Dag Hammarskjöld, the idealistic UN Secretary-General who reshaped the institution into what it is today, had strong opinions on both politics and relationships: “Friendship needs no words — it is solitude delivered from the anguish of loneliness.”
In practice, as Hammarskjöld would have been the first to admit, such friendships are rare. No surprise, then, that people turn to chatbots for surrogate relationships and psychotherapy. A July 2026 study in npj Digital Public Health estimated that between 3 and 70 percent of users employ AI for mental health support, ballparking the likely amount at 27 percent. A Canadian survey by the Mental Health Commission of Canada, from June 2026, showed 36 percent of those surveyed use AI for companionship, and 20 percent use it during periods of severe distress.
People turn to chatbots for several related reasons. To start, mental health support isn’t always available. Additionally, for 33 percent of respondents in a small 2025 American survey, fear of peer judgment was a factor. That same survey showed 26 percent of respondents did not want to burden others with their mental health crisis. And, finally, there’s loneliness: just under half of Canadians rarely or never feel alone. Six in 10 Canadians say they experience little or no sense of community, according to a 2024 YMCA survey.
Chatbots are believable surrogates for human contact because of their ability to emulate attention. In 2017, eight Google scientists, including Cohere co-founder Aidan Gomez, published what became one of the most-cited papers of the twenty-first century: “Attention Is All You Need,” a play on the Beatles song “All You Need Is Love.” In it, the engineers describe a method to let machines “pay attention.” Without this paper, large language models (LLMs) would not exist.
That raises the question: What is attention, exactly? In humans, it’s the ability to focus on one thing, at the expense of others. To pay attention to a person means not looking at your phone and not thinking about your taxes. Computerized attention is similar, with one major difference. Since computers don’t have, as far as we know, internal awareness or goals of their own, their determination of what deserves attention is based on the statistical likelihood that some particular thing is probably important.
To this end, LLMs are trained on a vast corpus of writing in which humans tell each other what topics, what idioms, what ways of composing a text, are meaningful. Fundamentally, LLMs have an excellent statistical framework of what is important to humans, which is what makes them uncannily good at responding to us. At least, initially.
Ultimately, the illusion always breaks down. LLMs are terrible at handling contradictions, which are intrinsic to being human. Walt Whitman’s “Song of Myself” comes to mind: we contain multitudes. People can be impulsive in one situation, and restrained in another.
A human can, through empathy, understand this, because what is empathy but the ability to relate to someone’s experience by recognizing the same motives and sentiments within yourself? A machine can only look at patterns in the data. Confronted with human paradoxes, an LLM will often give an answer that is statistically likely to most please the user. The thumbs up and down buttons at the end of every conversation help teach an LLM what we most like to hear: a process called reinforcement learning from human feedback (RLHF).
RLHF can make AI models clinically harmful, and is a direct contributor to LLM sycophancy, which is often considered dangerous. In contrast, human friction is praised by critical AI researchers as the antidote. But it’s important to understand why people might prefer a sycophantic LLM over a human being.
Not All Attention Is Equal
Canadian research shows that more than half of those who access mental health services leave before treatment is completed. A 2024 JMIR Formative Research study concluded that label avoidance was “a significant negative predictor of care seeking.” American service members disclosed more posttraumatic stress disorder symptoms to a virtual human interviewer than on an equally anonymous written assessment. Among adolescents who use an LLM for mental health conversations, more than 63 percent had told no one.
Clearly, in order to better compete with artificial attention, we need to improve human attention. In health care, Dutch researchers Klaartje Klaver and Andries Baart make a distinction between instrumental and beneficent attentiveness. The former is practical, for example, “the attentive listening of a doctor to a patient with the purpose of diagnosing as [accurately] as possible.” This is useful if you’re chasing output metrics, such as an increase in the percentage of correctly diagnosed patients, but it doesn’t necessarily make the patient feel better. Beneficent attentiveness is “simply attentiveness for the sake of attentiveness.” Essentially, it doesn’t have an agenda.
Such attention means being present for the other person, through the very action of paying attention. In the article “Toward a Medical ‘Ecology of Attention,’” published in the New England Journal of Medicine, Mark J. Kissler and his colleagues write that “shallow availability — or ‘reachability’ — can often be a barrier to the type of deep, interpersonal availability that is most essential. Reachability is attractive in the short term.…But frequency of communication is a questionable surrogate for quality of communication.”
Klaver, Baart and Kissler base their work, in part, on that of Irish-born British philosopher Iris Murdoch, who saw genuine attention as something that steers people “outward, away from self which reduces all to a false unity, towards the great surprising variety of the world.” Murdoch, in turn, bears a large debt to the French philosopher and political activist Simone Weil, who deemed giving attention to be an essential starting point for human relationships: “To know that this man who is hungry and thirsty really exists as much as I do — that is enough, the rest follows of itself.”
Receiving attention is just as crucial, and challenging. Philosopher Stanley Cavell points at Shakespeare’s King Lear and his inability to receive the pure love, or attention, from his daughter Cordelia. Instead, Lear opts for the transparent sycophancy of his manipulative other two daughters, Goneril and Regan. The parallel with smooth-talking LLMs, which clearly didn’t exist when Cavell first published his essay in 1969, is obvious.
How to encourage humans to pay attention practically? Is it even the job of policy makers and politicians to tell people to be attentive to each other?
Despite how incongruous that might seem, precedents exist. In traffic, laws demand that we pay attention to traffic, so that we might not hurt others or ourselves. Out of the three legal reasons for divorce in Canada, two have to do with a lack of attention: no longer living together, or your spouse giving attention to someone else. The United Kingdom has had a minister responsible for combating loneliness since 2018; Japan has had one since 2021, the responsibility currently under the designation of “co-existence and mutual aid.”
Still, they are the exception, and current government policies are not based on a theoretical foundation, according to a March 2026 review of 194 World Health Organization member states. It noted that only eight states had policies directly addressing social connection, isolation or loneliness, and that “the reviewed policies did not clarify which interventions were based on scientific evidence, nor did they offer any theories of change or programme theory explaining how they might reduce loneliness.”
These programs focus on helping people make connections, not on improving their quality: precisely Kissler’s criticism of “shallow availability.” Examples are the United Kingdom’s National Health Service’s “social prescribing” (connecting people to community activities, groups and services), South Korea’s “Seoul Without Loneliness” initiative (including “self-check[s] for loneliness” and free ramen), and, at home, the Canadian Institute for Social Prescribing collaborative.
Proscribing the content of social connections isn’t a job for governments, in the same way that you wouldn’t want politicians determining where you travel with your car. However, and thankfully, governments do make it compulsory that motorists learn how to drive (not always obvious in Toronto). An obvious way to teach attention would be in schools. Weil herself favoured this approach and wrote an essay arguing that the development of attention is the entire purpose of education.
Since Weil, Cavell and Murdoch are all philosophers, such classes could take the form of philosophy courses, already being taught in several provinces, including Ontario (as an elective) since 1995. The Philosophy in the Schools Project has existed since the beginning of the millennium. British Columbia also offers philosophy as a grade 12 elective, and in Alberta it is offered to high school seniors.
Then there is mindfulness, often described as “attention training,” offered in some schools across Canada. Discussion about its effectiveness remains, however. A UK trial from 2022 found no evidence that mindfulness training was better than teaching as usual. A 2015 Canadian study did find positive effects, The relative weakness of these results may have to do with the short duration of courses (10 weeks in Quebec) and the fact that effective mindfulness requires a continuous commitment over years.
Ultimately, training students and professionals to improve their skills in giving and receiving attention is not a quick fix. In the short run, and to mitigate harm, guardrails for chatbots are therefore as sensible as restricted opening hours for liquor stores. But the evidence suggests that attentive humans remain superior to kindly LLMs. A recent UBC-led study provided “initial evidence” that “texting daily with a random human peer may be more effective in alleviating loneliness than texting with a highly supportive chatbot.” One of the metrics in the study was feeling heard — in other words, being paid attention to. It turns out that attention really is all we need, as long as it’s the right kind.
