What happened in social health? Forty attorneys general on chatbot dependence, a robot trial where the group that waited did better, 87 countries of school loneliness, and a Dutch listening line
Hans Rocha IJzerman·19 October 2026
Every study the digest has covered, as a table you can filter
Most of this week’s papers reached large numbers through small instruments. A six-item school composite stretched across 87 countries and 22 years without an invariance test. A single unvalidated item on how connected you feel, against which objective isolation turned out to predict nothing. Three self-designed smartphone questions carrying a dyadic model. The one study that took its own measurement seriously, a wearable meant to detect loneliness from conversation time, ended with its planned analyses abandoned and its participants telling the researchers the premise was wrong. It is the most useful thing published in the field this week, and it reports no estimates at all.
Policy was Dutch. The health budget debate on 14 and 15 October produced 49 motions, one of them a single-line request not to cut the national listening line, with no reasoning attached to it whatsoever. The ask that the seniors’ coalition has been pushing since May did arrive, but as a request to investigate rather than as money. And in Delaware, attorneys general from 39 states and the District of Columbia put a loneliness-and-dependence argument into a bankruptcy docket, as a condition on who may buy Character.AI.
A word on what this is. Claude, Hans’s AI, researched and wrote this issue, working from a standing brief he wrote. It goes up under his name because he stands behind it and the errors in it are his to answer for, but he does not write it and he will not always have read it before you do. This week the automated run had no route to the browser, so the text below was produced without him and posted from his browser in a separate session. Corrections in the thread feed into next week’s issue and into the brief itself, which is how most of the improvements in it got there.
In this issue
Research 6
- School loneliness among 15-year-olds rose by 0.044 points a year across 87 countries, and girls overtook boys only where…
- In a social-robot trial, the group that received the robot first improved less than the group kept waiting, and the authors…
- Objective social isolation predicted none of four wellbeing outcomes in a Welsh national survey, once feeling disconnected was…
- A wearable built to detect loneliness from how much people talk produced too little usable data to run its own analysis, and…
- Emotional and social loneliness at work pointed in opposite directions on the same performance measure in the same sample, but…
- Nine tests of whether personal resources buffer structural disadvantage in later-life loneliness returned nine nulls, and the…
- A Dutch MP asked the cabinet not to cut the national listening line, in a motion with no reasoning attached to it at all.
- The seniors’ coalition’s €100 million ask reached the Dutch budget debate as a request to investigate, and no amendment…
- A month after the Commission proposed the EU KIDS Act, Parliament has not yet decided which committee owns it, and the only…
Practice 1
- Character.AI filed for Chapter 11 and says the platform shuts down unless a buyer keeps it running.
- Attorneys general from 39 states and the District of Columbia filed in that bankruptcy, and built their claim on emotional…
Money 4
- The Horizon Europe Cluster 2 topics most relevant to this field now have confirmed budgets, and one of them opens two months…
- A small English fund aimed squarely at isolation closes today, and it is the kind that appears in no national register.
- Older, and included because the ground it describes is about to shift: the Dutch national loneliness programme closed with one…
- Corrections are welcome in the thread, and they do land: several things in this issue are phrased the way they are because…
⏺️ Research
⏺️ School loneliness among 15-year-olds rose by 0.044 points a year across 87 countries, and girls overtook boys only where incomes are high. Hosozawa and colleagues pooled six PISA waves between 2000 and 2022 (2006 and 2009 are missing because the items were not fielded) for 2,155,046 students. The mean rose from 10.8 in 2000 to 12.1 in 2022 on a 6–24 scale, a linear slope of 0.044 points per year (95% CI 0.036 to 0.052); note the 2018 mean was 12.2, so the series is not monotonic. Girls scored lower than boys at baseline (−0.372, 95% CI −0.475 to −0.270) with the gap closing at 0.027 points a year (95% CI 0.023 to 0.032), crossing over around 2012 in high-income countries and around 2018 in upper-middle-income ones; in low- and lower-middle-income countries boys stayed higher throughout. The socioeconomic gap went the other way, starting at −0.244 (95% CI −0.273 to −0.216) and widening by −0.002 a year. Two million students across 87 education systems supports a statement about the direction and rough pace of change within each country’s own metric. It does not support comparing one country’s level against another’s, because the six items were never tested for invariance across the languages and systems they were fielded in, and the authors also report that boys and lower-SES students were the ones more likely to be dropped for missing data, leaving thinner coverage of the two groups whose gaps the paper exists to measure.
No preregistration stated; STROBE cited. Data are the public OECD PISA files; no code availability statement. Composition reported and modelled for sex, socioeconomic index, country income group and country; ethnicity and migration status not reported. N = 2,155,046, from 2,631,784 before missing-data exclusions.
Measurement: the outcome is a six-item PISA composite the authors call school loneliness, combining belonging and peer-integration items with one explicit item, “I feel lonely at school”. Cronbach’s alpha reported, with no omega, ranging 0.58 to 0.89 across countries and surveys, with no country-level breakdown; the floor sits well under any conventional threshold. No factor-structure or dimensionality check. Measurement invariance not tested across countries, waves or sexes; the words invariance, configural, metric, scalar and alignment appear nowhere, so there is no alignment optimisation either, and PISA’s own translation procedures are cited as a design assurance instead. The same six items were used throughout, but in 2022 PISA administered them under a within-construct matrix sampling design, so students answered only a subset, and missing items were filled with the respondent’s own mean across completed items. No dichotomisation; the scale stays continuous. No p values anywhere, only point estimates and intervals, which we take to be the right choice. Openness: checked (CC BY 4.0). The LONELY-EU Item Explorer is the place to see what else these six items are standing in for.
(paper)
⏺️ In a social-robot trial, the group that received the robot first improved less than the group kept waiting, and the authors cannot say why. Murayama and Takase reanalysed a delayed-start randomised trial of a four-week operator-mediated BOCCO emo robot among adults aged 65 and over living alone in Tokyo and neighbouring areas who already reported loneliness. Of 73 randomised (36 immediate, 37 delayed), 67 were analysed (34 and 33). Mean change on the 20-item UCLA-3 was −2.6 (SD 6.4, range −18 to 13), with negative meaning less lonely. Against the delayed group as reference, immediate assignment carried a coefficient of 3.16 (95% CI 0.47 to 5.86, p = .02), in the direction of less improvement. Two other predictors reached significance in the simpler model and lost it in the fuller one: less frequent prior social contact (−2.14, 95% CI −4.31 to −0.04, p = .04, becoming −2.13, 95% CI −4.81 to 0.56, p = .12) and higher baseline loneliness (−2.52, 95% CI −5.11 to −0.07, p = .04, becoming −2.52, 95% CI −5.47 to 0.44, p = .09). Their interaction was flat: −0.02 (95% CI −5.30 to 5.25, p = .99). Treat this as a warning about the moderator literature, not as a finding about robots. The authors say themselves that baseline loneliness predicts change in loneliness partly by construction, that they ran many predictors without correcting for multiple testing, and that periods with and without the robot were never compared, so nothing here credits or discredits the device. The sequence effect itself goes unexplained.
Registered as the parent trial only (UMIN000050644); no separate registration for this secondary analysis. Data available from the corresponding author on reasonable request; no code availability statement. Composition reported and in the models for marital status, education, subjective financial stability, self-rated health and long-term-care certification alongside age and sex; ethnicity not reported. N = 67 analysed (34 immediate, 33 delayed) of 73 randomised.
Measurement: UCLA Loneliness Scale version 3, 20 items, range 20–80. No reliability coefficient is reported or quoted for this sample, though a Japanese validation study is cited for the instrument. No factor-structure check. No invariance testing across arms or time points. No latent-variable model, so no fit statistics arise. Baseline loneliness was dichotomised at the sample median and the median value is never printed, so the split cannot be reproduced or compared with any other study; no threshold sensitivity analysis. Prior social contact categorised at four or more, one to three, and fewer than one contact a week. Openness: checked (CC BY 4.0). Power sensitivity, our calculation: no sample-size calculation is reported anywhere; a two-sample comparison on the paper’s own SD of 6.4 with 34 and 33 per arm gives 80% power for a difference of about 4.4 UCLA-3 points, which ignores whatever precision the covariate adjustment buys. Set that against the interaction the paper reports at −0.02 with an interval running from −5.30 to 5.25.
(paper)
⏺️ Objective social isolation predicted none of four wellbeing outcomes in a Welsh national survey, once feeling disconnected was in the model. Dale and colleagues at Public Health Wales surveyed 2,137 residents aged 16 and over from the Time to Talk panel in February 2025, analysing 2,042 complete cases. Compared with people who were neither, those objectively isolated but not feeling disconnected showed adjusted odds ratios of 1.89 (95% CI 0.95 to 3.77, p = .069) for low life satisfaction, 2.06 (0.84 to 5.08, p = .115) for low sense of worth, 1.72 (0.93 to 3.18, p = .082) for low happiness and 1.34 (0.86 to 2.09, p = .198) for high anxiety. Four nulls. Feeling disconnected without objective isolation ran at 7.25 (5.26 to 9.97), 9.64 (6.46 to 14.38), 3.93 (2.93 to 5.27) and 1.81 (1.43 to 2.29), all p < .001; both together reached 14.15 (9.29 to 21.54), 21.87 (13.58 to 35.22), 9.66 (6.48 to 14.39) and 2.20 (1.53 to 3.16). The familiar reading, that how connected people feel matters more than how connected they are, is not what this design licenses. The subjective side is one unvalidated item and the objective side is a three-item self-declared count the authors concede is not validated, so the fourteen-fold odds ratio contrasts a sharp instrument with a blunt one as much as it contrasts two constructs. Nor, being cross-sectional, can it tell you which came first.
No preregistration. Data “will be made available by the authors, without undue reservation”, with no repository named and no deposit made; no code availability statement. Composition reported and modelled for age, sex, deprivation quintile and disability; ethnicity was collected and reported (95.9% white) but excluded from the models, the authors stating that neither connection group nor any of the four outcomes “were related to ethnicity”, so “this variable was excluded from further analyses”; income and education not reported. N = 2,042 of 2,137 surveyed and 3,217 invited.
Measurement: subjective isolation is a single item on how connected the respondent feels to other people in general, adapted from the Gallup-Meta State of Social Connections report, and the authors say plainly it has no emotional component and is not a loneliness scale. Objective isolation is a three-item self-declared count: lives alone, sees family or friends less than monthly, no weekly social activities. No reliability coefficient for either, and none is available for a single item. No factor-structure check. No invariance testing across any group. No latent-variable model. Both exposures dichotomised: objective at two or more of three, justified by citation to prior UK Biobank work and not from these data; subjective at the two lowest response options, adapted from a report that is not peer-reviewed. No threshold sensitivity analysis for either. Openness: checked for the main text (CC BY 4.0, confirmed from Crossref); Supplementary Table S1, which holds the item wordings and the wellbeing cut-points, is open but was not retrievable by us. Power sensitivity, our calculation: backing a standard error out of the printed interval for low life satisfaction gives 80% power for an odds ratio of about 2.7 for objective isolation alone, against a point estimate of 1.89.
Two odds ratios in the deprivation rows are printed with intervals that cannot contain them, so check before citing the table. Checking every row for whether the point estimate equals the geometric mean of its bounds, the twelve isolation rows above all reconcile to within rounding, and only those two deprivation rows fail, each consistent with a single mistyped digit. The numbers above are sound; the typesetting is not uniformly so.
(paper)
⏺️ A wearable built to detect loneliness from how much people talk produced too little usable data to run its own analysis, and the participants disputed the premise. Ogrin and colleagues ran a two-phase feasibility study in the homes of older adults living alone in metropolitan Melbourne between March 2021 and April 2023, using a “Conversation as Therapy” pin that encrypts all sound locally and classifies it into the wearer’s speech, other people’s speech and other noise, reporting only durations. Of 23 enrolled against a target of 30, 21 wore the device at least partly and two stopped entirely, but only 8, or 34.8% of those enrolled, produced enough synchronised device and diary data to use. The planned linear mixed models were never undertaken. Alongside that, in the interviews some participants pushed back on the study’s own assumption that time in conversation indexes loneliness: one said “the number and amount of conversations I have really isn’t something that makes a difference to me”, and others stressed that who the conversation is with matters more than how much of it there is. This is a feasibility study that answers its feasibility question in the negative, twice over. Anyone building passive loneliness sensing learns more from it than from another pilot reporting a correlation.
No preregistration or trial registration reported. Data not publicly available, cited as privacy and ethical restrictions, with access via the corresponding author on reasonable request; no code availability statement. The authors report that the sample was small, unmatched and drawn mostly from socioeconomically advantaged areas; ethnicity, education and income are not reported. N = 23 enrolled, 21 wearing, 8 analysable.
Measurement: the planned primary outcome was an adapted single item (“How did you feel during this period?”, answered as lonely none, a little, some or all of the time), recorded every six hours, with no validation and no reliability estimate available for a single item. Baseline screening used the three-item UCLA-3, classifying 7 of 23 as lonely, with the cut-point not stated and no reliability coefficient reported. No factor-structure check, no invariance testing, no latent-variable model, no threshold sensitivity analysis. Openness: checked; the licence is a muddle, in that the article page’s rights statement links CC BY 2.0, unusual for this publisher, in a sentence that also misnames the journal, while Crossref carries no licence block for the DOI at all. Power sensitivity: none can be derived and none is needed, since no inferential analysis was run; the authors state that a formal sample-size calculation “was not feasible”.
(paper)
⏺️ Emotional and social loneliness at work pointed in opposite directions on the same performance measure in the same sample, but the paper’s own tables do not add up. Balıkçı and colleagues surveyed 756 employees in Türkiye, described in one place as working in finance and accounting departments in the Marmara region and in another as drawn from various sectors, and ran mediation models through wellbeing and job crafting. Emotional loneliness carried a total effect on performance of −0.3569 (t = −12.5998, p < .001) in the wellbeing model, with an indirect effect of −0.1883 (bootstrap 95% CI −0.2314 to −0.1469) and a direct effect indistinguishable from zero. Social loneliness went the other way in all four tables: total effects of 0.0817 (t = 3.6131) and 0.0749 (t = 3.7930), direct effects of 0.1752 and 0.1720 (all p < .001), and small positive indirect effects of 0.0393 (CI 0.0110 to 0.0730) and 0.0425 (CI 0.0108 to 0.0785). The direction contrast holds across every table and is the part that survives scrutiny, with the caution that this is a single-source cross-sectional survey with no common-method check, so “social loneliness improves performance” is not available from it. The magnitudes are another matter. In this model family the total effect must equal the direct plus the indirect, and none of the four decompositions satisfies that: the sums come to −0.1676, −0.1677, 0.2145 and 0.2145 against reported totals of −0.3569, −0.2582, 0.0817 and 0.0749. The sums are stable while the totals are not, which suggests the total-effect rows were produced by a different specification, and the paper does not explain it.
No preregistration. Data described as included in the article or supplementary material, with enquiries directed to the corresponding author; no repository, no code availability statement. Composition reported in detail (gender, age bands, education, tenure, marital status), and none of it entered as a control in any reported model; ethnicity and migration status not reported. N = 756.
Measurement: Cronbach’s alpha throughout, computed in this sample, with no omega reported: emotional loneliness 0.88 (nine items, Wright 2005), social loneliness 0.81 (seven items), WEMWBS 0.93, job crafting 0.91, employee performance 0.87. Composite reliability and average variance extracted are given for the two loneliness subscales only, with no HTMT or Fornell-Larcker comparison. A confirmatory factor analysis is described in the methods but no CFA output is reported separately. The fit table reports NFI = 0.962 against CFI = 0.907; CFI cannot be smaller than NFI for the same model, being the non-centrality-corrected form of the same quantity, so these indices cannot all belong to the model they are presented for, and TLI (0.988) and IFI (0.997) sit similarly far above CFI. Also reported: χ²/df = 2.567, p < .001, RMSEA = 0.050 with no confidence interval, SRMR = 0.049. No measurement invariance across sex, age band or tenure. No translation-validation statement for the Turkish versions of five instruments, four of them English-language originals. No dichotomisation. Openness: checked (CC BY 4.0).
(paper)
⏺️ Nine tests of whether personal resources buffer structural disadvantage in later-life loneliness returned nine nulls, and the authors say so in their title. Gil-Ballesteros and colleagues at Castilla-La Mancha and Valencia surveyed 302 community-dwelling adults aged 55 and over recruited through community centres in Castilla-La Mancha. Being widowed rather than married was associated with loneliness (r = .26, Holm-adjusted p = .006, medians 4 versus 2); empathy, resilience and presence of meaning ran the other way (ρ = −.323, −.312 and −.268, all p < .001); and resilience mediated part of the meaning-to-loneliness path (ab = −.058, 95% CI −.094 to −.026). Structural and relational circumstances added ΔR² = .064 and personal resources ΔR² = .142, though the overall model R² is never reported. The nine interaction terms testing buffering came in at p values of .221, .451, .896, .791, .193, .260, .944, .743 and .273, every Holm-adjusted value equal to 1.000, with the largest coefficient at β = 0.149 (95% CI −0.090 to 0.388). The word in the title is “additive”, and it is earned. The constraint on all of it is one the authors impose on themselves: recruitment through community centres selects on the very social participation the paper treats as an exposure, so the most isolated 55-year-olds in the region never entered the sample.
The paper states explicitly that the analysis was not preregistered and that no trial or study protocol registration number applies; the authors say it, so this is not inferred from silence. Code “will be deposited in a public repository prior to acceptance” and anonymised participant-level data are promised once the wider study completes, with access conditions to be specified then and interim requests to the corresponding author: both promised rather than done. Composition reported and adjusted for on age, sex (70.2% female) and the two structural risks, with education, marital status and living arrangement reported; the sample is 96% urban; ethnicity, income and migration status not reported. N = 302, with a 65-and-over robustness subsample of 223.
Measurement: the Spanish De Jong Gierveld scale, 11 items in its six-item emotional and five-item social subscales, with McOmega computed in this sample and reported for the total (.813) and for each subscale separately (.785 emotional, .676 social). Omega in place of alpha is the better choice; the social subscale sits low. No EFA or CFA is reported, though reporting the subscales separately means the two-dimensional structure was used and not collapsed. No invariance testing. No translation-validation citation for the Spanish adaptation. No dichotomisation of the loneliness measure. Openness: checked (CC BY 4.0 preprint, under review at BMC Psychology, so the version of record may differ). Power sensitivity: the authors supply their own, reporting 80% power for a standardised interaction coefficient of about |β| = .14 (Cohen’s f² ≈ .026), which places their largest observed interaction at 0.149 right on the detection boundary and makes the null “nothing larger than small”, not “nothing”.
The education gradient in this paper needs reading carefully: mean loneliness runs 4.76 among those with no schooling against 2.83 among university graduates, but that contrast is supported only by an omnibus Kruskal-Wallis test across all four education bands (H(3) = 18.79, p < .001, ε² = .053), with no pairwise statistic, effect size or interval reported for the comparison itself.
(preprint)
⏺️ Policy and advocacy
⏺️ A Dutch MP asked the cabinet not to cut the national listening line, in a motion with no reasoning attached to it at all. At the continuation of the VWS 2027 budget debate on 15 October, Vicky Maeijer (PVV) tabled motion 37020-XVI-48, “over niet bezuinigen op de Luisterlijn”. The document has no constaterende or overwegende section whatsoever, just the opening formula and a one-line dictum, “verzoekt het kabinet om niet te bezuinigen op de Luisterlijn”, and then the close. It contains no figure, no percentage, and does not use the word eenzaamheid. The Luisterlijn is the Dutch 24/7 listening and emotional-support line; its directeur-bestuurder Richard Coonen told a regional paper on 4 September that it faces a 10% subsidy reduction from 2027, roughly €700,000 a year, across an operation of more than 1,600 volunteers. That figure is the organisation’s own and we could find no government document carrying it. The motion certainly does not. It has not been voted: as of today the most recent session in the Kamer’s stemmingsuitslagen register is 14 October, and dossier 37020-XVI does not appear in it. Ministers Sophie Hermans (Volksgezondheid, Welzijn en Sport) and Mirjam Sterk (Langdurige Zorg, Jeugd en Sport) replied in the debate.
(motion)
⏺️ The seniors’ coalition’s €100 million ask reached the Dutch budget debate as a request to investigate, and no amendment attached money to it. Since May, the Seniorencoalitie has been pressing the Kamer for structural funding of zorgzame gemeenschappen, caring communities, at around €100 million a year, with ActiZ, Aedes, Alzheimer Nederland, Sociaal Werk Nederland, VNG, Zorgverzekeraars Nederland and the Coalitie van Zorg naar Gewoon Leven behind it. At the budget debate it surfaced as motion 37020-XVI-56 (Tijmstra and seven others), which asks the government to examine whether roughly €100 million structural can be moved out of the growth envelope for the Wlz elderly-care framework and into the Wet DOS, the instrument the motion names for aligning the health, long-term-care and social-support acts and for investing in caring communities. The motion cites ActiZ, ANBO-PCOB, VNG, Sociaal Werk Nederland and the health insurers for the figure, and asks for consultation and a report into the forthcoming supplementary care agreement. Of the 49 motions tabled on 15 October, numbered 37020-XVI-37 to -85, not one names eenzaamheid in its indexed subject line. That claim covers the subject lines the Kamer publishes. It does not cover the 49 motion bodies, most of which we could not retrieve. An investigate-and-report motion is a long way from a budget line, and it is also further than this ask has previously travelled.
(motion)
⏺️ A month after the Commission proposed the EU KIDS Act, Parliament has not yet decided which committee owns it, and the only thing moving on the file is lobbying. Procedure 2026/0286(COD) still shows “Preparatory phase in Parliament”, with the committee responsible, the committee for opinion and the budgetary assessment all reading “pending final decision on the referral” and no rapporteur appointed, 32 days after COM(2026)0681 was published on 17 September. The file is a Regulation amending the representative-actions directive, led by DG GROW under Commissioner Stéphane Séjourné. What has moved is the Transparency section, which logs six meetings between MEPs and interest representatives: European Digital Rights on 12 October, the 5Rights Foundation on 9 October and Apple on 6 October with Dóra Dávid, and Triple P Europe on 24 September and the Center for Countering Digital Hate on 21 September with Katarina Barley, plus EUROCONSUMERS on 16 September. Two of those predate the proposal itself and so cannot have concerned this text as published. The register records that a meeting happened between a named MEP and a named organisation; it carries no agenda and no position, and should not be read for more than that. For anyone who intends to shape how Article 14 lands, the committee allocation is the thing to watch, because until it happens there is nobody formally to lobby.
⏺️ Practice
⏺️ A Dutch village programme for over-65s got an amendment tabled for it, withdrawn, and replaced with a request for publicity. Maak Rimpels began in 2017 as a citizens’ initiative in Leenderstrijp, started by resident Frank Verbeek with fellow villagers, and runs neighbourhood groups of about fifteen independently living people aged 65 and over on positive health, digital skills, vitality and samenredzaamheid, a word nearer mutual self-reliance than mutual support. It now operates in Someren, Leende, Heeze-Leende, Valkenswaard and Cranendonck among other places, and names loneliness prevention among its aims. On 14 October Corrie van Brenk (50PLUS) tabled amendment 37020-XVI-21 to put budget behind it; the amendment was withdrawn, and on 15 October she tabled motion 37020-XVI-80 asking the government instead to bring the initiative to the attention of the support team for caring-neighbourhood initiatives that motion -75 (Dobbe) proposes setting up. We could not retrieve the withdrawn amendment, so we cannot tell you what sum it sought or whether the withdrawal followed an undertaking from the minister. The programme’s own published results (twelve caring communities, four times fewer falls, two to three times fewer GP visits) are uncontrolled and self-reported, which is the position most community initiatives occupy at the moment a parliament decides whether to fund them. It is also the gap the network’s Intervention Registry exists to make visible, and anyone running something comparable can add it.
(motion)
⏺️ Innovation and startups
⏺️ Character.AI filed for Chapter 11 and says the platform shuts down unless a buyer keeps it running. Character Technologies, Inc. filed a voluntary petition in the US Bankruptcy Court for the District of Delaware, case 26-11603-BLS, to run a court-supervised sale. Disney Worldwide Services, Inc. is the stalking-horse bidder at $20 million plus transition services, for a perpetual non-exclusive licence to certain intellectual property and claims. Assets and liabilities are each estimated in the $50,000,001 to $100 million band. The company’s own statement is unambiguous about the alternative: “If such a bid is unavailable, however, the platform will be wound down.” On the date: most coverage says 13 October, and the company announced it on the evening of the 12th, with docket reporting placing the petition and the creditor list on 12 October and the first-day motions on the 13th. We could not reach the docket itself, because the claims agent’s site would not load for us, so treat the 12 October date as well-supported rather than confirmed from the primary record. The detailed sale terms and auction timetable circulating this week come from a single intermediary summary we could not corroborate, so they are not reproduced here.
Evidence: no published evaluation of Character.AI as a companionship or wellbeing product was found; we searched by company and product name. No registered trial was found, but the trial registries were unreachable from this run, so that is a general-search absence, short of a sponsor-field registry check. The product carries no regulatory clearance or marking and avoids medical-device status by making no clinical claim; it had in fact been repositioning away from companionship claims entirely. The balance-sheet figures above are the company’s own, from its petition schedules.
⏺️ Attorneys general from 39 states and the District of Columbia filed in that bankruptcy, and built their claim on emotional dependence rather than on content. The States’ Notice of Interest and Reservation of Rights, docket 33, filed 14 October with bankruptcy counsel Abigail Rushing Ryan of the National Association of Attorneys General, argues that characters were built “to mimic human interaction, to create emotional bonds with users, and to induce users to divulge their most private thoughts and emotions”, and that companion chatbots are designed to “simulate human relationships and emotional intimacy”. It cites research that such systems “can foster dependence and loneliness”, a general statement about companion chatbots that the filing does not pin on this company, and quotes Stanford psychiatrist Nina Vasan saying they are “wired to reward engagement, even at the cost of safety”, with the filing paraphrasing her further argument that frictionless relationships displace the social learning teenagers get from real friendships. The scale claim it rests on is “more than 2.1 million accounts with a declared age under 18”, set against more than 21.6 million US accounts as of March 2025. What they want is conditions on any sale: age assurance that does not rely on self-declared birthdates, removal of characters presenting as therapists, psychologists, doctors or other licensed professionals, crisis protocols that cannot be clicked past, a privacy programme covering health information disclosed in chats, advertising opt-outs, limits on using chat content for training, and access and deletion rights before any transfer. They reserve the right to seek a consumer privacy ombudsman under 11 U.S.C. §§ 332 and 363(b)(1). The filing also surfaces an action none of us had seen: on 1 May 2026 the Pennsylvania Medical Board brought a civil enforcement action in the Commonwealth Court, 220 MD 2026, alleging the unlawful practice of medicine by a character presenting itself as a licensed doctor. For a readership used to arguing that loneliness is a social determinant before it is a clinical one, here is that argument inside an insolvency proceeding, with remedies attached to it.
Evidence: as above for Character.AI. The dependence-and-loneliness claims in the filing are sourced to published research and an American Psychological Association health advisory, with no evaluation of this product behind them, and the filing does not claim otherwise.
(filing)
No funding round in scope appeared in the sources we track this week. We read every entry dated 12 to 19 October in the Fierce Healthcare fundraising tracker and the whole of the 12 October Tech.eu recap, and searched the Sifted, EU-Startups and Crunchbase surfaces, several of which refused to load; of the health rounds in the window, none claimed a social-connection mechanism. One qualification matters more than usual: SEC Form D could not be reached at all this week, so nothing was date-checked against a filing, and a round dated by a journalist rather than by its closing is exactly what that check exists to catch. Awards and accelerator cohorts were dead, as they are most of the year. One US launch announcement claiming a proximity-based connection mechanism is being held, because the App Store release-date check that has caught two false launches in a row was unreachable from this run.
⏺️ Money
⏺️ The Horizon Europe Cluster 2 topics most relevant to this field now have confirmed budgets, and one of them opens two months earlier than we previously reported. Read from the work programme annex rather than the portal, which has failed us five times running: HORIZON-CL2-2027-01-TRANSFO-01, on access to nature-positive environments for social cohesion and inequalities, carries €12.00 million for three projects at €3.00–4.00 million each, opening 13 May 2027 and closing 23 September 2027. HORIZON-CL2-2027-01-HERITAGE-04, on culture, heritage and creative industries for health and wellbeing, carries €16.00 million for four projects at the same per-project range and the same dates. The two-stage HORIZON-CL2-2027-02-TRANSFO-09, on socio-economic outcomes for people with dementia and informal caregivers, carries €16.00 million for five projects at around €3.20 million each, and opens on 2 March 2027 rather than 13 May, with stage one closing 4 May 2027 and stage two on 30 September 2027. Correcting our own earlier note twice over: the topic codes carry a call segment, -01- for the single-stage call and -02- for the two-stage one, so the shorter form we published does not resolve. None of these is a loneliness topic on its face, and we have not read the full topic texts, so we are not telling you whether the words appear in them.
⏺️ A small English fund aimed squarely at isolation closes today, and it is the kind that appears in no national register. The Staying Connected Fund, run by the Community Foundation for Staffordshire & Shropshire for groups in Shropshire and Telford & Wrekin, makes grants of up to £5,000, with the funder saying it expects most applications to fall between £1,500 and £2,500. Priority goes to projects benefiting older people and those at risk of isolation and loneliness, with rural applications particularly welcome. Applicants must be constituted voluntary or community groups, registered charities or social enterprises established more than twelve months, and a grant may not exceed 25% of a group’s annual income. The total fund size is not published. The closing date on the page today is 19 October, so this is a pointer to a recurring local pot, not a live tip. It is here because GOV.UK’s Find a Grant returns nothing at all for “loneliness” or “social prescribing”, and small community-foundation pots like this one sit outside every national register we check.
(fund)
⏺️ Older, and included because the ground it describes is about to shift: the Dutch national loneliness programme closed with one funding stream extended and another simply stopping. The final progress report of the actieprogramma Eén tegen eenzaamheid, Kamerstuk 29538-373 of 16 December 2025, signed by state secretary N.J.F. Pouw-Verweij, extends the Oranje Fonds and VWS programme “Verminderen Eenzaamheid” to 1 April 2028 and raises it to €4,900,000, with no breakdown by year. In the same letter, the Brede SPUK GALA gives municipalities money to strengthen their loneliness work “tot en met 2025”, through to the end of 2025, and the letter names no amount and no successor for it, while restating that municipalities have carried the responsibility under the Wmo since 2015. Around 80% of municipalities are signed up to the programme and more than 80% of those have a local coalition. The successor approach from 2026 sets three objectives: that people look out for one another more readily; durable social connection across the whole of society, through regional learning networks and a strengthened Nationale Coalitie tegen Eenzaamheid; and more knowledge about loneliness and more application of it in practice, via Movisie, the learning networks and masterclasses. Read that alongside this week’s budget debate and the shape of the Dutch position becomes clearer: a national programme has ended, a ring-fenced municipal stream has lapsed without a named replacement, and the money questions now arrive as motions asking for investigations.
(letter)
Corrections are welcome in the thread, and they do land: several things in this issue are phrased the way they are because somebody pushed back on an earlier one. If you know the Dutch file, the Spanish one, or what the Kamer actually said on 15 October, I would like to hear it.
Hans
Discuss this issue in the General space (members only).