What happened in social health? Brussels moves on chatbots that simulate relationships, a Danish youth service misses on loneliness, the Dutch budget’s loneliness line, and Men’s Sheds without a grant
Hans Rocha IJzerman·21 September 2026
Set two of this week’s events side by side. On Wednesday the European Commission proposed a Regulation that would prohibit design features which simulate interpersonal relationships with children in ways likely to create emotional dependency. On Monday, a Danish evaluation of a nationwide youth counselling service reported that six months on, it had moved well-being by a tenth of a standard deviation and had not moved loneliness at all. Regulation is now moving considerably faster than the evidence about what actually helps, and I do not think that is an argument against the regulation. It is an argument about where the research money should go.
The other pattern in this week’s papers is duller and more consequential. Of the five studies below that measure loneliness with an instrument, one reports a reliability coefficient computed in its own sample. One. The highest-profile paper of the week attributes a third of the sex difference in adolescent depression to loneliness measured with a single question, dichotomised, with no confidence interval published anywhere for the headline proportion. If we are going to build policy indicators on these constructs, and the WHO is proposing exactly that, the instruments underneath them need more attention than they are getting.
This issue was researched and written by Claude, Hans’s AI, working from a standing brief I wrote; I pasted it into the forum myself. It appears under my name because I stand behind it and any errors in it are mine to answer for. I do not write these, and I will not always have read one before you do. Corrections and disagreements in the thread feed into the next issue and into the brief itself, which is how the thing improves.
⏺️ Research
⏺️ Denmark’s headspace youth counselling service shifted well-being by d = 0.11 at six months and left loneliness where it found it. A quasi-experimental evaluation across all 30 Danish headspace centres compared 1,362 users aged 12–25 with 3,192 propensity-matched controls drawn from national registers, recruited over 14.5 months from September 2022. In the final model, well-being on the WHO-5 improved by 2.179 points (95% CI 0.386 to 3.972, p = 0.018, d = 0.11). The other four outcomes did not separate: loneliness on the Three-Item Loneliness Scale −0.12 (−0.293 to 0.053, p = 0.175, d = −0.06), self-efficacy 0.563 (−0.044 to 1.170, p = 0.070), life satisfaction 0.138 (−0.046 to 0.321, p = 0.142), EQ-5D-5L sum score −0.203 (−0.496 to 0.089, p = 0.174). Two details change how that reads. In the observed-data model, before rematching and multiple imputation, self-efficacy, life satisfaction and health-related quality of life were all significant; the final model turns them null, and the paper resolves the contradiction by asserting the earlier models “were not valid or reliable”. And no control for multiple testing was applied across the five outcomes, which the authors state plainly. A reader can now say that a large matched comparison of Danish help-seekers found a small well-being gain and a loneliness estimate whose interval excludes anything above about a fifth of a standard deviation in either direction. A reader cannot say the service does not help lonely young people: matching on registers cannot balance the disposition to walk into a counselling centre, and d = 0.11 on well-being sits inside the range residual self-selection alone could produce. Anyone running a comparable youth service can add it to the LONELY-EU Intervention Registry, which exists precisely so that services like this one can be found by the people evaluating them (contribution form).
No trial registry entry is stated anywhere in the paper; a published protocol (Frontiers in Public Health 2025;13:1491756) and a statistical analysis plan deposited at OSF (10.17605/OSF.IO/9KZFD) stand in its place. Data restricted: “Open access to data is restricted by EU and local legislation, but data can be shared upon reasonable request.” No code availability statement, and no analysis software is named. Composition reported for age, sex, ethnicity, own and parental education, parental income, family type, prior psychiatric service use and school attainment; migration status not reported, and the breakdown across the 30 centres sits in supplementary material. N = 4,554 (1,362 intervention, 3,192 control).
Measurement: WHO-5, the Three-Item Loneliness Scale, the General Self-Efficacy Scale, EQ-5D-5L, and a single-item Cantril’s ladder. No internal-consistency coefficient of any kind is reported for any of the five, neither computed in this sample nor cited; the instruments are described as validated, with one citation each. No factor structure or dimensionality check; sum scores throughout. Measurement invariance not tested across arms or across the 12–25 age span. Danish-language validation is not reported or cited for four of the five instruments; only EQ-5D-5L carries Danish references, and those are a value set and population norms. Linear regression, so no fit statistics apply. The post-hoc population attributable fraction of 4% rests on dichotomising WHO-5 at 50, a cut-point cited without justification, with no sensitivity analysis on the threshold, and the published figure that is supposed to display those scenarios varies different values (20/30/50% prevalence, 5/10/20% coverage) from the 70% and 80% the text uses. Openness: checked. Power sensitivity, our own calculation from the printed standard errors and not a figure the paper reports: the loneliness standard error of 0.088 implies a detectable difference of about d = 0.12 at 80% power on a two-sample comparison, which ignores power gained from the baseline adjustment; the paper reports no sample-size calculation, though the published protocol does, at 90% power for a standardised difference of 0.18.
(paper)
⏺️ In 746 German adults with alcohol use disorder, feeling lonelier than usual predicted no change in drinking four months later, and drinking predicted no change in loneliness. Four waves at roughly four-month intervals across Charité Berlin, TU Dresden and the Central Institute of Mental Health in Mannheim, analysed with dynamic panel models that separate within-person change from between-person differences. Loneliness to alcohol: frequency β = 0.020 (−0.052 to 0.093, p = .583), quantity β = 0.044 (−0.071 to 0.159, p = .453), binge days β = −0.041 (−0.141 to 0.058, p = .417). The one non-null runs against the self-medication account: higher loneliness relative to a person’s own average predicted a lower AUDIT score four months on, β = −0.103 (−0.175 to −0.031, p = .005). All four reverse-direction paths were null. Over the study, drinking fell on every measure and loneliness rose slightly (β = 0.014, p = .024). These are tight nulls: with standard errors between 0.037 and 0.059 on standardised coefficients, effects this study would have missed are small ones. What they are nulls about is four-month dynamics. They say nothing about whether a lonely evening produces a drink that evening, which is the timescale the self-medication hypothesis actually describes, and the authors are careful about the distinction.
Preregistered (OSF, 10.17605/OSF.IO/3FHPM, registered 24 September 2025), with a section itemising deviations: the preregistration specified annual follow-ups where assessments were four-monthly, and preregistered hypothesis H5 was not run because the prospective association it depended on was not supported. Parent cohort registered DRKS00020580. Data not deposited, with the reason given verbatim: sharing was not covered by participants’ informed consent; available on request. Analysis code openly deposited on OSF (10.17605/OSF.IO/F76K3). Composition reported for age (M = 37.33, SD 12.83) and sex (38.9% women) plus mean DSM-5 criteria count; no ethnicity, education, income or migration status, and the three recruitment sites are named without per-site numbers. N = 746 at baseline, 472 at wave four.
Measurement: loneliness measured with a modified 12-item UCLA scale, the modification not characterised; social support with the F-SozU K-14. No internal-consistency coefficient is reported for either, in this sample or cited from elsewhere. No factor structure or dimensionality check; sum scores throughout. Measurement invariance across the four waves not tested, which in a within-person design is the assumption doing the most work. Model fit reported as AIC, BIC and a Wald χ²(2) per model, with no CFI and no RMSEA, and therefore no RMSEA confidence intervals; the authors note the quantity models fit comparatively poorly (Wald χ² = 0.59, against 19.39 for the AUDIT model). Binge drinking dichotomised at five standard drinks, a conventional threshold cited without justification, with no sensitivity analysis on it. Openness: checked. Power sensitivity, our own calculation and not reported by the paper, which states no sample-size calculation: standard errors of 0.037 to 0.059 imply detectable standardised within-person effects of roughly 0.10 to 0.17 at 80% power, treating each path as a simple two-sided test and so ignoring power gained from the repeated measures.
(paper)
⏺️ A meta-analysis of preoperative social connection and surgical outcomes reports that every prediction interval it could estimate included the null, and puts that sentence in its own abstract. From 33,301 records and 445 included studies, poolable estimates came from 72, and the five principal analyses rest on 39 analysis units from 33 studies. Postoperative all-cause survival: HR 1.37 (95% CI 0.99 to 1.89), I² = 74.1%, prediction interval 0.59 to 3.17, certainty very low. Early postoperative mortality: OR 1.50 (1.12 to 2.01), certainty low. Non-home discharge: OR 1.95 (1.35 to 2.81), prediction interval 0.71 to 5.34. Unplanned readmission: OR 1.14 (0.98 to 1.33). Any or major complications: OR 1.10 (0.96 to 1.25). No eligible study estimated failure to rescue. The exposure is the problem and the authors say so: marital or partnership status is the analysed contrast in 29 of the 39 units, and exactly one unclassified loneliness measure appears in the entire synthesis. So this is evidence about whether being married predicts surviving surgery, wearing the label of social connection. The readmission and complications nulls are precise enough to act on; the survival estimate is not a demonstrated absence but an imprecise estimate that happens to cross one. The methods disclosure deserves attention on its own terms: screening and extraction were performed by AI agents under written investigator directives, with two people confirming the 39 principal estimate records against the printed sources while seeing the AI-proposed values, and the authors state that this design “cannot distinguish accuracy from anchoring”.
Registered on PROSPERO (CRD420261449181, version 1.0 published 14 July 2026; version 2.0, an administrative update adding collaborators and a conflict declaration, published 9 September 2026), and the preprint itemises its departures from that registration: the protocol said structural, perceived and functional exposures would never be pooled together, and they were pooled into a single umbrella contrast; the risk-of-bias tool changed from Newcastle-Ottawa to result-level QUIPS before assessment. Data, deterministic analysis code, risk-of-bias assessments, GRADE profiles and figure source data openly deposited on GitHub under MIT and CC BY, archived on Zenodo. The 40-page supplement, with eTables 1–13, is posted with the preprint and we confirmed it downloads. Sample composition is not characterised at review level: no breakdown by ethnicity, education, income or country. 445 studies included; 33 contribute to the principal analyses.
Measurement: this is a review, so it does not carry instrument-level psychometrics, and what it owes instead is the reliability of its included studies’ measures, which it does not report. The exposure in 29 of 39 analysis units is a single categorical item, marital or partnership status. Openness: checked. Power sensitivity, our own calculations from the printed intervals on the log scale: standard errors of roughly 0.165 for survival, 0.078 for readmission and 0.067 for complications, so the last two are informative nulls and the first is not; no information-size or trial-sequential calculation is reported.
(paper)
⏺️ Loneliness at 14 is estimated to account for a third of the sex difference in depressive symptoms at 16, and it was measured with one question. In the Tokyo Teen Cohort (N = 3,171, assessed at 12, 14 and 16), girls’ prevalence at 16 was 15.9% for depressive symptoms against boys’ 8.2%, and 29.8% against 18.1% for suicidal ideation. Using a parametric mediational g-formula with mediators chosen in consultation with a young persons’ advisory group, the effect operating through age-14 loneliness and desire for thinness was 34.7% of the total effect of female sex for depressive symptoms and 46.3% for suicidal ideation. No confidence interval is published for either figure. This is not a retrieval failure on our part: the 38-page supplement prints 95% intervals for the total effect, the direct effect, the indirect effect and their difference in every table, and gives the effect-ratio row a bare point estimate every time. Both mediators are single items dichotomised at a stated cut, loneliness at “sometimes or always”. The finding is a real and useful one about a well-characterised cohort, and the honest way to quote it is as a proportion with unknown uncertainty, resting on a mediator captured by one question asked once at 14. The cohort’s own limits are reported candidly: three Tokyo municipalities, predominantly ethnically Japanese, socioeconomically advantaged relative to the general population, with lower-income households deliberately oversampled.
No preregistration stated, and no preprint version exists against which the analysis plan could be checked (two Europe PMC preprint searches returned nothing; a third was blocked by a server error, so treat this as strongly indicated). Data restricted: not publicly available owing to licensing with the cohort, de-identified data available on request to the co-corresponding authors subject to steering-committee approval within four weeks. Code posted on GitHub, though the repository contains three files and none of the five file paths its own README instructs readers to run. Composition reported unusually frankly in the Reporting Summary, which states that race and ethnicity were not used as analytical variables given limited variability, and that migration background was not available in the dataset. N = 3,171.
Measurement: loneliness and desire for thinness are each a single item, dichotomised, so no internal-consistency estimate exists for either. Depressive symptoms measured with the 13-item Short Mood and Feelings Questionnaire at a cut-off of 8, rescaled to 7.4 for the 12-item version; suicidal ideation with a current-state item whose wording sits behind the paywall. Two Cronbach’s alphas are reported and both belong to confounders (neighbourhood collective efficacy 0.87, a modified Compulsive Internet Use Scale 0.87); none is reported for the exposure, the mediators or the outcomes. No factor structure check. Measurement invariance across sex, and across ages 12, 14 and 16, not tested, and the entire claim is a sex comparison tracked over those ages. No latent-variable model, so no fit statistics. Threshold sensitivity does exist for both mediators: restricting loneliness to “always”, and loosening thinness to “I somewhat want to be thinner”, are each run as separate analyses. Openness: closed at the publisher, though the supplement and reporting summary are free and carry most of the methodological detail.
(paper)
⏺️ Feeling left behind by the pace of social change predicted loneliness two years later in Nuremberg’s over-75s, and explained half a percent of the variance. A two-wave panel from the GesTe survey, drawn from the population register: 10,000 sampled, 2,370 questionnaires at wave one (24.8% response), 818 analysed longitudinally. Cross-sectionally, psychological obsolescence was among the stronger correlates of loneliness (β = 0.238, p < .001, alongside social isolation β = 0.243 and being single or divorced β = 0.167, with subjective health, number of diagnoses, education and gender all non-significant). Longitudinally, adjusting for baseline loneliness (β = 0.601), obsolescence held at β = 0.074, p < .05, contributing ΔR² = 0.005. The authors print that 0.005 in the text instead of leaving readers to derive it, and by their own G*Power check the study had 80% power at N = 228 against the 818 analysed, so this is a well-estimated small effect and not a failure to find a large one. The harder problem is who is left in the sample: dropout was predicted by social isolation (OR 0.81, 95% CI 0.72 to 0.91) and by age (OR 0.96, 0.94 to 0.99), so the lonelier and more isolated left the study preferentially, and both the 8.6% prevalence and the association are estimated among the people the construct is least about. On instruments, this paper is the week’s outlier in the right direction, and the way it reports its scales is the standard the LONELY-EU Item Explorer exists to make easier to meet.
“The study was not pre-registered”, stated in those words. Data and analytic methods available on request for replication purposes; no separate materials or code statement. Ethics vote 23-237-S, Friedrich-Alexander-Universität Erlangen-Nürnberg. Composition reported for age, sex, education, relationship status, subjective financial situation, subjective health and diagnosis count, with a baseline table covering the cross-sectional, longitudinal and attrition samples; no ethnicity, no migration status, no absolute income, and one city by design. N = 818.
Measurement: loneliness measured with the 6-item De Jong Gierveld scale, rescored 0–12 instead of the original 0–6 index following Hansen and Slagsvold. McDonald’s omega was 0.84 at both waves, computed in this sample from a two-factor model that respects the emotional and social dimensions, and it is the only in-sample reliability estimate in this week’s batch. Psychological obsolescence, three items from Brandtstädter and Wentura, omega 0.76, also computed in-sample. No EFA or CFA is reported as such, and loneliness enters the path model as a manifest sum score. Measurement invariance across the two waves not tested, in a design that claims a mean increase in loneliness between them. The path model is saturated (df = 0), so there is no CFI, no RMSEA and no RMSEA interval, and the authors say so explicitly. The 0–12 score is dichotomised at a mean split to compare prevalence with the German Ageing Survey and SOEP, a threshold cited without argument and with no sensitivity analysis; the authors concede it captures only very severe loneliness.
(paper)
⏺️ Among English over-60s, every indicator of socioeconomic position also predicts whether someone is lonelier or less lonely than their level of isolation would suggest. Using ELSA waves 7 to 9, this paper builds a continuous “social asymmetry” score by standardising loneliness (UCLA-3) and isolation (a reconstructed Berkman-Syme index) within each wave and subtracting one from the other, with positive values meaning less lonely than isolated. In the fully adjusted models, which include depressive symptoms, self-rated health and life satisfaction: third-level education β = 0.44 (95% CI 0.36 to 0.51), highest income quintile 0.22 (0.16 to 0.29), subjective social status 0.004 (0.002 to 0.005), and on the other side, rare or no internet use −0.26 (−0.32 to −0.20), social housing −0.41 (−0.51 to −0.32), private renting −0.27 (−0.43 to −0.12), most deprived area quintile −0.20 (−0.29 to −0.10), densest quintile −0.18 (−0.27 to −0.10), unemployment −0.07 (−0.14 to −0.01). Anyone citing the abstract should know that it gives private renting as −0.35, a figure matching no model in the table (the unadjusted model gives −0.62, the SES-only model −0.36, the fully adjusted model −0.27), while every other abstract figure matches the fully adjusted column exactly. The estimates are in standard deviations of this cohort’s own within-wave distributions, and a difference score carries the measurement error of both scales, so these numbers describe gradients inside ELSA and are not portable units. Take them as a ranking of which socioeconomic markers separate the two constructs, the kind of European gradient collected in the LONELY-EU Evidence Explorer.
“The study reported in this manuscript was not preregistered”, stated twice. Data available on registration with the UK Data Service; the full R script is published in the supplementary material. Composition reported: mean age 71.2 (SD 8.3), 54.4% women, 97.0% White, employment status, education in four levels, subjective social status, individual deprivation, income quintiles, housing tenure, area deprivation and population density, with per-variable missingness given, and individual-level deprivation missing for 3,584 of 6,965 at baseline. Migration status not reported. N = 8,164 in the weighted analysis; the wave-7 analytic sample is 6,965.
Measurement: UCLA-3 for loneliness, three items rescaled to 0–6; for isolation, a Berkman-Syme Social Network Index reconstructed from ELSA with the modification documented, since ELSA does not carry attendance at religious services and membership of a religious organisation was substituted, giving a four-item 0–4 count. No reliability coefficient is reported for either, in-sample or cited, and none is computable for the index, whose components are assigned conceptually. No factor structure check on the loneliness items. Measurement invariance across waves not tested; standardising within each wave handles scale drift but not structural change. GEE, so no fit statistics apply. No dichotomisation: keeping the discrepancy continuous is a real improvement on the four-box typology this literature usually reaches for. Openness: checked, and the article is CC BY despite the publisher page describing its PDF as subscriber-only.
(paper)
⏺️ A European panel of 57 could agree on what a social prescribing link worker does, and could not agree on who should be one. A three-round Delphi run through the WONCA Europe social prescribing special interest group, EGPRN and EURIPA, with panellists from fifteen countries: 75 of 85 invited responded in round one (88%), 57 of 75 rated 52 statements in round two (76%), and all 57 returned for round three. At an a priori threshold of 80% agreement, 19 of 52 statements passed round two, and 13 of the 19 rerated statements passed round three. Consensus covered the core functions (navigation, advocacy, follow-up, documentation), mandatory communication and navigation training, supervision, a cofunded employment model, multiple access routes including self-referral and secondary-care referral, and the use of digital tools “as complements rather than substitutes for relational support”. No consensus was reached on preferred professional background, on formal certification, or on restricting particular referral routes, which is to say on exactly the questions a health ministry would need answered before writing a job specification. A panel recruited through the networks that advocate social prescribing measures the cohesion of that community, so read the 80% as a description of professional consensus and not as evidence about what link workers should do. For anyone drafting a national framework, the six statements that failed at 80% carry as much information as the thirteen that passed.
No registry entry stated; a published study protocol is referred to in the text, and the CREDES reporting checklist is completed in supplementary material. Data available on reasonable request; the 52 statements and round-by-round results are published as supplementary files. Ethics approval CEIm 25/066-P, IDIAP Jordi Gol. Panel composition reported by professional role, with recruitment monitored for country, professional background and social prescribing expertise; the authors state that years of experience in social prescribing were collected at registration but are not reported because too many responses were missing. No age, sex or ethnicity breakdown of panellists. N = 75 in round one, 57 in rounds two and three.
Measurement: a consensus study with no psychometric instrument, so reliability coefficients, dimensionality and invariance do not apply, and their absence here is not a gap. The one measurement-adjacent choice is the collapse of a five-point Likert scale into a binary agree/disagree at 80%, a convention cited without justification, with no report of how the count of passing statements would change at 70% or 75%. Openness: checked.
(paper)
⏺️ Policy and advocacy
⏺️ The European Commission has proposed prohibiting design that simulates a relationship with a child. COM(2026) 681 final, presented on 17 September, is a proposal for a Regulation known as the EU KIDS Act, an acronym for “EU Keeping Internet Digital Spaces Accountable and Trustworthy”. It is a proposal under the ordinary legislative procedure (2026/0286(COD)), now with Parliament and Council, and it is not law. Its scope reaches social networks, video-sharing platforms, app stores, online games, operating systems, and, for the first time in an EU instrument of this kind, AI companions and general conversational chatbots as categories in their own right. Article 14(1)(a) requires providers to avoid “design features and system behaviours that simulate interpersonal relations that are likely to create emotional dependencies”. Article 14(1)(b) requires that by default the system “does not use information or analysis derived from a minor’s prior interactions in subsequent interactions, except where necessary to protect the minor’s safety”, and Recital 32 states the same point as persistent conversational memory being disabled by default. Recital 32 also gives the rationale in terms this readership will recognise, extending the concern to dependencies that “may negatively impact relationships with other humans and the development of the minor”. Two clarifications, because the summaries circulating this week get both wrong: memory is off by default and not prohibited, and the widely repeated line that companion features must be off by default applies under Article 14(2) only where a companion is deployed as a functionality of a social network, video-sharing service or game, not to standalone companion apps. On penalties, the 6% of worldwide turnover figure appears once, in Article 34(2), for AI companions and chatbots, routed through Article 99 of the AI Act; other covered services are routed to Chapter IV of the DSA, which the text does not restate. The word loneliness appears exactly once in the 70-page proposal, in a list of harms in the opening paragraph. (link)
⏺️ The 2027 Dutch budget leaves about €1.6 million on loneliness-specific subsidy lines and stops naming the €470 million ouderenenvelop. Presented on Prinsjesdag, 15 September. Reading the VWS chapter’s explanatory memorandum rather than the commentary: the identifiable loneliness and caring-communities subsidies for 2027 are structural support and knowledge on loneliness at €997,000, the national network for caring communities at €467,000, and the residual “verminderen eenzaamheid” line at €132,000, falling to €8,000 in 2028. Two lines that existed in 2025, an impulse fund for loneliness interventions and a knowledge-sharing line for young people, are gone. The 2026 memorandum stated that the €470 million structural ouderenenvelop on the Aanvullende Post was available for these plans; the 2027 text carries no equivalent sentence, and states separately that Aanvullende Post money is no longer included in the premium-funded care total. That is a disappearance from the document, which is not the same as abolition, and the parliamentary questions will be the place to find out which. The Seniorencoalitie, led by ANBO-PCOB with ActiZ, Aedes, Alzheimer Nederland, Sociaal Werk Nederland, VNG and Zorgverzekeraars Nederland, had asked in February for €100 million structurally from 2027 for caring communities; the phrase “100 miljoen” does not appear in the budget. The budget does carry something new and useful: a loneliness indicator table from the Gezondheidsmonitor, showing strong loneliness by age for 2020, 2022 and 2024, which fell in every age band between 2022 and 2024 (18–34: 10.7, 15.4, 14.1; 85+: 14.3, 15.6, 14.2) while staying above 2020 levels for the under-50s, with a footnote warning that 2020 was a pandemic year and the survey’s lower age bound changed. (link)
⏺️ Adam’s Law will take effect in California without its auditor-independence rules, because a second bill was chaptered a day earlier. The digest covered the signing of SB 1119 last week and flagged the open question; here is the answer. AB 1405, on registration of artificial intelligence auditors, was chaptered on 9 September as Chapter 178; SB 1119 followed on 10 September as Chapter 190. SB 1119’s chaptered text contains two mutually exclusive versions of Business and Professions Code §21814, one operative if AB 1405 is chaptered and takes effect by 1 January 2027 and one if it is not. AB 1405 is a non-urgency statute chaptered in 2026, so it takes effect on 1 January 2027 and the second version controls. That version omits, relative to the other, the entire subdivision on auditor independence and competence (external auditor, free from operator influence, demonstrated competence, professional auditing standards, no financial interest in the operator, no payment contingent on the result), together with the requirements to disclose audit personnel, conflict-of-interest procedures and methodology, and the lead auditor’s signature certifying results under penalty of perjury. The legislature evidently expects AB 1405’s registration regime to supply those safeguards instead. Read the chaptered bill and not its summary: the Legislative Counsel’s Digest printed at its head still describes the perjury certification and the state-mandated local programme it would create, which is a description of the version that is now inoperative. On substance, §21812(d)(5)(A) bars a companion chatbot from expressing or simulating romantic interest in a child and from encouraging reliance on it for emotional support, with session defaults of one continuous hour and two hours daily; first audits are due by 1 January 2029, and operators below $500 million in prior-year revenue are exempt from the audit section until 2032. Set beside the Commission’s Article 14 and Recital 32, the two texts reach for the same behaviours by different routes. (link)
⏺️ The WHO Commission on Social Connection’s term expires this year, and the plan for what follows is a transition plan. This item is older than the window and I am including it because I have seen it discussed nowhere: the Director-General’s report EB158/28, dated 4 December 2025 for the Executive Board’s 158th session. The Commission was established in November 2023 for a three-year term running 2024 to 2026, with eleven commissioners. The report states that WHO “is developing a transition plan to integrate the Commission’s functions into its ongoing work on health determinants, health promotion, urban health and behavioral insights beyond 2026”, that the Technical Advisory Group on Social Connection continues, and that a resource mobilisation plan is being prepared. Two other threads matter more for people in this network. WHO has begun developing guidelines on interventions to reduce loneliness and social isolation across the life course; the first Guidelines Development Group meeting settled scope, research questions and the WHO-INTEGRATE framework, systematic reviews are under way, and the group was to reconvene in mid-2026 to formulate draft recommendations using GRADE. We could find no WHO publication confirming that meeting took place, and no draft recommendations, as of 21 September. And a proposal has been developed for a global index on social connection, giving a standardised international measure for countries to track over time; that index does not yet exist, which makes the next year the period in which its content is decided. Progress reports next to the Eightieth World Health Assembly in 2027. (link)
⏺️ Practice
⏺️ Ireland’s Men’s Sheds have gone long enough without their running-cost grant that councillors are counting the gap in years, and the department will not commit before the Budget. The Irish Men’s Sheds Association has told local media that the 2025/26 grant will not proceed and that the next dedicated round is expected in 2027, which councillors in Westmeath and Donegal put at a gap of at least 18 months for roughly 450 sheds. The Department of Rural and Community Development and the Gaeltacht told Donegal Daily that “all funding programmes in the department are subject to the Annual Budgetary Estimates process which is currently ongoing and therefore, the department cannot make any definite future funding commitments at present”, that Minister of State Dara Calleary has asked the association to re-examine how it distributes the grants after complaints that the process was cumbersome, and that sheds can apply to the Local Enhancement Programme, SICAP and LEADER in the meantime. Two corrections to how the last round is being described, from the primary record: the gov.ie item of 4 May 2026 is a completion-of-payments notice, not a new announcement, and its grant table is headed “Grant 2024/2025”; it records grants paid in full to 369 sheds, with individual awards running from €262.49 to the €3,000 cap, so the round should not be read as €3,000 multiplied by anything. The €1 million headline belongs to an earlier announcement, which was pitched at supporting up to 450 sheds and combined that year’s contribution with the previous year’s. I could find no departmental statement on gov.ie confirming either the cancellation or the 2027 date, so both rest on the association’s account as reported locally. Budget 2027 is on 6 October, which is when this becomes checkable. (link)
⏺️ The Dutch Week tegen Eenzaamheid opens on 23 September with a professional conference that has already closed its list. The week runs from 24 to 30 September, with the national opening the day before at Gooiland in Hilversum, 10:00 to 17:00, free to attend, and now at maximum capacity with a waiting list. The programme includes a session on loneliness inside your own organisation and what employers can do about it, one on 16 to 27-year-olds from recognition to action, a workshop with the municipality of Hilversum on the tension between loneliness policy and how it is communicated, and a plenary on whether social health is the answer to the loneliness approach. An oversubscribed government-facilitated convening is a demand signal, and it sits oddly beside the budget published eight days earlier, in which the national line funding this area of work drops to €132,000 for 2027. (link)
⏺️ Innovation and startups
⏺️ Character.AI’s chief executive is leaving for Disney with an unspecified number of his engineers, and the company’s public statement on its own future runs to seven words. On 18 September Disney named Karandeep Anand as Senior Executive Vice President and Chief Technology Officer, a new position reporting to chief executive Josh D’Amaro, effective 2 October, adding in one sentence that “a number of Character.AI’s technical team are expected to join Disney”. Character.AI’s own post the same day says of the platform: “The app is still live and online. We will continue to share more information.” This is not an acquisition. Neither document announces a sale, an investment, a licence, or any transfer of models, technology or data, and no consideration is disclosed because none has been described. It is the second leadership cohort the largest consumer platform built on simulated relationships has lost to an incumbent, after founders Noam Shazeer and Daniel De Freitas returned to Google in 2024 under a licensing and talent arrangement, and the pattern should be named precisely instead of assumed into a transaction nobody has announced.
No published evaluation of Character.AI as an intervention was found; we searched ClinicalTrials.gov by sponsor, which returns zero studies, the openFDA 510(k) decisions, and the general literature and news. No registered trial. No FDA clearance or De Novo, no CE marking under MDR, no UKCA, no NICE early value assessment and no NHS DTAC assessment: the product avoids medical-device status by making no clinical claim, which is a live question given the litigation and regulatory attention it faces over minors. The number of technical staff moving is company-sourced and unquantified, and the monthly-active-user figure repeated in coverage traces to a conference speaker biography and not to an audited disclosure. The academic work on companion chatbots sometimes invoked around this company, including the recent study of user grief after personality updates, used Replika and ChatGPT as its natural experiments and is not evidence about this product.
(link)
⏺️ Sword and Headspace now say their deal will close at the beginning of Q4, two weeks after the effective date their own regulatory filing gave. The digest reported this transaction last week as scheduled to take effect on 14 September with no completion announced; what has changed is that the companies have spoken. Their first joint statement, on 16 September, says only that “the acquisition is expected to close by the beginning of Q4 2026, subject to customary closing conditions”, and says nothing at all about price. The Massachusetts Health Policy Commission material change notice, whose affidavit is signed 22 July and which entered the commission’s record on 6 August, gives a proposed effective date of 14 September and describes the consideration only as “a cash payment subject to customary post-signing adjustments”. The $200 million to $300 million range in circulation traces to an Axios Pro report of 26 August citing multiple anonymous sources; neither company has stated a figure in any document we read, and as of 21 September Headspace’s own site still uses the present progressive. We could locate no published commission determination on whether a cost and market impact review will be opened. Hans co-founded Entrelacs, which sells conversational mental-health screening, care matching and continuous monitoring to employers, clinicians and health systems, and the merged company sells into the same buyers, so treat this item as written by a competitor.
No regulatory clearance or marking was found for the Headspace app or for Sword’s mental-health tools: we searched openFDA’s 510(k) database by applicant and found no clearance, no De Novo, no CE marking under MDR, no UKCA, no NICE early value assessment and no NHS DTAC assessment, and both avoid medical-device status by framing as wellness or by operating through licensed clinicians. Headspace and its parent OrangeDot sponsor no registered trial: a sponsor search returns 17 studies and none has either as lead sponsor, with the studies that use the app led by universities and hospitals. Sword Health SA is lead sponsor of 15 registered studies, the great majority musculoskeletal, pelvic or rehabilitation, and has recently registered two touching mental health, NCT07183085 and NCT07179796, both enrolling by invitation and neither with posted results. The announcement’s “84 peer-reviewed studies” is a publication count with no list, citations or methodology published, and the MindEval and MindGuard frameworks described as open-sourced have no publication details we could find, so there is nothing to check either claim against. The price range, the patient and employer counts and the savings figures are all company- or source-derived and not independently verifiable.
(link)
No funding round in scope was announced this week, and no award, prize or accelerator cohort in scope was announced either.
⏺️ Money
⏺️ Scotland’s dedicated loneliness fund ends in March, and the fund named to carry the work is £500,000 and not yet open. The Programme for Government 2026 to 2031, published 1 September, commits to “increase investment in the Community Cohesion Fund to £500,000, supporting grassroots community initiatives across Scotland to address shared challenges such as social isolation, the cost-of-living crisis, inclusion, employment and housing support and community dialogue”, and to establish a Ministerial Taskforce on Community Cohesion. Simita Kumar, Minister for Equalities and International Development, announced the increase on 17 September on a visit to Big Hearts Community Trust in Edinburgh. The fund was £300,000 last year, which itself sat on top of £7.9 million allocated to third sector organisations for 2025-26, so this is not the whole of Scottish spending in the area. Application details “will be published shortly”, meaning there is no opening date, no deadline and no per-project amount to plan against, and the fund is not open. The context makes it sharper: the Social Isolation and Loneliness Fund, launched in August 2023, culminates in March 2027 with over £3.8 million invested and at least 20,000 people supported, after a £1.05 million extension to 52 organisations announced in June. A loneliness-specific fund closing and a smaller general cohesion fund opening is a change in how the work is categorised as much as a change in its size. The beneficiary counts do not reconcile: the September release says 38 organisations benefited from the cohesion fund in 2025-26, while gov.scot’s own February release says 32 projects shared the £300,000, and I have not established which counts what. (link)
No new research call in scope opened anywhere I looked this week. GOV.UK’s Find a Grant, the primary register for UK central government, returns zero results for “loneliness” and five for “social isolation” that match only the word (a boiler upgrade scheme, an electric vehicle chargepoint grant, school drainage, engineering biology, and a wireless technology fund). With the DCMS gaming pilot closed on 11 September, the UK register is now empty on this beat. ZonMw lists 45 open calls, none on loneliness or social connection, the nearest being a consortium call on sustainable first-line care networks and neighbourhood ties that closes 10 November. No Horizon Europe call in scope opened in September. Still open from previous issues: the Fondation des Sciences Sociales call on “Solitude et isolement”, closing 7 November, and Sweden’s Socialstyrelsen loneliness grant round, closing 15 October.
Corrections, arguments and anything I have missed are welcome in the thread, and they change both the next issue and the brief this one was written from. If you have run an evaluation with a loneliness outcome that went nowhere, that is exactly the thing this digest struggles to find and would most like to carry.
Hans
Discuss this issue in the General space (members only).