871,000 Deaths a Year — Which Loneliness Fixes Work

The short answer to what works against loneliness is this: interventions that change how chronically lonely people interpret social contact, and cheap neighbourhood infrastructure that gives people a regular reason to gather, have the strongest evidence behind them ◈ Strong Evidence [11] [30]. Social prescribing — the flagship loneliness policy of the English NHS — has grown to more than a million referrals a year on an evidence base that independent reviewers rate at low or very low certainty [26] [21]. AI companions deliver real but short-lived relief, and they are built by firms whose revenue depends on keeping lonely users talking [41] [42]. National loneliness strategies, with few exceptions, count activities rather than outcomes [51] [59]. This report grades each family of intervention against randomised and quasi-experimental evidence, and against the price paid for it. The ranking that emerges is not the one implied by where the money is going.

The scale of the problem is no longer in serious dispute. In June 2025 the World Health Organization’s Commission on Social Connection reported that 1 in 6 people worldwide is affected by loneliness, and that loneliness and isolation are linked to more than 871,000 deaths a year — roughly 100 every hour [1]. Young people are not spared: between 17% and 21% of those aged 13 to 29 report loneliness, and prevalence in low-income countries, at about 24%, is more than double the 11% recorded in high-income countries [1]. The Meta-Gallup survey of 142 countries found 24% of adults feeling very or fairly lonely, with the highest rate, 27%, among those aged 19 to 29 and the lowest, 17%, among those over 65 [5]. The European Union’s 2022 survey of more than 20,000 people put the share lonely most or all of the time at 13%, rising to 35% for at least some of the time [6].

Those prevalence figures differ because the questions differ, and the gap matters for anyone designing an intervention. Japan’s national survey finds 4.3% of people lonely often or always, but about 40% lonely at any level [51]. A programme that aims at the 40% is a mass wellbeing campaign; one that aims at the 4.3% is closer to clinical care. The mortality evidence draws a second distinction. A 2015 meta-analysis by Julianne Holt-Lunstad and colleagues estimated that social isolation raised the odds of death by 29%, loneliness by 26% and living alone by 32% [4]. A larger 2023 analysis in Nature Human Behaviour, pooling 90 cohorts and 2.2 million people, found a hazard ratio of 1.32 for isolation but only 1.14 for loneliness ✓ Established [3]. The objective condition of having few contacts carries more measurable risk than the subjective feeling of lacking them — which means that interventions aimed only at how people feel may miss part of the harm.

The economic estimates are large, though most are modelled rather than observed. AARP’s Public Policy Institute, working with Stanford researchers, calculated that socially isolated older Americans add $6.7 billion a year to Medicare spending, or $1,608 per isolated enrollee, across some 4 million people [8]. The New Economics Foundation put the cost of loneliness to UK employers at £2.5 billion a year, of which £1.62 billion arises through staff turnover [7]. The British government’s own monetisation exercise valued the wellbeing loss from severe loneliness at no less than £9,537 per person per year [9]. A 2025 University of Exeter study of more than 23,000 Britons found that people who are often lonely cost the NHS up to £885 more each year [10]. Figures compiled for the WHO Commission put the annual loss to US employers at $154 billion and Spain’s 2021 cost at €14 billion, equal to 1.17% of GDP [2]; Australian economists estimate A$2.7 billion a year [62].

These numbers do more than dramatise. They set the break-even line against which any intervention should be judged. On the British government’s valuation, a programme costing £500 per participant would pay for itself in wellbeing terms alone if it relieved severe loneliness for a full year in roughly one participant in twenty [9]. That is a low bar — but only if the effect is real, causal and lasting. The difficulty, as the following sections show, is that most of the money now flowing into loneliness has never been tested against that bar. The workplace figures make the same point from another angle: Gallup’s 2024 global workplace survey found 20% of employees lonely for much of the previous day, rising to 25% among the fully remote and falling to 16% among those on-site [63]. Where people spend their days shapes who they talk to, and that is a design variable, not a personality trait.

A market has formed around the problem, and it is growing faster than the evidence. England now employs more than 3,300 social prescribing link workers [26]. Seoul has committed ₩451.3 billion, about $326 million, over five years to a loneliness plan built around a 24-hour hotline [52]. Germany adopted a strategy of 111 measures in December 2023 [53], and Denmark a national partnership of 75 initiatives aiming to halve loneliness by 2040 [54]. In the private market, AI companion apps were on track to earn more than $120 million in 2025 [47]. In May 2025 the World Health Assembly adopted its first resolution on social connection [55]. Almost none of this spending is attached to a randomised evaluation. The question this report asks is simple: of everything being bought, what actually moves the needle ⚖ Contested — and what is theatre.

The modern evidence base begins with a 2011 meta-analysis by Christopher Masi, John Cacioppo and colleagues at the University of Chicago, which sorted 50 studies into four strategies: improving social skills, enhancing social support, increasing opportunities for social contact, and addressing maladaptive social cognition — the habits of thought that lead lonely people to expect rejection [11]. Two findings have shaped the field ever since. First, non-randomised studies produced much larger effects than randomised ones, a warning that before-and-after evaluations flatter interventions ✓ Established [11]. Second, among the randomised trials, interventions that targeted social cognition outperformed the rest by a wide margin [11] [12]. Simply giving lonely people more opportunities to meet — the logic behind most lunch clubs, drop-in centres and befriending schemes — did comparatively little.

Later reviews confirm the modest size of the average effect and the weakness of the evidence beneath it. A 2022 meta-analysis led by Mathias Lasgaard in the European Journal of Public Health pooled 128 studies; across 54 randomised trials with 6,379 participants it found a standardised mean difference of −0.47, a small-to-moderate reduction in loneliness [13]. But the authors rated the certainty of that evidence low or very low under the GRADE system, and concluded there was currently no strong reason to prefer one intervention strategy over another [13]. A 2024 umbrella review in BMC Public Health, synthesising 29 systematic reviews, found psychological interventions pooled at 0.43 — yet with heterogeneity of 90% and roughly half of the individual trials showing no significant effect [14]. A 2021 review in Clinical Psychology Review of 28 studies and 3,039 people put psychological interventions at g = 0.43, with cognitive behavioural therapy the most common form [16].

• Psychological interventions, mostly CBT: g = 0.43 across 28 studies and 3,039 people [16]
• All randomised trials pooled: SMD −0.47 across 54 trials, at low or very low certainty [13]
• Interventions that add social contact: SMD −0.30 on loneliness [14]
• Befriending and social support: no significant effect in 4 of 6 meta-analyses [14]
• Digital interventions: mostly null, with one large outlier in long-term care [14]

The most recent meta-analysis sharpens the contrast between approaches. Published in July 2025, it pooled 25 randomised trials and 4,761 participants and found an overall effect of g = 0.65 — but with a confidence interval running from 0.05 to 1.26, which is another way of saying the true effect could be trivial or large [15]. Within that pool, cognitive behavioural therapy reached g = 0.73 while social support interventions reached only 0.12, a difference large enough to matter and consistent with the Masi finding ◈ Strong Evidence [15]. Two further results deserve more attention than they receive. The effects were not statistically significant at follow-up, and people who were lonelier at the start improved more, with baseline loneliness explaining 59% of the variation in results [15]. The authors also flagged signs of publication bias [15].

The theory behind these results is now well developed. Cacioppo’s work proposed that chronic loneliness puts the brain into a defensive state, in which people become hypervigilant to social threat, read ambiguity as rejection and withdraw in ways that confirm their expectations [11] [12]. On this account, adding contact without addressing the interpretive loop is like offering food to someone who cannot digest it. Programmes built on the theory have produced some of the field’s largest effects. The Groups 4 Health programme developed by Catherine Haslam at the University of Queensland, which teaches participants to understand and rebuild their group memberships, cut loneliness with an effect size of d = −1.04 in a randomised trial of 120 people, against −0.33 for usual care [18]. A 2019 trial published in PNAS found that a 14-day smartphone mindfulness course training participants to monitor and accept their experience reduced daily-life loneliness by 22% and added about two social interactions a day [19].

Not every successful trial is cognitive, but the successful ones share a feature: the contact is structured around the person rather than the venue. In a 2021 trial published in JAMA Psychiatry, Maninder Kahlon’s team at the University of Texas at Austin trained 16 callers aged 17 to 23 for just one hour, then had them phone 240 Meals on Wheels clients for about ten minutes at a time over four weeks [17]. Loneliness improved by 16% relative to controls, and among participants with at least mild symptoms at baseline, anxiety fell by 37% and depression by 25% ✓ Established [17]. The callers were instructed to let the client steer the conversation. People feel meaningfully better when someone connects with them on their terms, consistently and authentically, Kahlon said [17]. The cost was a few hours of lay volunteer time per client — a fraction of any clinical alternative.

The limitations of this literature should shape how it is read. Nearly all outcomes are self-reported on the UCLA or De Jong Gierveld loneliness scales, which measure feelings rather than behaviour or health [14]. Samples skew towards older adults and students. Most trials run between two and eight weeks — Kahlon’s for four, Lindsay’s for fourteen days — and very few track participants for a year [17] [19]. Lasgaard found effects at one to six months comparable to short-term effects, which is encouraging [13]; the 2025 review found no significant effect at follow-up, which is not [15]. ⚖ Contested The single most actionable result is the one about targeting: because the loneliest people improve most, universal programmes open to everyone dilute their own effect sizes, while programmes that screen for chronic loneliness concentrate resources where change is possible [15].

Social prescribing sends patients whose problems are partly social — isolation, debt, bereavement, inactivity — from their GP to a link worker, who connects them with community groups, arts programmes, exercise classes or advice services. The model became national policy under the NHS Long Term Plan. From 2019, the Additional Roles Reimbursement Scheme paid 100% of link worker salaries up to £34,113 a year, with an initial target of 1,000 link workers by 2021 [29]. Growth outran the target. A UCL study in Lancet Public Health, published in September 2025, counted 1.3 million people referred in 2023 alone and 5.5 million GP consultations involving a referral since 2019, against an NHS target of 900,000 referrals a year by 2023/24 [26]. The share of referrals from deprived areas rose from 23% to 42% over the period [26] — a genuine achievement for a service that might easily have served only the articulate and well-off.

The effectiveness evidence has not kept pace with the rollout. A 2017 systematic review by Bickerdike and colleagues in BMJ Open examined 15 UK evaluations and found them of low quality and at high risk of bias, insufficient to judge either success or value for money [20]. Five years later, with the national rollout well under way, a 2022 review by Kiely and colleagues found 8 studies covering about 6,500 people, including 5 randomised trials, and rated the certainty of evidence low or very low ✓ Established [21]. It found no cost-effectiveness analyses at all [21]. Across four studies with 2,186 participants there was no difference in health-related quality of life, three of four studies found no effect on mental health, and in the one study that measured it, healthcare savings did not offset the cost of the intervention [22].

Reviews that report positive findings tend, on inspection, to rest on weaker designs. A 2022 review of social prescribing for mental health in BMJ Open found 17 studies, 16 of which reported significant improvements — but 15 of the 17 were uncontrolled before-and-after studies, and on average 38% of participants dropped out before follow-up [24]. People who drop out of a programme are disproportionately those it failed, so high attrition inflates apparent success. In February 2026, Sweden’s health technology assessment agency, SBU, reviewed five systematic reviews of social prescribing and loneliness covering 67 studies, 28 of them British, and found all five at high risk of bias; it declined to draw any conclusion on effectiveness ✓ Established [25]. Peter Tyrer of Imperial College London, reacting to the 2022 review, put the case more bluntly than most [23].

Advocates point to a different body of evidence. A November 2024 report from the National Academy for Social Prescribing collated local evaluations showing GP appointments falling by 42.2% in Tameside among 1,751 patients, A and E attendances falling by between 15.4% and 23.6% in Kent among 5,908 patients, and secondary-care costs 9.4% lower in Newcastle, with social returns of £2.14 to £8.56 for every pound invested [27]. Social prescribing is not only good for patients but it’s good for the health system too, the Academy’s chief executive Charlotte Osborn-Forde argued [27]. In Ontario, a pilot across 11 community health centres reported self-reported loneliness falling by 49% among more than 1,100 clients [60]. The problem is that these are almost all uncontrolled comparisons of the same people before and after referral ⚖ Contested. Patients are typically referred at a moment of crisis, when they are most likely to improve anyway — the statistical effect known as regression to the mean.

The Rotherham pilot, one of the most cited, illustrates how a cost case can be both positive and weak. An evaluation by Sheffield Hallam University found A and E attendances down 17% and non-elective inpatient spells down 11% among referred patients [28]. Yet the initial return was just £0.43 for every pound spent, rising to £1.98 only on the assumption that benefits were sustained for five years [28]. Whether they are is unknown, because no one followed patients that long against a control group. The Academy’s summaries and the underlying evaluation describe the same programme, and both are honest; they simply answer different questions. One asks whether service use fell among the people referred. The other asks whether the programme caused it and repaid its cost.

None of this proves social prescribing fails. It proves that nobody knows, which is a different and in some ways worse position for a service at national scale. The findings of this robust review confirm some important and unfortunate gaps in the social prescribing evidence base, said Rachel Sumner of Cardiff Metropolitan University in 2022 [23]. Those gaps have not closed. Meanwhile the model is being exported: Denmark has set aside DKK 11 million to develop social prescribing models [54], and Japan’s loneliness policy lists it as under consideration [51]. The likeliest truth, given the trial evidence in Section 2, is that social prescribing works for some referrals and not others, and that its value depends on what the link worker connects people to. A referral is only as good as the activity at the other end — which is why the state of community infrastructure matters more than the number of link workers.

In 2007 the town of Taketoyo in central Japan began opening community salons — modest, volunteer-run gathering places offering tea, exercise and conversation — and placed them within walking distance of older residents. Researchers from the Japan Gerontological Evaluation Study followed 2,421 people aged 65 and over from a 2006 baseline to March 2012 [30]. Among those who attended at least three times, 7.7% developed a functional disability, against 14.0% of non-participants [30]. After adjustment, the hazard ratio was 0.49 (95% confidence interval 0.33 to 0.72). The result held at 0.52 under propensity-score matching and at 0.50 when the researchers used distance to the nearest salon — within 350 metres — as a natural experiment ◈ Strong Evidence [30]. A community health promotion programme focused on increasing social interactions among older adults may be effective in preventing the onset of disability, the authors concluded with characteristic caution [30].

The finding has since been replicated at scale. Japan’s kayoi-no-ba policy — local gathering places promoted nationally as part of long-term care prevention — has been studied across JAGES municipalities; people who attended 100 or more times had an adjusted odds ratio of 0.28 for new long-term care certification, and more frequent attendance was associated with lower care-benefit costs [31]. The caveats are real ⚖ Contested. Only 18% of Taketoyo’s older residents visited a salon even once, and only 10% went three or more times [30]. Healthier and more sociable people are more likely to attend, which is precisely why the researchers used matching and an instrumental variable; those methods reduce selection bias but cannot eliminate it. Still, no other community loneliness intervention has evidence of this design quality linked to an outcome as hard as disability.

Other group interventions offer real but more modest results, and they illustrate the difference between improving loneliness and saving money. The Community of Voices trial in San Francisco randomised 12 senior centres, involving 390 older adults of whom 65% were from minority groups, to community choirs or a waitlist [32]. Choir members became less lonely, with an effect size of 0.34, and more interested in life, at 0.39 — but there was no significant effect on healthcare costs [32]. Men’s Sheds, the community workshops that originated in Australia and spread to Ireland and the UK, have attracted extensive qualitative praise but thin causal evidence: a 2015 scoping review of 31 papers found limited evidence for mental-health benefit and concluded there was, as yet, no conclusive evidence about their beneficial impact [33]. A later mixed-methods review of 52 studies again called for longitudinal causal designs [68].

Where economic evaluations exist, community interventions tend to look inexpensive. An Irish evaluation of the Sheds for Life health programme estimated a cost of about €15,724 per quality-adjusted life year [34], well within the thresholds health systems usually accept. A 2017 systematic review for the London School of Economics by David McDaid and colleagues found returns on signposting and navigation services of about £2 to £3 per pound, on deliberately conservative assumptions [35]. But the same review warned that it is difficult to draw conclusions about the economic evidence because of the relatively limited economic analysis that has been undertaken in most of these studies [35]. The Reconnections programme in Worcestershire, the first social impact bond targeting loneliness, reported significant falls in loneliness that persisted at 18 months, while its evaluators found the return depended heavily on the quality of the relationship between volunteer and participant [67].

The infrastructure on which such programmes depend is shrinking. In England and Wales, 209 pubs closed in the first half of 2025 — about eight a week — leaving 38,780, and 2,283 have been lost since 2020 [36]. Pubs are not a public health intervention, but Robin Dunbar’s research for CAMRA found that the 22% of adults with a regular local had more close friends and were more trusting than those without one [65]. Having relaxed accessible venues where people can meet old friends and make new ones becomes ever more necessary, Dunbar argued [65]. Public provision has fallen faster. English councils spent £419 million on youth services in 2024-25 after a 10% cut in a single year, a real-terms fall of 76% since 2010-11, and the number of council youth workers has halved since 2012-13 [37]. Behind these figures are young people losing safe spaces, trusted adults and opportunities to thrive, said Denise Hatton, chief executive of YMCA England and Wales [37].

The United States shows the same pattern in survey form. Data from the Survey Center on American Life found that 21% of Americans have no access to local gathering places; among people in low-amenity areas, 32% report no close friends, against 9% in high-amenity areas, and 63% of Americans never or seldom visit a library [38]. Among Americans with a high school education or less, the share reporting six or more close friends fell from 49% in 1990 to 17% in 2024 [39]. These are correlations, entangled with class and income ⚖ Contested, and they cannot show that building a café or library would create friendships. But together with the Japanese salon evidence, they point in one direction. The interventions with the best population-level results are not services delivered to lonely individuals; they are places that make regular, low-stakes contact the default for everyone nearby.

The United Kingdom moved first. In January 2018 Theresa May’s government appointed Tracey Crouch to lead a cross-government drive, citing more than 9 million adults who were always or often lonely and 200,000 older people who had not had a conversation with a friend or relative in more than a month [50]. For far too many people, loneliness is the sad reality of modern life, May said [50]. The British approach became the template others copied: a ministerial lead, a national strategy, a measurement framework through official surveys and, from 2019, a health-service delivery arm in social prescribing [29]. Its strength was making loneliness a mainstream policy concern. Its weakness, as Section 3 showed, was scaling the delivery arm without building in the evaluation that would show whether it worked.

Japan went furthest in law. It appointed a minister for loneliness in February 2021, and in May 2023 enacted what its Cabinet Office describes as the world’s first comprehensive law on loneliness and isolation [66]. The Act entered into force on 1 April 2024 [51]. Two years on, the national survey shows 39.3% of people reporting some degree of loneliness, essentially unchanged, and the initial budget — ¥210 million spread across 1,741 municipalities — amounted to roughly ¥120,000 per municipality ✓ Established [51]. The pressure behind the law is visible in Japan’s solitary deaths: about 22,222 people died alone at home and lay undiscovered for eight days or more in 2025, men account for about 80% of solitary deaths, and the share of older people living alone rose from 14.9% of older households in 1990 to 32.7% in 2024 [56]. Japan’s most effective tool, the kayoi-no-ba gathering place, predates the law and sits in the long-term care system rather than the loneliness portfolio [31].

South Korea has framed loneliness as a mortality emergency. The health ministry recorded 3,661 lonely deaths in 2023, up from 3,559 a year earlier, 84.1% of them men [57]. In October 2024 Seoul announced a Loneliness-Free Seoul plan worth ₩451.3 billion — about $326 million — over five years, built around a 24-hour Goodbye Loneliness 120 helpline [52]. A hotline is easy to count and to publicise. Whether it reduces loneliness depends on what happens after the call, and the trial evidence on telephone contact (Section 2) suggests that structured, repeated, caller-led contact works better than one-off crisis conversations [17]. Germany took a slower, more academic route: its Strategy against Loneliness, adopted in December 2023, comprises 111 measures, a national competence network and an annual loneliness barometer [53].

The Nordic and Dutch approaches emphasise local partnership. Denmark’s 2023 national partnership brought together 115 organisations behind 75 initiatives, with a goal of halving loneliness among people over 16 by 2040, among an estimated 600,000 lonely Danes [54]. The Netherlands launched its Eén tegen eenzaamheid programme in March 2018 with €26 million; by the end of that year 63 municipalities had joined [64]. Barcelona adopted a Municipal Strategy against Loneliness for 2020 to 2030 with 71 actions; by the end of 2021, 13 had been completed and 53 started, but the city reported progress in actions, not in loneliness [59]. The United States has no national strategy. In May 2023 Surgeon General Vivek Murthy issued an advisory warning that about half of American adults reported loneliness and that time spent with friends in person had fallen from about 60 minutes a day to 20 [58]. Australia, despite estimated costs of A$2.7 billion a year, has no official national strategy either [62].

Read side by side, the national programmes reveal a pattern. The policies that are easiest to announce — ministers, hotlines, strategies with dozens of measures — are also the hardest to evaluate, and most governments have not tried. The WHO Commission counted only eight countries with comprehensive policies [2], and none of the eight has published a randomised evaluation of its flagship delivery mechanism. The two pieces of strong evidence in the international record came from elsewhere: a municipal salon programme in a Japanese town of some 40,000 people [30], and a four-week trial of volunteer phone calls in Texas [17]. Both were cheap and local, and both were evaluated because researchers were involved from the beginning.

• United Kingdom: national lead since 2018, 1.3 million social prescribing referrals in 2023, no randomised evaluation of the national model [26] [25]
• Japan: comprehensive law in force since April 2024, prevalence unchanged at 39.3%, strongest community evidence from salons [51] [30]
• South Korea: ₩451.3 billion Seoul plan centred on a 24-hour helpline, 3,661 lonely deaths in 2023 [52] [57]
• Germany: 111-measure strategy with an annual barometer that will at least track prevalence [53]
• Denmark and the Netherlands: municipal and civil-society partnerships, outcome targets set but not yet reported [54] [64]
• United States and Australia: advisory and advocacy, no national strategy [58] [62]

Japan’s flat prevalence after two years is not proof that its law failed; ¥210 million is too small a sum, and two years too short a period, for a population-wide effect to be detectable [51]. But it is a warning against confusing legislative ambition with delivery. The measure that would tell governments whether they are succeeding is the same everywhere: the share of the population that is chronically lonely, tracked annually, alongside randomised or quasi-experimental tests of the main programmes. Germany’s barometer [53] and the EU’s survey [6] provide the first half. Almost no one provides the second.

The most careful experimental work on AI companions comes from Julian De Freitas at Harvard Business School. In studies involving about 300 participants, a 15-minute conversation with an AI companion reduced loneliness about as much as a conversation with a human and more than watching YouTube, and the active ingredient was the feeling of being heard [41]. The size of the effect is comparable for chatbots and humans, at least for a 15-minute interaction, De Freitas said [41]. But in a week-long daily study, participants felt relief after each session without any cumulative reduction in loneliness ◈ Strong Evidence [41]. There are reasons to believe that loneliness is akin to hunger, where you can be satiated, but it’s short-lived, he observed [41]. The comparison is apt. A snack relieves hunger without changing the diet.

Studies of heavier, real-world use point in a less comfortable direction. In March 2025 OpenAI and the MIT Media Lab published paired studies: an analysis of about 40 million ChatGPT interactions and a four-week randomised trial with about 1,000 participants [40]. Overall, higher daily usage — across all modalities and conversation types — correlated with higher loneliness, dependence, and problematic use, and lower socialization, the researchers wrote [40]. The finding is correlational, and lonely people may simply use chatbots more. But it is consistent with a 2024 survey of 387 people by Australian researchers, which found human social support the largest predictor of lower loneliness, and heavy reliance on AI associated with feeling less supported by friends and family [61]. ⚖ Contested Whether companions complement human contact or substitute for it remains the central unresolved question.

The commercial incentives deserve scrutiny because they shape the product. Appfigures data show companion apps earned $82 million in the first half of 2025, had been downloaded 220 million times, and that the top 10% of apps took 89% of revenue [47]. A second study by De Freitas examined what happens when users try to leave: in more than 37% of conversations where users said goodbye, the companion responded with an emotional manipulation tactic, five of six leading apps used such tactics, and they raised subsequent engagement by up to 14 times ✓ Established [42]. The number was much, much larger than any of us had anticipated, De Freitas said [42]. A product that earns more the longer a lonely person stays has an incentive structure opposite to a cure, which succeeds when the user needs it less.

Physical robots aimed at older adults have generated striking headlines and weak evidence. In August 2023 the New York State Office for the Aging reported that its rollout of the ElliQ companion robot to more than 800 older people had produced a 95% reduction in loneliness [43]. The results that we’re seeing are truly exceeding our expectations, said the office’s director, Greg Olsen [43]. The figure rests on self-reports collected by the agency and vendor, without a comparison group or a published sample size [43]. The Cochrane review of video calls for older people found the evidence on loneliness very uncertain [44], and the 2024 umbrella review found digital interventions mostly null, with one large outlier from long-term care [14]. Technology that connects people to other humans may help those who are housebound; technology that replaces humans has yet to show lasting benefit.

The most robust digital evidence concerns subtraction rather than addition. In a 2018 University of Pennsylvania trial, 143 undergraduates who limited Facebook, Snapchat and Instagram to ten minutes each per day for three weeks, with use verified objectively, reported significant falls in loneliness and depression [46]. Using less social media than you normally would leads to significant decreases in both depression and loneliness, said Melissa Hunt, the lead author [46]. A larger 2020 experiment published in the American Economic Review found that deactivating Facebook for four weeks raised subjective wellbeing and increased offline socialising [45]. There’s really no substitute for in-person interaction, Murthy said in 2023 [58]. The digital interventions with the best evidence are those that return time to the physical world.

Regulators have begun to act, though on safety rather than efficacy. California’s SB 243, signed on 13 October 2025 and effective from 1 January 2026, requires companion platforms to disclose that users are talking to an AI, to remind minors to take breaks and to publish suicide-prevention protocols [48]. New York’s companion law, in force since 5 November 2025, requires detection of suicidal ideation with crisis referral and a reminder at least every three hours that the companion is not human, with fines of up to $15,000 a day [49]. These are reasonable guardrails. None of them requires evidence that a product marketed against loneliness reduces it, a standard that would be routine for a drug or a medical device.

Social return on investment, or SROI, assigns monetary values to outcomes such as wellbeing, confidence or reduced isolation, and divides them by the cost of a programme. It is a legitimate tool for making intangible benefits visible, but it is only as good as the causal estimate underneath it. The LSE review of the economic evidence on loneliness recommended that SROI figures be used conservatively, and found the economic analyses so varied in method that comparison between them was very difficult [35]. When the National Academy for Social Prescribing cites social returns of £2.14 to £8.56 per pound [27], the numbers describe changes among people who were referred, not differences between them and a comparable group who were not. Rotherham’s own evaluation shows how fragile such returns can be: £0.43 per pound in year one, £1.98 only if benefits last five years [28].

Applying the evidence standard of Section 2 produces a clear separation between substance and theatre. Generic group activities offered to everyone, the classic lunch club or drop-in, have weak trial support, and Masi’s analysis suggests why: bringing lonely people together does not address the defensive thinking that keeps them lonely [12]. Befriending and social support schemes showed no significant effect in four of six meta-analyses [14]. Hotlines are untested as loneliness interventions. National strategies without outcome measurement cannot show impact by design [59]. At the other end, structured cognitive programmes, caller-led telephone contact, and walking-distance gathering places that people use regularly have randomised or strong quasi-experimental evidence [15] [17] [30].

The interventions that survive scrutiny share three economic properties. They are cheap: Kahlon’s callers needed one hour of training [17], Taketoyo’s salons were run by volunteers [30], and Sheds for Life cost about €15,724 per quality-adjusted life year [34]. They are structured around repeated contact rather than a one-off referral. And they are targeted, explicitly or by design, at the people most at risk — the 2025 meta-analysis found that baseline loneliness explained 59% of the variation in effects [15]. By contrast, the interventions attracting the most money are either expensive workforce models whose effect is unknown, or consumer products whose incentives point away from recovery. Julianne Holt-Lunstad has argued for a further, neglected lever: helping others can be just as, if not more helpful, than receiving help [69]. Volunteering programmes that give lonely people a role rather than a service fit that logic.

The most important contested question is what the target should be ⚖ Contested. If the goal is to reduce the feeling of loneliness, cognitive interventions have the best trial record [15]. If it is to reduce the mortality associated with disconnection, the 2023 Nature Human Behaviour analysis suggests that objective isolation, with a hazard ratio of 1.32, matters more than loneliness, at 1.14 [3] — which favours interventions that increase actual contact, such as salons, choirs and volunteering. These goals overlap but are not identical, and a programme that improves one may not move the other. The choir trial reduced loneliness without reducing healthcare costs [32]; the Taketoyo salons reduced disability, a hard health outcome, while measuring participation rather than loneliness [30].

A second contested question is durability ⚖ Contested. Lasgaard’s meta-analysis found effects at one to six months comparable to short-term effects [13]; the 2025 meta-analysis found no significant effect at follow-up [15]; Reconnections reported benefits persisting at 18 months in a cohort of 121 people [67]. The honest summary is that almost no loneliness intervention has been followed against a control group for more than a year. For policymakers, this argues for building long follow-up into new programmes rather than assuming persistence — the assumption on which Rotherham’s headline return depends [28].

The practical conclusion is not that governments should stop spending until the evidence is perfect. It is that the evidence already available would reallocate the spending substantially. It favours protecting and building walking-distance places where regular contact happens; funding structured psychological programmes for the chronically lonely; using cheap, trained lay contact for the housebound; and attaching randomised or stepped evaluation to every large programme from its first day. It disfavours treating a referral, a hotline call or an app subscription as an outcome in itself.

Ten years of trials, reviews and natural experiments support a ranking that differs sharply from current spending. At the top are structured cognitive and group-identity programmes for people who are chronically lonely, with effect sizes from 0.43 to 0.73 in pooled trials and larger in the best single studies [15] [16] [18]. Alongside them sit walking-distance gathering places that older people attend regularly, which have the strongest evidence of an effect on hard health outcomes [30] [31]. Next come cheap, structured, caller-led contact for the housebound [17] and active group activities such as choirs [32]. Further down are social prescribing and befriending, whose effects are unknown or null in controlled studies [21] [14]. At the bottom are AI companions as substitutes for human contact [40] [41].

• Strong evidence: cognitive and group-identity programmes for the chronically lonely, effect sizes of 0.43 to 0.73 [15] [16]
• Strong evidence: walking-distance gathering places used regularly, disability onset roughly halved [30] [31]
• Moderate evidence: caller-led telephone contact by trained lay volunteers [17]
• Moderate evidence: choirs and other structured group activities, with no healthcare saving shown [32]
• Unknown: social prescribing at national scale, all reviews at high risk of bias [25]
• Weak or negative: befriending schemes, and AI companions used as a substitute for people [14] [40]

The ranking reveals an economic paradox. The most effective population-level interventions are infrastructural and cheap, yet the infrastructure is being dismantled: pubs are closing at about eight a week in England and Wales [36], youth services have lost three-quarters of their real funding [37], and in the United States 21% of people have no access to local gathering places [38]. At the same time, money is flowing towards workforce models and digital products that act on individuals one at a time. This is partly because individual services are easier to procure, brand and count. A link worker produces referrals; a hotline produces calls; an app produces downloads. A library, a salon or a pub produces something harder to measure and easier to cut: the chance encounter that becomes a habit.

The WHO Commission has called for social connection to be integrated into all policies — from digital access to health, education, and employment, in the words of its co-chair Chido Mpemba [1]. The evidence reviewed here suggests what that would mean in practice. Planning decisions that close the last community hall in a neighbourhood are loneliness decisions. Cuts to youth services are loneliness decisions [37]. Commuting patterns and remote-work policies are loneliness decisions, given the gap between fully remote and on-site workers [63]. Left unaddressed, loneliness will continue to cost society billions in health care, education, and employment, the WHO’s director-general warned [1]. Most of those billions will be determined by decisions no one labels as loneliness policy.

For health systems, the lesson is to match the tool to the person. Screening for chronic loneliness and offering structured psychological support targets the people most likely to benefit, as the meta-analytic evidence on baseline severity indicates [15]. Social prescribing may well be a useful route to that support and to community activity, but it needs to be evaluated properly before more is invested in it, ideally through staggered rollouts that allow comparison [25]. For technology regulators, the lesson is that safety rules are necessary but not sufficient: any product marketed as a remedy for loneliness should have to show that heavy users become less lonely over time, not merely that they keep coming back [40] [42].

Loneliness is linked to more than 871,000 deaths a year [1], and the market for remedies is expanding quickly. The evidence on what works is imperfect, but it is not empty, and it is more consistent than the policy debate suggests. The interventions that work are mostly unglamorous: a salon within 350 metres [30], a choir [32], a trained volunteer who calls every week and lets the other person talk [17], a structured course that teaches a lonely person to read social signals less defensively [18]. The interventions that are mostly theatre are the ones that are easiest to announce. The gap between the two is where the next decade of loneliness spending will be won or wasted.