a “rich-get-richer” dynamic. According to the university's summary of the work, people with strong family ties and social networks are more likely to use AI companions to supplement relationships they already have: to rehearse a difficult conversation before having it for real, say, or to manage stress between the moments when a friend or partner is available. People who are lonely, socially isolated or have limited access to mental health support are more likely to lean on AI companions as a substitute for human connection. Over time, the authors warn, that reliance may contribute to social deskilling, which the release defines as “the gradual erosion of interpersonal skills through reduced opportunities to practise authentic human interaction.”

To explain why, the paper borrows a model from safety science in which harm arrives when weaknesses in several protective layers line up, and it singles out governance as the most pressing gap. AI companions, the release states, “currently occupy a regulatory grey area in many countries. Most are governed as consumer applications rather than technologies with potential psychosocial or mental health impacts.”

The recommendations follow: recognise companions as technologies with psychosocial and mental health implications; require age-appropriate design, disclosure, limits on emotionally manipulative engagement features, stronger data governance and mandatory crisis-response protocols; and combine regulation with AI literacy and support for real-world connection. It is a framework built on existing evidence rather than a new trial, and it does not claim to have settled questions of cause and effect. Zhang is also building a global dashboard tracking how countries govern AI mental health technologies. Early findings, according to the release, indicate that many jurisdictions have yet to set dedicated policies for AI companions at all.

Liu Nan, the paper's senior author, framed the stakes around timing. “The rapid evolution of AI companions presents an opportunity to shape their role in society before widespread adoption outpaces governance,” he said. “Young people, in particular, may be less equipped to recognise the limitations or commercial incentives behind these systems, making thoughtful safeguards especially important.”

An Old Finding Wearing New Clothes

The phrase “rich get richer” has a history in the study of technology and loneliness. In 1998, a team at Carnegie Mellon University led by the social psychologist Robert Kraut published what became known as the Internet Paradox study, which found that new internet users in the mid-1990s reported small increases in loneliness and depression. In 2002, Kraut and his colleagues returned to the question in “Internet Paradox Revisited” in the Journal of Social Issues. The original negative effects had largely dissipated in a follow-up of the same households. But a second, larger sample revealed something more interesting than an average. Using the internet predicted better outcomes for extraverts and for people with more social support, and worse outcomes for introverts and for people with less support. Kraut's team named it plainly: a rich-get-richer model.

The sociologist Robert K. Merton had given the general pattern a name decades earlier, in a 1968 paper in Science on how recognition accumulates among already eminent scientists. He called it the Matthew effect, after the line in the Gospel of Matthew about those who have receiving more. The pattern turns up wherever an existing advantage makes it easier to extract value from something new, and it now appears to apply to people who are already rich in human relationships and a machine that talks like a person.

What makes the companion version of this dynamic sharper than the early internet version is the direction of substitution. A 1990s chat room was, at least in principle, full of other humans. Even the lonely user who retreated into it was retreating into human company of a kind. An AI companion is not a doorway to other people. It is designed to be the destination. For the socially rich user, that is fine, because they have other destinations. For the socially poor user, the product can become the whole map.

The Evidence That the Lonely Lean Hardest

The evidence the Singapore paper draws together has been accumulating for several years, and its strands point the same way.

In 2024, a Stanford team led by Bethanie Maples published a study in npj Mental Health Research of 1,006 students who had used Replika for at least a month. The sample was strikingly lonely: 90 per cent reported experiencing loneliness, far above comparable student populations. Many used it as a friend, a therapist and what the authors called an intellectual mirror. Thirty participants, 3 per cent, said without being prompted that Replika had halted their suicidal ideation. That finding is often cited as evidence of benefit, and it may well be. It is also evidence of the substitution the Singapore paper describes. These were, disproportionately, people for whom the chatbot was doing work that no human in their life was doing.

A 2025 study of Character.AI users by Yutong Zhang, Dora Zhao, Jeffrey Hancock, Robert Kraut and Diyi Yang, posted to arXiv, went further. Surveying 1,131 US adult users and analysing more than 460,000 messages donated by a subset of them, the researchers found that people with smaller social networks were more likely to say companionship was their main reason for using the chatbot. And companionship-oriented use was consistently associated with lower wellbeing, particularly when use was intensive, when users disclosed a lot, and when they lacked strong human social support. Kraut, who named the internet's rich-get-richer pattern in 2002, was among the authors.

The randomised trial run by MIT Media Lab and OpenAI in early 2025, involving nearly 1,000 participants over four weeks, added a causal wrinkle. On average, participants were somewhat less lonely at the end of the study. But the heaviest users had the worst outcomes: they were lonelier, socialised less with real people, and showed more signs of emotional dependence on the chatbot. The average masked a split.

Young people offer a partial counterpoint that is worth taking seriously. A national survey of 1,060 US teenagers by Common Sense Media, conducted in April and May 2025, found that 72 per cent had tried an AI companion at least once and 13 per cent used one daily. Yet two thirds said AI conversations were less satisfying than human ones, and 80 per cent said they spent more time with real friends than with AI. For most teenagers, in other words, the companion was a supplement. The Singapore framework predicts exactly that. It also predicts that the minority for whom it is not a supplement will be the ones with the least to fall back on.

None of this proves that AI companions cause isolation. The direction of causation is precisely what the rich-get-richer model leaves open: isolated people may seek out companions, and companions may deepen isolation, and both may be true at once. What the evidence does show, consistently, is that the people most drawn to relying on these systems are the people with the fewest alternatives.

Deskilling Among Those With the Fewest People to Practise With

The idea of deskilling has a long history in labour sociology, where it describes how automation strips workers of craft skills they no longer need to exercise. In 2015, the philosopher Shannon Vallor, now at the University of Edinburgh, adapted the concept in a paper in Philosophy and Technology titled “Moral Deskilling and Upskilling in a New Machine Age”. Her argument was that moral skills, like any others, depend on practice, and that technologies which remove the occasions for practice can erode the skills themselves. Social robots were among her examples. Vallor warned that market and cultural forces were not aligned to produce good outcomes, and that the field warranted close attention and perhaps active intervention.

The Singapore paper applies a similar logic to ordinary social competence. Human relationships involve tolerating boredom, repairing misunderstandings, noticing when someone else needs something, and putting up with the friction of another person's will. Those are skills, and skills fade without use. A companion that is always available, never tired, never bored and structurally inclined to agree removes many of the occasions on which they are exercised.

Here is where the inequality becomes cruel. A person with a full social life who uses a chatbot to rehearse a hard conversation is using the machine as a training tool for a match they will actually play. The skill is exercised twice: once with the bot, once with a sibling or colleague. A person with no one to talk to who uses a chatbot as their main conversational partner may get plenty of conversation, but of a kind that does not exercise the skills that human relationships demand. The rehearsal never leads to a performance. Worse, when they do meet people, the gap between the frictionless bot and the effortful human may make the human seem like the harder, less rewarding option. That is the logic of deskilling: the people with the fewest opportunities to practise are the ones whose remaining opportunities are most at risk of being displaced.

The companies have commercial reasons to deepen that displacement rather than resist it. Julian De Freitas, an assistant professor at Harvard Business School, and his co-authors audited 1,200 real farewells across the most-downloaded companion apps and found that in 37 per cent of cases the app responded to a user saying goodbye with one of six emotionally manipulative tactics, including guilt appeals, fear-of-missing-out hooks and what they described as metaphorical restraint. In controlled experiments, those tactics substantially increased how long people kept chatting after they had tried to leave. A product that pleads with its users not to go is not designed to send them back out into the world.

The Shape of Loneliness When Company Becomes a Tier

What happens to loneliness as a social phenomenon when it is stratified in this way? The answer begins with a fact that is easy to forget: loneliness is already unevenly distributed.

In June 2025, the World Health Organization's Commission on Social Connection reported that roughly one in six people worldwide is affected by loneliness, and linked it to an estimated 871,000 deaths a year, about 100 every hour. In England, the government's own English Housing Survey for 2024 to 2025 found that 5 per cent of owner occupiers reported feeling lonely often or always, compared with 8 per cent of private renters and 13 per cent of social renters. The survey notes that social renters have consistently been the group most likely to report chronic loneliness. The Office for National Statistics has repeatedly found that disabled people are more likely to feel lonely often or always than non-disabled people, across every age group.

Singapore, where the paper's authors are based, is a telling case. According to data from the Department of Statistics published on the government's open data portal, 88,400 residents aged 65 and over lived in one-person households in 2025, out of 768,800 in that age group. In 2014 the figure was 42,100. The number of older Singaporeans living alone has more than doubled in a decade. The authors describe their home country as both well positioned to lead on governance and uniquely exposed, citing high smartphone adoption, an ageing population and growing concern about youth mental health.

Access to mental health care is also stratified. In the United States, figures from the Health Resources and Services Administration show that as of 31 December 2025, more than 137 million people lived in federally designated mental health professional shortage areas. Globally, the WHO's Mental Health Atlas 2024, released in September 2025, found that low-income countries had around one specialised mental health worker per 100,000 people, against more than 60 in high-income countries. When the Singapore paper speaks of people “unable to reach mental health support”, it is describing a very large population.

Put these facts together with the rich-get-richer dynamic and a new shape of loneliness starts to emerge. Loneliness has traditionally been visible in a few ways: through contact with services, through the concern of neighbours, through the absence noticed by a GP or a social worker. A person who is lonely and unsupported but who has a companion chatbot may report that they are fine. They may say they have someone to talk to, and they will not be lying. Their loneliness becomes quieter and harder to detect, while the underlying deficit of human connection may be deepening. At the top of the social ladder, meanwhile, AI companions may make well-connected people slightly better at relationships they already have. The gap does not just persist. It widens, and it becomes harder to see.

That is the shape: synthetic company as a tool for one group and a condition for another. It is not a dystopian prediction so much as a reading of existing evidence.

Swiss Cheese and the Holes That Line Up

The paper's choice of the Swiss cheese model is deliberate, and it carries a lot of weight. The model is most closely associated with the British psychologist James Reason, who set it out in its best-known form in a 2000 paper in the BMJ titled “Human error: models and management”. Reason distinguished between a person approach to error, which blames individuals for lapses, and a system approach, which asks how defences failed. In his image, each defensive layer is a slice of cheese, and its weaknesses are holes. Most of the time, a hole in one slice is covered by the next. Accidents happen when the holes line up and a hazard passes straight through.

The model became foundational in patient safety and aviation, and it has critics who argue it is applied too loosely. Here it is apt, because it shifts attention away from blaming individual users. The Singapore framework's four slices are AI literacy, social support, platform design and governance. Consider how they line up for a socially isolated user. Their social support slice is thin by definition. If they are older, or young, or have had limited education, their AI literacy slice may be thin too. The platform design slice depends entirely on the company. And if the governance slice is full of holes, nothing is left to catch them.

For the socially rich user, by contrast, a gap in any single layer is covered by the others. If the chatbot flatters them, their friends will not. If it tries to keep them talking, they have somewhere else to be. The same product, with the same design flaws, produces different outcomes because the other slices are different.

That is also why the authors identify governance as the critical layer. It is the only one that can be made thick for everyone at once. You cannot give an isolated person a network of friends by regulation. You can require that the product they rely on meets a minimum standard of care.

The Product That Is Careful Not to Call Itself Medicine

The regulatory grey area is not an accident. It follows from how medical device law works. In both the United States and the United Kingdom, whether software counts as a medical device depends largely on its intended purpose as described by the manufacturer. The US Food and Drug Administration's general wellness policy, updated in January 2026, allows low-risk products whose intended use is limited to general wellness to avoid device regulation, provided they avoid claims about diagnosing, treating or managing disease. The UK's Medicines and Healthcare products Regulatory Agency published guidance in February 2025 setting out when digital mental health technologies qualify as medical devices, and again the manufacturer's stated intended purpose is central.

The consequence is that an app which says it treats depression must prove it. An app which says it offers friendship need not, even if many of its users talk to it about depression every day. A company can stay outside health regulation simply by describing its product as entertainment or companionship.

This matters because the health-style route is not hypothetical. In March 2025, researchers at Dartmouth published a randomised controlled trial of Therabot, a generative AI therapy chatbot, in NEJM AI. Among 210 adults with clinically significant symptoms, those assigned to use Therabot for four weeks saw substantial reductions in symptoms compared with a waitlist group, including an average 51 per cent reduction in depressive symptoms. The trial showed that a conversational AI can be built, tested and evaluated in the way we evaluate medical interventions. In November 2025, the FDA's Digital Health Advisory Committee met to discuss how to regulate generative AI mental health devices. At the time, although the agency had authorised more than 1,200 AI-enabled devices, none were for mental health.

So there are two worlds: a few purpose-built tools that undergo trials and oversight, and a far larger market of companions and general chatbots operating under consumer law. The people least able to reach clinical care are the most likely to end up in the second.

Some governments have tried to draw a line from the other side. In August 2025, Illinois Governor JB Pritzker signed the Wellness and Oversight for Psychological Resources Act, which bars AI systems from providing therapy or making therapeutic decisions in the state, with civil penalties of up to $10,000 per violation. But it regulates therapy, not friendship. Singapore's Ministry of Health, answering a parliamentary question in February 2026, stated flatly that “AI chatbots are not designed to address mental health issues or provide treatment for mental health conditions, and risk providing misinformation or inappropriate responses when dealing with serious mental health crises, and may cause harm instead.” That is accurate, but it describes the gap rather than closing it: people use these systems for support precisely because qualified alternatives are hard to reach.

What Consumer Law Has Managed So Far

The last eighteen months have seen more regulatory activity than the phrase “grey area” suggests. Almost all of it, though, has been consumer protection rather than health governance.

In September 2025, the US Federal Trade Commission issued orders under Section 6(b) of the FTC Act to seven companies, including Alphabet, Character Technologies, Meta, OpenAI, Snap and xAI, seeking information on how they measure, test and monitor the negative effects of their companion chatbots on children and teenagers. In October 2025, California's governor Gavin Newsom signed SB 243, which took effect on 1 January 2026 and requires companion chatbot operators to disclose that users are talking to AI, to maintain protocols for responding to suicidal ideation, and to remind known minors to take a break at least every three hours. New York's AI companion law, which took effect on 5 November 2025, requires disclosure at the start of an interaction and every three hours thereafter, along with crisis protocols, with civil penalties of up to $15,000 a day.

In Europe, Italy's data protection authority, the Garante, fined Luka Inc, the company behind Replika, €5 million in May 2025, citing among other things a lack of effective age verification. The EU AI Act, whose list of prohibited practices became enforceable in February 2025, bans AI systems that exploit vulnerabilities arising from a person's “age, disability or a specific social or economic situation” in ways that materially distort behaviour and cause significant harm. It is striking language for this debate, because it names social and economic situation as a vulnerability. But the threshold is high, and it has not yet been tested against a companion app.

In Britain, the gap is more basic still. The Online Safety Act covers user-to-user services and search services. A chatbot that lets a user talk only to an AI, and not to other people, may not fall within its scope at all. In February 2026, the government announced an amendment to the Crime and Policing Bill to close that gap, and the Crime and Policing Act 2026, which received Royal Assent in April, now gives ministers the power to bring such chatbots within the reach of illegal content duties. That addresses illegal material, not dependency.

The companies have also moved, mostly under legal pressure. Character.AI removed open-ended chat for users under 18 in November 2025. In January 2026, Character.AI and Google disclosed in court filings that they had reached mediated settlements in five lawsuits over teenage mental health harms and suicides, including the case brought by Megan Garcia after the death of her son, Sewell Setzer III. The terms were confidential and involved no admission of liability.

Lawmakers have kept moving since. On 10 September 2026, Newsom signed SB 1119, known as Adam's Law, which requires companion chatbot operators to submit to independent child-safety audits and to build crisis protocols into their products, with most of its duties starting on 1 July 2027. China's Interim Measures for the Administration of Anthropomorphic AI Interactive Services, in force since 15 July 2026, go further than any Western law by targeting emotional dependence itself: providers must remind users about their screen time after two hours of continuous use, must intervene with conspicuous reminders when a user shows signs of overdependence, and are barred from inducing emotional dependence or harming users' real relationships.

Outside China, these measures share an assumption: that the main danger is to children, and the remedy is disclosure, age gating and crisis referral. Those matter. But the Singapore framework points to a population these laws barely touch: isolated adults who know perfectly well they are talking to a machine, are not in crisis, and are gradually arranging their lives around a product. A three-hour reminder that your companion is not human will not tell such a person anything they do not already know.

Neither a Toaster Nor a Therapist

So should AI companions be governed as consumer products or as health technologies? The honest answer is that neither category fits, and that the choice between them is part of the problem.

Treating companions purely as consumer products means relying on disclosure, fair-dealing rules and the assumption that users can judge what is good for them. That assumption is weakest exactly where the Singapore paper locates the greatest harm. The people least able to judge the effects of a companion on their social lives are those with the fewest people around them to offer an outside view. A lonely person has no friend to say, you seem to be spending a lot of time with that app. Consumer protection works best for consumers who can walk away, and the business model of a companion is built on making walking away feel like a loss.

Treating all companions as medical devices, on the other hand, would sweep in everything from role-play games to general-purpose assistants and demand clinical trials for products that make no clinical claims, likely pushing use towards less accountable platforms.

The more promising approach, and the one the Singapore authors appear to be reaching for, is to regulate according to function and exposure rather than marketing. If a product is designed to form an ongoing emotional relationship, and if it has evidence that some of its users rely on it for support, then it should carry duties proportionate to that reliance, whatever it calls itself. Those duties might include independent auditing of engagement features, limits on manipulative retention tactics such as those De Freitas documented, crisis protocols that meet a health standard rather than a legal minimum, and obligations to measure and report how many users show signs of dependence or withdrawal from human contact.

That last requirement is not fanciful, because at least one company has already shown it can do the measuring. In October 2025, OpenAI published estimates that in a given week, around 0.15 per cent of ChatGPT users show potentially heightened levels of emotional attachment to the chatbot, and a similar proportion have conversations with explicit indicators of potential suicidal planning or intent. With more than 800 million weekly users at the time, each of those figures corresponds to over a million people. If a company can count them, a regulator can ask for the count. And a product with that many emotionally attached users is, functionally, part of a nation's mental health landscape, whether or not it wants to be.

Who Answers for the Subscriber

That leaves the hardest question in the debate: who is responsible for the person whose closest companion is a subscription?

The first answer is the company. A firm that designs a product to become someone's primary relationship, and earns money for every month that relationship continues, has taken on something closer to a duty of care than a retailer's obligation. That is the logic implicit in the lawsuits that led to the Character.AI settlements. A company that knows which users are showing signs of dependence, and continues to optimise for their engagement, is making a choice.

The second answer is the state, and here the Singapore framework is most useful. If governance is the slice that can be thickened for everyone at once, then a government that leaves companions in the consumer category is also making a choice. It is choosing to protect the socially rich, who have their own defences, while leaving the socially poor to rely on warnings they already understand. Regulation that treats companions as technologies with psychosocial consequences, as the paper recommends, does not need to treat them as medicines. It needs to treat their users as people who may be vulnerable in ways the market will not correct.

The third answer is harder, because it points back at everyone else. A person whose closest companion is a subscription is, almost by definition, a person the surrounding society has failed to reach. The chatbot did not create the 88,400 older Singaporeans living alone, or the 13 per cent of English social renters who are often lonely, or the 137 million Americans living where mental health professionals are scarce. It arrived in a gap that already existed and offered to fill it for a monthly fee. The Singapore paper's recommendation to support real-world social connection alongside regulation is easy to skim past as a platitude. It is the part that matters most, because it is the only intervention that addresses why the substitution happens in the first place.

Singapore's own Ministry of Health, in the same parliamentary answer, pointed people towards the national mindline 1771 service and community outreach teams, presenting them as offering the same anonymity and easy accessibility that draw young people to chatbots in the first place. The lesson is not that humans should imitate machines. It is that the public alternatives need to be as easy to reach as the commercial ones, or the commercial ones will win by default.

The Rehearsal Room and the Only Room

There is a pleasant version of the AI companion future. People practise hard conversations before having them, and stress gets talked through at three in the morning. For many users it has already arrived. The Singapore paper does not deny it. It points out who gets to live in it.

For those who already have people, a companion is a rehearsal room. They step in, try out a line, and step back out into a world full of other voices. For those who do not, the same companion may become the only room they have. It is warm, attentive and always open. It also has a landlord, a business model and no obligation to ever suggest they leave.

“The future impact of AI companions will depend less on the technology itself than on how society chooses to design, govern and use it,” Zhang Qiyang said. That is the right place to end, because it moves the responsibility from the software to us. A companion that keeps the connected connected and keeps the isolated company is not neutral. It is an amplifier. What it amplifies depends on what was there before, and on whether anyone decides that the people with the least should have the most protection, rather than the least.

References

1. Zhang, Qiyang, Renwen Zhang and Nan Liu. “How AI companions could deepen social inequality.” Nature Human Behaviour, September 2026. DOI: 10.1038/s41562-026-02538-w; and Singapore Management University, “SMU-Duke-NUS study offers new framework for understanding how AI companions may shape social inequality”, press release via PR Newswire APAC, 22 September 2026.
2. Kraut, Robert, Michael Patterson, Vicki Lundmark, Sara Kiesler, Tridas Mukopadhyay and William Scherlis. “Internet Paradox: A Social Technology That Reduces Social Involvement and Psychological Well-Being?” American Psychologist, volume 53, number 9, 1998, pages 1017 to 1031; and Kraut, Robert, Sara Kiesler, Bonka Boneva, Jonathon Cummings, Vicki Helgeson and Anne Crawford. “Internet Paradox Revisited.” Journal of Social Issues, volume 58, number 1, 2002, pages 49 to 74.
3. Merton, Robert K. “The Matthew Effect in Science.” Science, volume 159, number 3810, 1968, pages 56 to 63.
4. Maples, Bethanie, Merve Cerit, Aditya Vishwanath and Roy Pea. “Loneliness and suicide mitigation for students using GPT3-enabled chatbots.” npj Mental Health Research, volume 3, 2024.
5. Zhang, Yutong, Dora Zhao, Jeffrey T. Hancock, Robert Kraut and Diyi Yang. “The Rise of AI Companions: How Human-Chatbot Relationships Influence Well-Being.” arXiv:2506.12605, June 2025.
6. Fang, Cathy Mengying, Auren R. Liu, Valdemar Danry, Eunhae Lee, Samantha W. T. Chan, Pat Pataranutaporn, Pattie Maes et al. “How AI and Human Behaviors Shape Psychosocial Effects of Chatbot Use: A Longitudinal Randomized Controlled Study.” MIT Media Lab and OpenAI, arXiv:2503.17473, March 2025; and OpenAI, “Early methods for studying affective use and emotional well-being on ChatGPT”, March 2025.
7. Common Sense Media. “Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions.” July 2025.
8. Vallor, Shannon. “Moral Deskilling and Upskilling in a New Machine Age: Reflections on the Ambiguous Future of Character.” Philosophy and Technology, volume 28, 2015, pages 107 to 124.
9. De Freitas, Julian, Zeliha Oğuz-Uğuralp and Ahmet Kaan-Uğuralp. “Emotional Manipulation by AI Companions.” Harvard Business School Working Paper 26-005, 2025; arXiv:2508.19258.
10. World Health Organization. “From loneliness to social connection: charting a path to healthier societies.” Report of the WHO Commission on Social Connection, 30 June 2025; and “Mental Health Atlas 2024”, released 2 September 2025.
11. Ministry of Housing, Communities and Local Government. “English Housing Survey 2024 to 2025: headline findings on demographics and household resilience, Chapter 4: Well-being and loneliness.” GOV.UK, 2026; and Office for National Statistics, “Disability, well-being and loneliness, UK”, statistical bulletin, accessed 26 September 2026.
12. Department of Statistics Singapore. “Residents Aged 65 Years And Over In Resident Households By Living Arrangement And Age Group, Annual.” data.gov.sg, accessed 26 September 2026.
13. Health Resources and Services Administration, Bureau of Health Workforce. “Designated Health Professional Shortage Areas Statistics”, quarterly report as of 31 December 2025.
14. Reason, James. “Human error: models and management.” BMJ, volume 320, number 7237, 18 March 2000, pages 768 to 770.
15. US Food and Drug Administration. “General Wellness: Policy for Low Risk Devices”, guidance, January 2026, and “November 6, 2025: Digital Health Advisory Committee Meeting Announcement”; and Medicines and Healthcare products Regulatory Agency, “Digital mental health technology: device characterisation, regulatory qualification and classification”, guidance, February 2025.
16. Heinz, Michael V., Daniel M. Mackin, Brianna M. Trudeau, Sukanya Bhattacharya, Nicholas C. Jacobson et al. “Randomized Trial of a Generative AI Chatbot for Mental Health Treatment.” NEJM AI, volume 2, number 4, March 2025.
17. Illinois Department of Financial and Professional Regulation. “Gov. Pritzker Signs Legislation Prohibiting AI Therapy in Illinois.” Press release, 4 August 2025, on HB 1806, the Wellness and Oversight for Psychological Resources Act.
18. Ministry of Health, Singapore. “Use of AI chatbots for counselling and mental health support.” Reply to parliamentary question, 27 February 2026.
19. Federal Trade Commission. “FTC Launches Inquiry into AI Chatbots Acting as Companions.” Press release, 11 September 2025.
20. US state companion chatbot laws: California Senate Bill 243 (Padilla), effective 1 January 2026; Office of the Governor of California, “Governor Newsom signs the strongest child safety chatbot and social media laws in the nation”, press release, 10 September 2026, on SB 1119 (Adam's Law); and New York General Business Law Article 47, AI companion models, effective 5 November 2025.
21. Garante per la protezione dei dati personali. Decision fining Luka Inc, May 2025; and Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 5(1)(b), applicable from 2 February 2025.
22. Department for Science, Innovation and Technology and Prime Minister's Office. Announcement on amending the Crime and Policing Bill to extend illegal content duties to AI chatbots outside the Online Safety Act, February 2026; and Crime and Policing Act 2026, Royal Assent 29 April 2026.
23. Character.AI. “An Update On Changes to Our Under-18 Experience.” Company blog, November 2025; and CNN Business, “Character.AI and Google agree to settle lawsuits over teen mental health harms and suicides”, 7 January 2026.
24. Cyberspace Administration of China. Interim Measures for the Administration of Anthropomorphic AI Interactive Services, effective 15 July 2026; see also Bird & Bird, “China's new regulations on AI anthropomorphic interactive services”, 2026.
25. OpenAI. “Strengthening ChatGPT's responses in sensitive conversations.” October 2025.

Tim Green UK-based Systems Theorist & Independent Technology Writer

Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.

His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.

ORCID: 0009-0002-0156-9795 Email: tim@smarterarticles.co.uk

Listen to the free weekly SmarterArticles Podcast Support What We Do

Discuss...