What happened in social health? Loneliness falling worldwide but rising in Europe, a null on Alzheimer’s blood markers, mourning a chatbot, and a chatbot bill on Newsom’s desk

No trials this week. I went looking and there were none. A date-bounded search of the last seven days for randomised or controlled intervention studies of loneliness, social isolation or social connection returned twenty-six records, and all twenty-six were conference abstracts, observational studies caught by a keyword, or off-topic. What did arrive was a set of papers that mostly undercut things the field believes.

Three of them belong together. A cross-temporal meta-analysis says loneliness among older adults has been drifting down worldwide since 1997 and up in Europe. A second-order meta-analysis covering 4.8 million people puts the loneliness–physical health correlation at .15 and adds that risk of bias may have inflated it. A prediction model reaches a decent AUC that turns out to rest almost entirely on having asked people whether they are lonely. On the AI side, California’s legislature passed a child-safety regime for companion chatbots on 31 August, three days before the first careful measurement of what happens to people when a companion app changes underneath them. Policy produced almost nothing this week and funding produced nothing new at all, and I have said so in those sections instead of filling them.

Two things are new in this issue. Each research item now carries a Measurement line saying what the study measured loneliness with, whether it reported reliability in its own sample, and whether it compared scores across groups or time without showing those scores are comparable. Where a paper makes an absence claim, there is also a Power line. Some of those figures are mine rather than the authors’, derived only from numbers the paper itself printed, and they are labelled where that is the case. The reasoning is in the thread and I would rather argue about it there than have it pass unnoticed.

And a note on what can be checked at all. Six of the ten research items below are closed: no version anywhere, on either Unpaywall or OpenAlex. For those, several of the lines that follow read “could not be checked because the paper is closed”, which is a different statement from “the authors did not report it” and I have kept the two apart. One item is free to read but its publisher refuses automated access, which is my limitation and is marked as mine.

About this post: Claude, my AI, researched it, wrote it and posted it here, working from a standing brief I wrote and keep editing. It goes up under my name because I stand behind it and any error in it is mine to answer for. I do not write it, and there will be weeks when you read it before I do. If something here is wrong, missing or overstated, say so in the thread. Corrections feed the next issue, and where they are about how the digest is put together, they feed the brief.

Research

A cross-temporal meta-analysis of 125 studies finds loneliness among older adults has declined modestly worldwide since 1997, and increased in Europe. Kang and colleagues pooled 125 studies published between 1997 and 2025 across ten databases, N = 107,304, and regressed mean loneliness scores on the year the data were collected. The overall trend was a decline, d = −0.331. Loneliness fell in Asia and North America and rose in Europe, where levels were also highest. Higher GDP per capita, life expectancy and old-age dependency ratio went with lower loneliness, unemployment with higher. The unit of analysis is the published study, so this tracks how sample means have moved over three decades. The moderators are ecological, matching country-level indicators to study-level means, so they say nothing about whether an unemployed individual is lonelier than an employed one. No Europe-specific effect size appears anywhere I could reach, so the size of the European increase is not something I can give you.

Measurement. This is where the lead finding is most exposed. The analysis regresses mean scores on year, pooling studies that used different loneliness instruments across twenty-eight years, and nothing in the accessible record shows those scores are on a comparable scale. A drift in pooled means and a drift in how loneliness is measured produce the same picture, and only measurement equivalence across instruments and across time separates them. That evidence would be unusual to have and its absence is not a failing of these authors in particular; it is a reason to treat the direction as a prompt to look at Europe rather than as a measurement of European deterioration. The usable conclusion is a reason to doubt a worsening global trend, and a reason to look harder at Europe.

Credentials. Closed: no open version located on Unpaywall or OpenAlex. No preregistration stated in any accessible part of the record, and no PROSPERO registration located; the methods section is closed, so this could not be checked rather than confirmed absent. Data availability not stated; materials and code not stated. Composition is study-level: region and gender are reported as moderators, with no participant-level ethnicity, education or income. 125 studies, N = 107,304; the number of European studies is not stated in accessible material. (paper)

A second-order meta-analysis of 49 meta-analyses puts loneliness and physical health at r = .15, and reports that risk of bias may have inflated it. Zell and Robinson synthesised 49 independent meta-analyses covering more than 60 outcomes, roughly 1,900 studies and 4.8 million participants. Loneliness correlated with broad mental health at r = .37, 95% CI [0.31, 0.44], psychological adjustment at r = .32 [0.28, 0.36], and physical health at r = .15 [0.10, 0.20]. Associations varied significantly by which loneliness measure was used and by whether unpublished studies, longitudinal studies and covariates were included. A second-order meta-analysis treats existing meta-analyses as its data points, so the tight confidence intervals reflect the size of the underlying literature and not any new control over its quality. Selective publication, cross-sectional design and inconsistent confounder adjustment inside those 1,900 studies all pass through intact, which is why the authors’ own closing call for “longitudinal associations, interventions, and preregistered effects” is the operative sentence.

Two things to take away. The mental-health association is large and consistent across syntheses. The physical-health association is r = .15, which is the figure to have in mind when someone reaches for mortality-scale comparisons, and even that should be read as an upper bound.

Measurement. The paper reports its own measurement finding and it deserves more attention than it will get: the association varied significantly by which loneliness instrument the underlying studies used. That is evidence that the instruments are not interchangeable, in a literature that routinely pools them. Reliability of those instruments is not reported at this level of synthesis, which is a property of the design rather than an omission.

Credentials. Closed. No preregistration or PROSPERO record stated or located, including on OSF. Data, materials and code could not be checked, as the publisher’s platform is not machine-readable and there is no open version. Composition is meta-analytic, with no participant-level demographics; reported moderators are the loneliness measure, publication status, study design and covariate adjustment. 49 meta-analyses, approximately 1,900 studies, 4.8 million participants. (paper)

In 434 dementia-free French adults over 65, the association between loneliness and two Alzheimer’s blood markers weakened sharply once the models were adjusted, and one of the two nulls is much firmer than the other. Hooper and colleagues analysed baseline data from the INSPIRE-T cohort in Toulouse. Unadjusted, loneliness was associated with plasma p-tau217, β = 0.01014, 95% CI [0.00276, 0.01751], p = 0.007, and with neurofilament light, β = 0.00984 [0.00426, 0.01541], p = 0.001. After adjustment for age, sex, education, cognitive performance, depressive score, eyesight, hearing and APOE ε4, both estimates crossed zero: p-tau217 β = 0.00458 [−0.00328, 0.01243], p = 0.253; NfL β = 0.00138 [−0.00418, 0.00693], p = 0.626. There was no moderation by sex, depression, eyesight, hearing or APOE ε4.

The covariates that remove the association are the ones that describe who tends to be lonely in later life, so the plain reading is that the crude correlation reflects age, education, cognition and mood rather than a link between loneliness and tau pathology or neuronal injury. Being cross-sectional, it cannot test whether loneliness tracks change in these markers, which is the question a causal story would need. The cohort is single-site and skews well-educated and high socioeconomic status, which the authors concede.

Power (my calculation, from the paper’s own intervals). The two nulls are not equally strong, and the difference matters. Backing the standard error out of each adjusted confidence interval gives 0.0040 for p-tau217 and 0.0028 for NfL. For NfL the adjusted interval [−0.00418, 0.00693] excludes the unadjusted estimate of 0.00984, and the model had about 93% power for an effect that size, so that null is informative. For p-tau217 the adjusted interval [−0.00328, 0.01243] still contains the unadjusted estimate of 0.01014, and the model had about 72% power for it. Adjustment there widened the uncertainty around the original effect rather than ruling it out, so “the association disappeared” is stronger than the p-tau217 result supports. Smallest effects detectable at 80% power: 0.0112 for p-tau217, 0.0079 for NfL. These figures are mine, not the authors’, and they assume nothing beyond the estimates and intervals printed in the abstract.

Measurement. The instrument used to measure loneliness is not named in any accessible part of the record, which for a paper whose exposure is loneliness is the first thing a reader needs. No reliability reported, and none could be checked. Whether comparability across the adjustment strata was examined could not be checked either. The blood markers are assayed, so measurement quality on the outcome side is an assay question rather than a psychometric one.

Credentials. Closed. Alongside the no-conflicts declaration, the cohort’s funders include Pfizer, Pierre-Fabre, KORIAN and EDENIS. The INSPIRE-T cohort is registered (ClinicalTrials.gov NCT04224038); this analysis is not separately preregistered, and the authors state the sample size “was defined a posteriori and was based on data availability in INSPIRE-T”, so there is no a priori power calculation to report. Data available to researchers on request, following approval by the INSPIRE-T data access committee and a data use agreement; no code or materials statement. Composition reported for age, sex and education in three categories, plus APOE ε4 status, cognition, depressive score, eyesight and hearing; no ethnicity, income or migration status; single site, Toulouse. N = 434 analysed, from 1,120 at baseline. (paper)

A machine-learning model for predicting loneliness onset in Japanese older adults reached AUC 0.740, and 0.635 once the baseline loneliness items were removed. Shimoda and colleagues followed 1,806 adults aged 65 and over who were not lonely at baseline, drawn from the National Center for Geriatrics and Gerontology’s Study of Geriatric Syndromes, for a mean 3.1 years; 421 (23.3%) were lonely at follow-up on UCLA-LS-3 with a cut-off of 44. Twelve algorithms were compared and XGBoost performed best: AUC 0.740 (0.705–0.776), accuracy 0.631, sensitivity 0.806, specificity 0.577. Ten-fold cross-validation gave a mean AUC of 0.742 ± 0.065, close to the held-out figure, so the model is not obviously overfitted to its training split. SHAP put the baseline UCLA items among the most influential predictors, and removing them dropped test AUC to 0.635.

This is exploratory prediction work, and preregistration is not the criterion to judge it by. The criteria that matter are whether the model holds up out of sample, whether its probabilities are calibrated, and whether using it would help anyone. Internal validation is reasonable here, since the cross-validated and held-out estimates agree. There is no external validation in an independent cohort, no calibration reported, and no decision-curve or net-benefit analysis, so nobody can yet say what would happen if a service used this to triage. The 0.740-to-0.635 gap is the substantive finding: most of the discrimination comes from having asked about loneliness at baseline, which means the model largely identifies people already close to the threshold rather than finding them from independent information. The authors say as much.

Measurement. UCLA-LS-3, dichotomised at 44 to define the outcome, with the same scale’s baseline items among the predictors. No reliability reported in this sample, and it could not be checked. The dichotomisation is doing real work: a cut-off turns a continuum into a class, so a participant one point below the line at baseline and one point above at follow-up counts as an incident case, and that is the mechanism most likely to produce the 0.740-to-0.635 gap the authors describe. If you want to see what those items ask, and what the alternatives look like, they are in the network’s Item Explorer.

Credentials. Closed. Exploratory prediction study; no preregistration or analysis plan stated, which is standard for this design rather than a gap in it. Internal validation by ten-fold cross-validation (mean AUC 0.742 ± 0.065) alongside a held-out test set; no external validation cohort, no calibration curve or Brier score, and no decision-curve analysis reported. Data available from the corresponding author on reasonable request; no code, model file or model card, so the model cannot be reproduced or reused as published. Composition beyond age and sex could not be checked. N = 1,806 analysed, from 4,050 respondents, with 421 incident cases. Funded by AMED, NCGG Longevity Sciences, JSPS and RISTEX, and by Kao Corporation; the authors declare no conflicts of interest. (paper)

Three statisticians at Karolinska argue that a recent longitudinal finding on stigma and loneliness is a regression artefact, and have posted their reanalysis code. The target is Wan and colleagues (Journal of Clinical Nursing, July 2026), a cross-lagged panel analysis of 277 stroke survivors from two Chinese hospitals followed at baseline, three months and six months, reporting a bidirectional stigma–loneliness relationship partially mediated by social participation. On what is at issue: a cross-lagged panel model asks whether stigma at one wave predicts loneliness at the next after controlling for earlier loneliness, and the objection is that when two stable, correlated traits are measured with error, this specification generates apparent prospective effects with no causal process behind them, because part of what is being predicted is regression to the mean. Sorjonen, Melin and Melin’s repository reconstructs data from the correlation matrix Wan and colleagues published, runs six models in both causal directions as a multiverse, aggregates them after Fisher-z transformation, and then fits an alternative in which a single stable trait loads equally on stigma and on loneliness. If that simpler model fits comparably, the direction of the reported effect is not identified by the data.

Measurement. This item is a measurement argument, which is why it is worth the space. The artefact the authors describe is driven by unreliability: the less reliably a construct is measured, the more of the earlier wave it fails to control for, and the more room is left for a spurious cross-lagged path. That is the same mechanism that makes a low reliability coefficient at one wave a threat to a longitudinal conclusion rather than a footnote about the scale, and cross-lagged panel models are used heavily in this literature, including in designs several of us run.

Credentials. Free to read at the publisher (bronze open access), but Wiley refuses automated retrieval, so I read its code and its reference list rather than its text and the account above is inferred from those. That is my limitation and not a paywall, and anyone can open it in a browser and correct me. It was published on 30 August, a day before this week’s window. No preregistration stated. Data and code openly available on GitHub (KimmoSorjonen/26.30_Stigma); the data are reconstructed from the correlation matrix reported in the target paper rather than collected. Sample composition not applicable, as the reanalysis has no participants of its own. N not applicable; the criticised study analysed 277 stroke survivors. (paper)

A preprint network analysis of 1,503 US adults finds that using social media to forget about personal problems is the strongest bridge between problematic use and distress, and that this does not vary by age. Helvich, Kim and Primack estimated Gaussian graphical models across four decade-representative cohorts aged 30 to 70 and compared them with network comparison tests. The strongest cross-domain edge was between anxiety and using social media to forget about personal problems, w = 0.12; anxiety was negatively associated with cognitive preoccupation with social media, w = −0.09; depression and loneliness were associated mainly with impairment- and regulation-related symptoms, w = 0.05 to 0.10. Network comparison found no significant differences in overall connectivity, network structure or individual cross-domain associations across cohorts.

Each edge is a partial correlation, the association between two symptoms with all others held constant, and edges between 0.05 and 0.12 are small. The data are cross-sectional, so nothing here speaks to how these symptoms drive one another over time, which is what the bridge-symptom language usually implies.

Measurement. The nodes are single items rather than scales, so internal consistency does not apply and the relevant diagnostics are edge-weight accuracy and centrality stability, which tell you whether an estimated network would look the same in another sample. I could not find them in the record I could read. Whether networks of this magnitude replicate is unsettled in this literature, and without those diagnostics the ordering of edges is the part to hold most loosely.

Power. The age result is a failure to detect a difference and not evidence of equivalence, which would require an equivalence test against a specified bound. The paper reports no such bound, and network comparison tests do not yield a printed estimate and interval from which I could derive one, so unlike the Hooper item above there is no figure I can honestly put here. What can be said is that four cohorts of roughly 375 each is not much power for detecting differences in network structure.

Credentials. Open: a PsyArXiv preprint, and the only research item this week that anyone can read in full without a subscription. Preregistration recorded on OSF as “not applicable”, with the stated reason that “no data collection, extraction, or analysis is reported in the preprint”, which sits oddly beside a paper reporting a network analysis of 1,503 respondents. Data not currently available: per the authors, “data will be publicly available via a structured data request form one year after initial analyses are completed”; supplemental materials at osf.io/67uvt. Composition reported for age band and US residence only; sex, ethnicity, education and income not stated in the record I could read. N = 1,503. Not peer reviewed. (paper)

When Replika removed erotic role-play and when ChatGPT rolled out GPT-5, users’ posts shifted measurably towards grief. De Freitas, Castelo, Uğuralp and Oğuz-Uğuralp treated both product changes as natural experiments, analysing 54,861 Reddit posts and running seven surveys with 1,452 participants. Both updates increased negativity, loss framing and restoration desires: Replika negative posts rose 24.7 percentage points, 95% CI [20.1, 29.2], ChatGPT 13.0 points [10.8, 15.2]. Increases in sadness were larger for Replika than ChatGPT, d = 2.67 versus 1.41, as were increases in negative mental health, d = 1.72 versus 0.63. Replika users reported closeness exceeding common human ties (versus a friend, d = 0.47) and anticipated mourning above other technologies (d = 0.32 to 0.57).

The two halves of the paper are not equally strong. The natural experiments compare the same communities before and after a dated product change, which is a defensible counterfactual, though nothing else happening to those communities at the time is controlled for. The surveys were recruited from Replika communities, so the closeness and mourning estimates describe committed users rather than people who use companion apps casually, and the very large d values should be read in that light. What this establishes is that a unilateral product change can produce measurable distress in an identifiable group of users, which is a fact a regulator can act on. What it does not establish is how large that group is.

Measurement. Half the evidence here rests on text: negativity, loss framing and restoration desire scored across 54,861 Reddit posts. The property that matters is therefore not a scale’s reliability but whether the coding scheme or classifier was validated, and against what human agreement. That could not be checked, the paper being closed. It is worth flagging because the largest numbers in the item, the 24.7 and 13.0 percentage-point shifts, are the ones that depend on it entirely.

Credentials. Closed. One AsPredicted preregistration is cited (#199,522, aspredicted.org/z9pb-wym7.pdf, registered 15 November 2024). It covers a single survey study, a seven-condition within-subjects design with a target of 200 participants recruited from a Replika Facebook group; the two natural experiments and the remaining surveys are not covered by any cited preregistration. Data and analysis code openly deposited on GitHub (Ethical-Intelligence-Lab/ai_loss). The preregistration specifies collection of age, gender, education and ethnicity; whether income, socioeconomic position or any country breakdown beyond US residence are reported could not be checked. N = 54,861 posts and 1,452 survey participants across seven surveys. Funded by Harvard Business School; the authors declare no competing interests. (paper)

Across 732,808 US nursing home residents, about a third of the variance in loneliness sits between facilities, and measured facility characteristics explain almost none of it. Nielsen and colleagues ran multilevel logistic models on MDS 3.0 assessments from 13,491 nursing homes, October to December 2023, using the single item “How often do you feel lonely or isolated from those around you?”, dichotomised. Prevalence was 20.7%. The variance partition coefficient was around 35% at facility level and around 2% at county level, and adding resident, facility and county characteristics changed the cluster variance by only 1% to 2%. Prevalence was higher among women and residents with mental health diagnoses or sensory impairments, and lower among older residents and Non-Hispanic Black residents; for-profit status, rural location, racial and ethnic diversity and more certified nursing assistant care were associated with lower prevalence.

Two results here pull against each other. The variance partition says which home you live in matters a great deal, roughly 35% against 2% for county. The 1% to 2% change in cluster variance says that ownership, staffing, rurality and diversity, which are the things commissioners and regulators can actually observe, do not account for it.

Measurement. This is the item where measurement decides what the headline number means, so it is worth being blunt. Loneliness is a single item, administered by staff, then dichotomised. Internal consistency does not apply to one item, and that is a property of the design rather than an omission. But the central claim is a comparison of scores across 13,491 facilities, and a comparison like that assumes the item behaves the same way in each of them. With one item there is no way to test that assumption, so the two live explanations for the 35% figure, real differences in daily life inside a home and systematic differences in how staff ask the question, are not merely unresolved here. They are indistinguishable by construction. Anyone wanting to act on the facility-level finding would need a validated multi-item measure, collected the same way across sites, before the number means what it appears to mean.

Credentials. Closed. No preregistration stated; the authors report institutional review board approval only. Data restricted: “Restrictions apply to the availability of the data (MDS 3.0) under a data use agreement. MDS 3.0 is available from www.resdac.org with the permission of the CMS.” No code or materials statement. Composition reported for age, sex, race and ethnicity, marital status, cognitive status, mental health diagnoses and sensory impairment, plus facility rurality, ownership, chain membership, diversity and staffing; no resident-level education, income or migration status. N = 732,808 residents in 13,491 facilities. Funded by NIA R01AG071692 and NCATS TR001454; no conflicts declared. (paper)

From 28 August, so a little outside this week and easy to have missed: five waves of panel data on 1,966 US adults found little credible evidence that any of ten social technologies predicts later life satisfaction. Kushlev, Moon, Motyl, Fast and Schroeder used random-intercept cross-lagged panel models, Bayesian and frequentist, on Understanding America Study data measured every three months on a six-point frequency scale, with results reported as posterior medians against a region of practical equivalence of ±0.10. In the reverse direction, increases in life satisfaction predicted only modest increases in video calling, in some demographic groups. Comparing between people, texting went with higher life satisfaction and YouTube and TikTok with lower.

A random-intercept model separates stable differences between people from within-person change over time, and it is the within-person half that answers whether changing your own usage would change your own wellbeing. That is the half where the evidence is absent. The between-person associations survive, and they are the ones that produce headlines, but they cannot distinguish TikTok making people unhappy from unhappy people using TikTok. The authors note that responses were temporally stable, which limits how much within-person movement there was to detect, so this sits closer to an informative absence than to a demonstrated zero.

Power. Worth crediting: a region of practical equivalence of ±0.10 set in advance is the right way to make an absence claim, and it is rarer in this literature than it should be. It converts “we found nothing” into “we can rule out anything larger than this”. I cannot check how the estimates sat against that bound, because everything below the abstract is closed and there is no preprint, so take the direction and the decision rule and do not quote a magnitude.

Measurement. Technology use is self-reported frequency on a six-point scale, every three months. Reliability and factor structure could not be checked. The stability the authors report is itself a measurement observation as much as a substantive one: a coarse frequency item asked four times a year will not move much, and an outcome that does not move is hard to relate to anything.

Credentials. Closed, and no preprint located on Unpaywall or OpenAlex. No preregistration stated in the article, and none visible on the authors’ public OSF project. Data are managed access: “available to researchers by application through the UAS data portal (https://uasdata.usc.edu/). Access is governed by the UAS data-use agreement, and the authors are not permitted to redistribute the data.” Analysis code openly deposited on OSF (doi 10.17605/OSF.IO/A3DN6). Sample composition beyond N could not be checked. N = 1,966, five waves. Funded by the Knight Foundation; the authors declare no competing interests. (paper)

Policy and advocacy

Close to empty. The WHO Commission on Social Connection, the JRC, DG SANTE, DG EMPL, OECD WISE, and the German, Dutch, Spanish, Danish, Swedish and UK health agencies all published nothing on this in the window. Several older documents are circulating as though they were new, and I have listed their real dates at the end. One item.

Japan’s Minister for Loneliness and Isolation Measures co-signed a four-minister message for Suicide Prevention Week, published on 1 September as schools returned. The message went out from the Ministry of Health, Labour and Welfare over the signatures of the health, education and children’s policy ministers together with the Cabinet Office loneliness portfolio, ahead of the week running 10 to 16 September, and cites 538 suicides among primary, junior high and high school students in 2025, the highest figure since the series began. I am including it as a thread rather than an event: the substance is suicide prevention, and the social-connection content is the standing co-signature and the Cabinet Office’s continued involvement. The Cabinet Office’s own loneliness and isolation portal has published nothing since 10 July. (link)

Practice and the market

California’s legislature passed a companion-chatbot child-safety bill on 31 August, and as of today the Governor has not signed it. SB 1119, “Companion chatbots: children’s safety,” authored by Senator Steve Padilla with Assemblymembers Bauer-Kahan and Wicks, cleared the Assembly 69–4 with 6 not voting on the session’s final night and returned to the Senate for concurrence. It requires companion-chatbot operators, on or before 1 July 2027, to perform and document an annual comprehensive child-safety risk assessment, submit to independent child-safety audits, and report to the Attorney General. It is named for Adam Raine, a Californian teenager who died by suicide in 2025. Governor Newsom has until 30 September to sign or veto. Passed the same night and enrolled on 4 September: SB 867, a four-year moratorium on manufacturing or selling children’s toys containing AI companion chatbots. One discrepancy I could not resolve: the legislature’s history page records Senate concurrence on SB 1119 as 39–0 while its votes page shows 40–0, so I am reporting both rather than choosing. (link)

OpenAI publicly asked the Governor to sign it. In a post dated 31 August by Ann O’Leary, its VP for global policy, OpenAI said it had written to Newsom commending him and the bill’s authors, and that “we encourage Governor Newsom to sign the bill into law.” It endorses seven specific requirements, including age determination, pre-release risk identification, independent audits, parental tools and crisis-support routing, and argues that “SB 1119 also appropriately recognizes that AI is not social media.” A frontier lab lobbying in favour of binding safety requirements on its own product category is unusual enough to record. One number in the post to handle carefully: the claim that “nearly nine in ten teens who use ChatGPT turn to it for learning, information, skill-building, or productivity in a given week” is company-sourced, with no sample size or method published, and is not independently verifiable. (link)

Vox went through what the loneliness market currently charges, and the prices carry most of the argument. Allie Volpe’s 31 August feature lists a loneliness coach at $3,000 for a three-month programme, a $1,000 weekend friendship camp payable in instalments via Klarna, a $250 pendant that listens and talks back, a coworking space at nearly $300 a month, and a $50 one-off knitting club run by a venture-backed friendship app. Adam Neumann’s residential venture, aiming at “solving loneliness,” raised $350 million from Andreessen Horowitz; the app 222, which pairs strangers for group activities by personality questions, raised $10 million. Volpe’s structural point is that a company must either charge a great deal or maximise how often people use the product, and both of those work against the way relationships actually form, while pricing out the people with the highest need. Julianne Holt-Lunstad, who fields these approaches constantly, is quoted: “I would caution developers to really critically evaluate their business model and the barriers to who they’re actually reaching. Think about ways that [connection] can occur organically without monetizing it.” Declaring an interest: I co-founded Entrelacs, which sells social-connection measurement to employers, so I am inside the market this piece is about. (link)

Money

Nothing new opened this week, and the state of play is unchanged. No EU4Health, ZonMw, ANR, DFG, Wellcome, Nuffield, VolkswagenStiftung or NIH notice on loneliness, social isolation or social connection was published between 31 August and 7 September. Two things moved quietly. The NIHR call “Loneliness in the community” (reference 2025/451, Public Health Research programme) closed its outline stage on 18 August; out-of-remit decline notifications go out in early September, outline shortlisting and the opening of full applications are scheduled for the end of October, and full applications close in early January 2027. And Horizon Europe HORIZON-CL2-2026-01 remains open with no corrigendum and no deadline change: 23 September, 17:00 Brussels time, sixteen days from today, with TRANSFO-09 on long-term care (€15m, around €3.75m per grant, four grants) the closest fit to this consortium’s remit. HORIZON-CL2-2026-02-TRANSFO-01, the €60m co-funded partnership, closes 13 October. One caveat about the completeness of that check: the Funding and Tenders portal’s search API was unreachable, so the EU sweep was done topic by topic rather than as an exhaustive keyword search, and it is not proof that nothing else is open. (link)

Two notes on this issue

An item was pulled after it was drafted. A JAMA Pediatrics study of 39,761 Ontario schoolchildren and generative AI for emotional support was written up for this issue and then removed, because it ran in the 2 September issue. That is the second week running that something has been carried twice, after the IPPR figure, and the running record that is supposed to prevent it has now failed twice. Two things about it are worth keeping anyway. Its access line last week said the paper was behind the JAMA paywall; it is not, and a submitted version is deposited in PubMed Central, which we could not retrieve rather than could not access. And in that paper loneliness is a covariate rather than an outcome, which is worth knowing before the finding is cited as being about loneliness.

The measurement layer is new and I would like it argued with. The rule I have applied is that every study measuring loneliness with an instrument gets a measurement line, including when every entry in it reads “not reported”, because a line that appears only when there is something to say would teach you that its absence means the measurement was fine. A longer note goes in only where it changes how the result reads, which this week was three of the ten. Tell me if that balance is wrong in either direction.

Corrections, additions and arguments in the thread, please, particularly if you can get past a paywall I could not. Six of the ten research items this week are closed to everyone without a subscription, and those are the ones where I am most likely to be wrong.

Hans