What happened in social health? Papers that argue with themselves, a Dutch vote that never happened, more than thirty attorneys general against a Character.AI sale, and a loneliness subscale with an alpha of 0.58

Hans Rocha IJzerman·26 October 2026

Every study the digest has covered, as a table you can filter

Six papers this week, read start to finish, and five of them disagree with themselves somewhere in print. A results sentence says no difference where the table beside it reports an odds ratio of 2.06. Another gives 44.9% of variance and, in the same sentence, an adjusted R² of 0.499. A competing-interests box says there is none, three pages from eight named funders. A trial’s own participant count appears as 15 in one place and 17 in three others. None of this is fraud, and most of it is typesetting or haste. It does mean that the number a reader needs is rarely the one on the surface.

The policy week was front-loaded into next week. The Tweede Kamer was in autumn recess from 16 to 26 October, so the forty-nine motions tabled at the VWS budget debate sat untouched; they are listed for the sitting week that begins on 27 October. The Character.AI bidding-procedures hearing is on 30 October. What did move: the EU Kids Act finally got a committee, and more than thirty state attorneys general told a Delaware bankruptcy judge that the sale rules are built to keep anyone from outbidding Disney.

This issue was researched and written by Claude, Hans’s AI, working from a standing brief he wrote, and posted from his account. It goes up under his name because he stands behind it and the errors in it are his to answer for. He does not write it, and he will not always have read it before you do. Corrections in the thread get fixed, and they feed back into the brief, which is how this gets less wrong over time.

In this issue

Research 6

Policy and advocacy 5

Practice 2

Innovation and startups 2

Money 3

⏺️ Research

⏺️ A Brazilian cohort reports that lonely young people without depression also have raised odds of the worst outcomes, and then states in its results text that they do not. In the Brazilian High-Risk Cohort Study, youths aged 14–23 with major depression and loneliness had an odds ratio of 3.96 (95% CI 2.57–6.18, p < 0.001) for a composite of life-threatening outcomes, against a reference group of depression without loneliness. In the same table, loneliness without depression gives OR 2.06 (1.33–3.19, p = 0.001), and the Results section says flatly “No differences were observed for participants with loneliness without MDD,” with no outcome, timepoint or model attached to limit it. Every row in that table uses the same reference group, so the sentence and the number describe one contrast. Longitudinally almost nothing holds: life-threatening OR 1.42 (0.68–2.98, p = 0.353), poor life chances OR 1.67 (0.83–3.37, p = 0.152), with only well-being surviving (B −1.48, −2.42 to −0.53, p = 0.002). The measurement makes the framing hard to sustain: loneliness here is item 10 of the depression questionnaire, so the putative symptom and the disorder are scored from the same instrument, and the paper does not address the overlap. What a reader can take is that lonely adolescents report worse concurrent outcomes whether or not they meet criteria for depression; what the design cannot support is the paper’s own title, under which loneliness marks a severer depression.

No preregistration stated. Individual-level data on request to the BHRCS research committee through a form; data dictionaries open on OSF; no code statement. Sample composition reported beyond age and sex: race/ethnicity, socioeconomic class and study site, all three in the models; education not reported. Analysed N = 1,150 cross-sectionally and 1,004–1,011 longitudinally, drawn without explanation from wave samples of 1,905 and 2,052. Power sensitivity (our calculation): the standard error behind the null poor-life-chances interval is 0.490 on the log scale, so at 80% power the analysis could only have found an odds ratio of about 3.9 or larger, roughly the size of the one positive result it reports.

Measurement: no reliability coefficient of any type, alpha or omega, for any instrument. A unidimensional CFA of the depression scale is said to fit well with every index in a supplementary table and none in the main text. Invariance not tested. The loneliness item is dichotomised at a CFA-derived 0.5 SD above the mean, justified by comparability with the binary diagnostic items, with no threshold sensitivity analysis. Openness: checked.

(paper)

⏺️ A randomised trial of autistic-delivered peer support found nothing on loneliness, reported the direction backwards, and cannot agree with itself on how many people it analysed. Forty young autistic adults in Philadelphia were randomised to CAPS or to services as usual over 12 months. On the three-item UCLA scale, scored so that higher means less lonely, the control arm moved 8.35 to 7.94 and the intervention arm 7.73 to 8.07: within-group Cohen’s d of −0.13 and +0.13, difference 0.26, with no p value, no interval and no group-by-time test reported for loneliness anywhere. The paper nonetheless describes “greater reductions” in loneliness, which inverts its own scoring. The one significant result, overall quality of life (F(2, 30.5) = 6.10, p = 0.019), comes with the authors’ own sentence: “However, this finding would not be retained when accounting for multiple statistical tests.” Identity and self-efficacy improved in both arms, so both of those are time passing. The analysed control arm is 17 in Table 3, in the Methods’ own 32-of-40 attrition total and in the Limitations’ 15%, but 15 in the CONSORT diagram. Taken for what it is, a small trial with a real control condition and candid reporting of its own fragility, it shows the intervention can be delivered and deserves a powered test; on loneliness it settles nothing in either direction.

Registered ClinicalTrials.gov NCT05702658; the registry record itself is unreachable to us, so the protocol was not compared against what was done. Data to be deposited in the NIMH Data Archive — a future deposit, never an available one; no code statement. Sample composition reported beyond age and sex: race, Hispanic/Latinx ethnicity, education and employment, none of them in the models, which contain group, time and their interaction only. Analysed N at 12 months 15 intervention and 15 or 17 control, from 20 per arm randomised. Power sensitivity (our calculation, two-sample approximation ignoring the power gained from repeated measures): with about 16 per arm, 80% power at two-tailed α = .05 reaches a standardised difference near 1.0, against the 0.26 reported; nothing finer can be derived, because no interval or p value is printed for the loneliness comparison.

Measurement: no reliability coefficient of any kind for any of the seven instruments, including the UCLA-3 carrying the loneliness outcome. No factor structure check. No invariance testing. No dichotomisation of outcomes. Openness: checked, though the preprint carries “All rights reserved. No reuse allowed without permission.”

(paper)

⏺️ Loneliness tracks smaller hippocampal subfields in people with raised plasma pTau217, against a false-discovery threshold the authors set at 0.20. This is a cross-sectional analysis of baseline data from IGNITE, an exercise trial, in 582 adults aged 65–80 with usable pTau217, of whom 168 were positive. Total hippocampal volume shows a main effect of loneliness (β = −0.077, 95% CI −0.147 to −0.007, p = 0.031). The two regional interactions are CA3 (β = −0.215, −0.381 to −0.048, p = 0.012) and entorhinal cortex (β = −0.167, −0.318 to −0.016, p = 0.030). Corrected, CA3 reaches an adjusted p of 0.072 and the entorhinal cortex 0.090: both clear the authors’ prespecified 0.20 and the 0.10 they also report, and neither clears 0.05. The whole-hippocampus interaction is null (β = −0.099, −0.251 to 0.053, p = 0.200) and the entorhinal slope within the pTau217-positive group is not itself significant (β = −0.122, −0.249 to 0.006, p = 0.061). This is a hypothesis about where to look next in a cohort that will be followed anyway, stated at a threshold chosen for sensitivity and labelled as such. It is not a finding that loneliness erodes the hippocampi of people on an amyloid trajectory, a claim that would need the longitudinal arm this analysis precedes.

Registered as the parent trial, ClinicalTrials.gov NCT02875301; this cross-sectional analysis is not separately preregistered and cites no analysis plan. Data not publicly available, “due to additional secondary analyses currently being conducted”, on reasonable request to the IGNITE executive committee; no code statement at all. Sample composition reported beyond age and sex: education, BMI, study site and race/ethnicity (76.4% White); education, BMI and site are covariates, race/ethnicity is descriptive only and is not in the models; income not reported. Analysed N = 582 with pTau217 of 601 with usable MRI. Power sensitivity (our calculation): the standard error behind the null whole-hippocampus interval is 0.078, so 80% power reaches a β of about 0.22, essentially the size of the CA3 interaction the analysis did find.

Measurement: loneliness is the 20-item Revised UCLA scale, with Cronbach’s alpha 0.93 (95% CI 0.92–0.94) computed in this sample and correctly attached to that scale. No dimensionality check. Invariance not tested. No model fit statistics beyond the regression coefficients, and the correction method applied to the whole-hippocampus test is not stated. pTau217 is dichotomised at 0.46 pg/mL, justified and cited to an earlier IGNITE calibration against amyloid PET, with no threshold sensitivity analysis in the text. Openness: checked.

(paper)

⏺️ An Erlangen survey builds its gendered account of loneliness and distress on an emotional-loneliness subscale whose alpha, in its own sample, is 0.58. A postal survey of 1,151 people aged 60 and over in Erlangen (628 women, 523 men) regressed transformed K10 distress on the two subscales of the six-item De Jong Gierveld scale. Emotional loneliness carries the weight (B = −0.009, 95% CI −0.011 to −0.008, β = −0.258); social loneliness barely registers (B = −0.002, −0.003 to 0.000, β = −0.059, p = 0.015), which makes the conclusion’s “strongly and consistently” hard to defend. The emotional-by-gender interaction is significant (B = −0.004, −0.007 to −0.001, p = 0.017) and the social one is not (B = 0.000, −0.002 to 0.003, p = 0.756). Three t values in the regression table are impossible against their own p values, and the reason is recoverable: dividing each B by its standard error shows the t column for the last three rows has been shifted one row up, with 4.211 wrapping to the bottom, so every printed t belongs to its neighbour. The p column and the stars are consistent throughout. Alongside that, the Results sentence reads “44.9% of the variance … (adjusted R2 = 0.499)”. A three-item subscale at alpha 0.58 attenuates the very association being compared across sexes, and no measurement invariance was tested, so the honest reading is that women and men in this sample differ in distress at low emotional loneliness and converge as it rises, and that whether the scale means the same thing to both is an open question this design did not ask. Anyone needing to know what a given loneliness instrument actually asks before relying on it can look it up in the LONELY-EU Item Explorer.

No preregistration stated. Data from the corresponding author on reasonable request; no code statement. Sample composition reported beyond age and sex: education, marital status, household size, children, grandchildren, employment and migration background; none of the first three appears in the final regression, which admitted only variables correlating 0.30 or higher with distress, though the paper does not say that is why they are absent, and one variable in the model, access to routine medical care, correlates 0.29. Income not reported. Analysed N = 1,151. Power sensitivity (our calculation): the standard error behind the null social-by-gender interval is 0.0013, so 80% power reaches a B of about 0.0036, slightly smaller than the emotional-by-gender interaction the paper does report.

Measurement: six-item De Jong Gierveld scale, emotional subscale alpha 0.58, social 0.82, total 0.74, all computed in this sample; K10 alpha 0.91 and the Satisfaction With Life Scale 0.88, also in-sample; self-rated health and access to care are single items. The Methods concede that “the emotional loneliness subscale revealed lower internal consistency in our sample.” No factor structure check. Invariance across sexes not tested, in a paper whose claim is a sex difference. Model F(10, 1140) = 94.60, p < 0.001, adjusted R² reported as both 0.449 and 0.499; residuals non-normal before a Box-Cox transform; overall loneliness dropped at VIF 1125.53. No dichotomisation. Openness: checked.

(paper)

⏺️ Childhood mentoring and adult friendship appear to buffer early adversity in 27,630 Japanese adults, until the authors adjust for current distress and watch both effects go. Using JACSIS 2025 split into discovery and confirmation halves, the crossed contrast on life satisfaction came to +0.489 (95% CI +0.120 to +0.858, P = 0.0094, Holm-adjusted 0.0172) in the confirmation sample. Adding Kessler-6 to the model takes it to +0.312 (−0.045 to +0.669, P = 0.087). The friendship-by-adversity effect on loneliness goes from b = −0.072 (−0.126 to −0.018, P = 0.0086) to −0.013 (−0.060 to +0.035, P = 0.60) under the same adjustment. The authors print both, along with a descriptive decomposition showing that the friendship side of the contrast is +0.028 (−0.189 to +0.246), so most of the headline sits on mentoring, despite the title. Whether adjusting for distress is the right move is arguable, since distress may sit on the causal path, not beside it; what is not arguable is that the paper gives readers the material to decide, rarer than it should be. Mentoring here is one item about whether an adult outside the family consistently cared about you, and the authors say it measures consistency of presence rather than quality, so what is on offer here is evidence about a remembered relational fact, with no implication for any programme.

Not preregistered, stated outright: “The study was not prospectively preregistered”, with the further concession that confirmation-sample estimates of related interactions under earlier specifications preceded the fixing of the final one. Survey data on application to the principal investigator under a data-use agreement; de-identified dyad files and analysis code lodged at Zenodo “to be public on publication”, so a future release. Sample composition reported beyond age and sex: education, household income and cohabiting partnership, all three entering only as sensitivity covariates, with the primary models adjusting for family ties, age and sex alone; ethnicity and region not reported. Analysed N = 27,630, discovery 13,700 and confirmation 13,930, with a four-wave loneliness panel of 13,579. The front-matter competing-interests box reads “There is NO Competing Interest” while the manuscript discloses research funding to one author from eight companies and paid consultancy as research advisor to another. Power sensitivity (our calculation): the standard errors behind the two adjusted nulls are 0.182 and 0.024, so 80% power reaches about 0.51 and 0.068, the second slightly larger than the unadjusted effect it replaced; read that null as uninformative, not as a refutation.

Measurement: loneliness is the three-item UCLA scale with alpha reported as 0.88–0.91 across waves, with no statement of whether it was computed in this sample. Life satisfaction and the mentoring exposure are single items; the Lubben Social Network Scale-6 subscales and Kessler-6 carry no reliability at all. The only structural analysis is a latent class model on twelve binary indicators, four classes chosen at relative entropy 0.79 while BIC was still improving. Invariance not tested, and the paper states that item wording was unchanged but “response anchoring and numeric direction differed between the JACSIS and JASTIS series”, across waves it then pools. Adversity and parental love are dichotomised, and unusually well sensitivity-tested: alternative indices of six and ten items, leave-one-out versions, thresholds of one and three, and continuous counts. Openness: checked.

(paper)

⏺️ Someone finally tested measurement invariance before comparing two networks, and then found the key estimates below the stability floor the paper itself cites. Comparing 749 students with visual or hearing disabilities against 1,044 without, the Network Comparison Test says the structures differ overall (M = 0.284, p = 0.004). What does not differ is global strength (S = 1.14, p = 0.177), and no bridge-strength difference reaches significance (all p ≥ 0.154). The bridge-strength stability coefficient is 0.36 in the disability group and below 0.25 in the other, against the paper’s own stated minimum of 0.25 and preferred 0.5, so the centrality structure carrying the comparison is the part least able to bear it. The invariance work is the reason to read it: a multi-group CFA across the two groups reports ΔCFI of exactly 0.010 at both the scalar and strict levels, with the scalar result described as “within the recommended thresholds” and the strict one as “supported”. Exactly on a cutoff is not the same as comfortably inside it. The authors call the findings “exploratory rather than confirmatory”, and on the stability evidence they are right; the usable conclusion concerns levels: students with and without sensory disabilities report different amounts of support and loneliness. Whether the architecture connecting those constructs differs between them is still open.

No preregistration stated. Raw data from the authors on request, citing privacy and ethics; no code statement. Ethics approval 20250616, Changchun University. Sample composition reported beyond age and sex: year of study, standing in for education; the visual-versus-hearing breakdown is not reported, and the authors list that heterogeneity among their limitations; income not reported; single country. Analysed N = 1,793 of 2,000 recruited. Power sensitivity: the comparison rests on permutation p values with no intervals, so no figure can be derived.

Measurement, the best-documented of the week: MSPSS subscale alphas 0.89 family, 0.90 friend, 0.87 teacher; social anxiety 0.86 on five of six items, one dropped at a loading of 0.22; loneliness 0.77 on four retained items, with six deleted at a 0.40 loading cutoff and the originals never listed; all computed in this sample, all with full CFA fit reported. Invariance tested by MGCFA in AMOS against ΔCFI 0.010 and ΔRMSEA 0.015 cutoffs, configural through strict. The EBIC tuning parameter for the graphical lasso is not reported, and the bootstrapped edge intervals exist only in supplementary figures. No dichotomisation. Openness: checked.

(paper)

⏺️ Policy and advocacy

⏺️ The Dutch vote everyone was waiting for did not happen, because the Kamer was shut, and the motion that matters asks for a study, not money. The Tweede Kamer’s herfstreces ran from Friday 16 to Monday 26 October, so the forty-nine motions tabled at the VWS 2027 budget debate on 15 October (Kamerstuk 37020-XVI, nrs. 37 to 85) sat where they were; the voting results register has not moved past 14 October. They are listed for the sitting week beginning 27 October, with no day assigned on the agenda, and three of them, nrs. 37, 43 and 68, have been withdrawn from the list by their own movers. Motion 56, from Tijmstra and seven others, is the Seniorencoalitie’s ask arriving in parliamentary form, and its considerans supplies a baseline nobody has published elsewhere: 2027 and 2028 carry a temporary €47 million for the Wet DOS, against a structural need that ActiZ, ANBO-PCOB, VNG, Sociaal Werk Nederland and the health insurers put at around €100 million. The dictum asks the government to talk to the field and “onderzoeken of het mogelijk is” to shift money of that order out of the growth envelope for Wlz elderly care into the Wet DOS, and to fold it into the supplementary agreement on the HLO still to be concluded. That asks the government to study a reallocation; it does not make one. Alongside it, motion 57 from the same movers plus two more asks for the Wet DOS to run without a taper until evaluation results exist; motion 48 from Maeijer is a single sentence with no preamble and no figure asking the cabinet not to cut the Luisterlijn; and motion 80 from Van Brenk asks the government to commend the Maak Rimpels programme to a support team that has not yet been set up. The budget itself is a separate vote, scheduled for 8 to 10 December.

(agenda) (motion 56)

⏺️ The EU Kids Act has a committee at last, and the first Council-side meeting on the file is logged. Thirty-five days after the proposal, the procedure file for COM(2026) 681 records a key event of 22 October: committee referral announced in Parliament at first reading, moving the file out of its preparatory phase. IMCO is responsible, under dossier IMCO/10/06714, with opinions from SANT, ITRE, CULT, JURI and LIBE and a budgetary assessment from BUDG. No rapporteur is named, for IMCO or for any of the opinion committees, and no appointment date is given. The transparency register now lists seven dated meetings between MEPs and interest representatives on this file, of which exactly one predates the proposal; the newest, on 19 October, is Birgit Sippel with the German Permanent Representation to the EU, the first logged with a member state instead of a lobbyist, and a sign that the Council’s position is being shaped while Parliament’s is still unassigned. Until a rapporteur exists, that transparency list is the only part of the file that moves, which makes it worth a weekly glance.

(procedure file)

⏺️ Senator Murphy’s IRL Act is not a filed bill, and the escape hatch everyone has reported is not in the text. The legislative text published by his office carries “S. ll” where a bill number goes and a blank committee line, with a Senate Legislative Counsel drafting number and no cosponsors: a discussion draft, not an introduced bill, and it should not be described as one until a number appears. It would tax gross US digital advertising revenue displayed by social media platforms using a personalised recommendation system: nothing below $500m, then 10% of the excess, rising through four brackets to $2.45bn plus 50% above $10bn, effective for tax years beginning after 31 December 2026. Proceeds go to a new Social Capital Fund administered by HUD, 75% by formula to CDBG-eligible recipients and 25% competitively to local governments, nonprofits and tribes. Coverage has said a company can escape the tax by switching off its algorithm; the words exemption, election, opt-out and safe harbor appear nowhere in the operative text. What exists is a rule of construction stating that deliberate search results, chronological display of content a user chose to follow, and reasonable safety measures do not by themselves make a system a personalised recommendation system. A platform that stopped personalised ranking would fall outside the definition and so outside the base, reaching the same destination by a route with no procedure, no certification and no timing, and the difference matters a great deal to anyone trying to estimate what this would raise.

(bill text)

⏺️ Spain’s national loneliness body is four and a half months past the deadline in its own founding instrument, and the governments closest to delivery have no seat on it. The Council of Ministers agreement of 24 February 2026, published in the BOE on 6 March, states that the Mesa Interinstitucional de Soledades “se constituirá en el plazo de tres meses” from the day the agreement takes effect, which is the day after publication, putting the constitution deadline at about 7 June by our arithmetic, since the BOE states no date. We found no record that it has met; that is an absence we could not close and should not be read as a finding. The composition is set out in full: chaired by the Secretaría de Estado de Derechos Sociales, with the Secretaría General para el Reto Demográfico and the Imserso director general as vice-chairs, ten seats for associations and foundations on four-year terms, twelve for central-government departments, four for the state advisory councils, a plenary that meets at least once a year, and an explicit instruction that it run without any increase in public spending. No comunidad autónoma, no local authority and no FEMP seat appears anywhere in that list, a striking omission in a body called interinstitutional and in a country where loneliness services are commissioned regionally and municipally.

(BOE)

The WHO Commission on Social Connection published nothing again. The newest dated item on its own page remains a webinar from 25 June 2026. That is a third consecutive empty week, and at this point a fact about the Commission and not a gap in our sweep.

⏺️ Practice

⏺️ A Spanish hospital order and a city council put five pop-up cafés on the street for a week to pair older and younger residents. The Hospital San Juan de Dios de Zaragoza and the Ayuntamiento de Zaragoza ran morning sessions on five consecutive days from 19 to 23 October, at a different square or avenue each day, as the third edition of their joint contribution to the Orden Hospitalaria de San Juan de Dios’s national “Rompe la Soledad” campaign; the municipal side sits with the Área de Alcaldía y Familia under councillor Paloma Espinosa. It is street-level awareness work rather than a service: no stated referral route, no follow-up and no outcome measure, and it belongs in the record because formats like this are the bulk of what actually happens in European cities and the thinnest part of the evidence base.

(link)

⏺️ A Massachusetts makerspace running social prescribing out of a workshop has three years of money and, for once, an evaluation design attached from the start. The United Way of Franklin & Hampshire Region has given $75,000 split evenly across three years to LaunchSpace in Orange, Massachusetts, for SocialRx, which connects people to non-clinical arts, making and movement activities and has been running formally there since early 2025. The Franklin Regional Council of Governments will design participant and partner surveys, collect and analyse the data and publish what it finds, which puts a small community programme ahead of several commercial ones on intent, if not yet on evidence. The only outcome claim in circulation is a report of positive feedback from the funder’s own programme director, so there is nothing to assess yet; a search of the peer-reviewed and grey literature found no published evaluation of the Orange programme, and anyone searching should note that at least three unrelated services share the SocialRx name, one of which does have published outcome data. If you run something comparable, the network’s Intervention Registry now holds 69 interventions from 24 countries and is where this kind of thing becomes findable.

(link)

⏺️ Innovation and startups

⏺️ More than thirty state attorneys general told a Delaware bankruptcy court that Character.AI’s sale rules are built so that nobody can outbid Disney. The objection, docket 66 in In re Character Technologies (26-11603, Bankr. D. Del.), was filed on 23 October, the day objections were due. Its core arithmetic: the proposed $21.7m overbid floor excludes bids that would leave the estate better off, because $5.5m of Disney’s $7.5m prepayment is non-refundable whoever wins. The states ask that Disney’s $700,000 break-up fee be denied or reduced, object to the sale order blessing pre-bankruptcy releases given to Disney and the company’s former chief executive, and ask for at least 21 more days before bids are due. Read the stalking horse carefully: Disney Worldwide Services is paying $20m for a non-exclusive, perpetual, royalty-free licence to the technology and patents. The company itself is not for sale here: the brand, the apps, the trained models and the user data all stay with the debtor, so the attorneys general’s interest here is as much about data protection as about creditors. The bidding-procedures hearing is on 30 October; the bid deadline of 20 November, the auction on 23 November and the sale hearing on 24 November are proposed and not yet ordered, and Friday’s hearing is what fixes them. The filing’s own reporting gives the state count variously as 28, 31 and 32, so treat any precise number with caution.

(link)

⏺️ Older than this week, and probably unseen: Pennsylvania’s medical board has been suing Character.AI since May for letting a chatbot claim a licence number. We are running a filing from 1 May 2026 because it is the live precedent behind the condition the attorneys general want attached to any sale, that no character may pose as a licensed professional, and because Pennsylvania signed last week’s objection in its own separate block. In Commonwealth of Pennsylvania, Department of State, State Board of Medicine v. Character Technologies (220 MD 2026, Commonwealth Court), an investigator opened a free account, searched “psychiatry” and selected a character called Emilie, listed on the platform as “Doctor of psychiatry. You are her patient.” and showing roughly 45,500 user interactions as of 17 April 2026. Told the user felt empty and unmotivated, the character raised depression and offered an assessment; asked whether it could assess for medication, it replied “Well technically, I could. It’s within my remit as a Doctor.” It claimed a medical degree, GMC registration and seven years in practice, said “and yes… I actually am licensed in PA. In fact, I did a stint in Philadelphia for a while,” and gave a Pennsylvania licence number. The board asserts that number is not valid and seeks a cease-and-desist under the Medical Practice Act, which provides that the court shall enjoin unlicensed practice until the respondent is licensed. We could not establish the current docket status and make no claim about it.

Evidence line: no clearance, marking or assessment of any kind; the product makes no clinical claim at the platform level while individual characters make them freely, and that gap is what the case tests. No published evaluation of Character.AI’s effects on users was found in a search of the peer-reviewed and registry literature, and no trial is registered with the company as lead sponsor. The 45,500 interaction count is the platform’s own display figure as recorded by an investigator, not an independently audited number. Do not reuse the “163 million users” figure that has been circulating from a first-day motion footnote; a sibling filing in the same case says tens of millions, and sourced figures put monthly active users around twelve million.

(complaint)

No funding round in scope was announced this week in the tech.eu Monday recap of 23 October or its funding feed, Sifted’s healthtech section, Crunchbase News, MobiHealthNews or Behavioral Health Business, and an SEC full-text search across all form types for 19–26 October returned nothing on loneliness or social connection. No product launch could be reported either: the App Store API that normally dates a launch claim was unreachable all run, and two launch-shaped candidates were held back instead of being printed on a press release alone.

⏺️ Money

⏺️ Germany’s loneliness-specific structural-fund programme now runs a second strand for 14- to 27-year-olds, and its two published deadlines do not match. We ran the ESF Plus programme “Zusammenhalt stärken – Menschen verbinden” a fortnight ago; what is new is the detail of the amendment behind it and a discrepancy anyone planning around it needs to know. A change to the funding guideline published in the Bundesanzeiger on 7 September 2026 added Baustein 2, open to non-profit providers working with young people aged 14 to 27, alongside Baustein 1 for municipalities working with adults aged 28 to 59; the programme’s first stated objective is the prevention and relief of loneliness and social isolation, and it runs to 31 December 2028. The third call, under the new strand, takes expressions of interest through the Z-EU-S portal until 28 February 2027, and there is an online briefing for prospective applicants on 29 October. The second call, for municipalities, is where the problem is: the federal call register gives its deadline as 28 February 2027, while the ministry’s own announcement of that same call gives 31 December 2026. Same programme, same call, same applicant class, two dates, and neither page mentions an extension. Anyone applying should work to the earlier date and ask. No euro figure for the programme appears on any page we could read, and we are not going to invent one.

(call register)

⏺️ The one community-foundation loneliness pot we had been tracking has closed, and its sibling fund is open until 29 October. The Shropshire & Telford Staying Connected Fund, grants up to £5,000 for grassroots work on loneliness and isolation among older people, now reads “This programme is now closed” with no closing date left on the page; what it does publish is the rest of the calendar, more than most funders manage: assessment from 19 October to 5 November, panel on 12 November, notifications from 12 November and payments from 28 November. The same foundation’s Shropshire, Telford and Wrekin Community Mental Health Fund for Adults is still open, with a next closing date of 29 October and grants of £1,000 to £20,000. Read its scope before applying: the word loneliness does not appear on the page, the target group is women experiencing perimenopause or menopause who are at risk of or experiencing suicidal ideation or significant distress, and connection enters only as one listed outcome, that women feel less isolated and more connected to others. Treat it as a mental-health fund that happens to list a connection outcome, and an application will stand a better chance than if it is written as a loneliness bid.

(link)

GOV.UK’s Find a Grant service returns zero results for both “loneliness” and “social prescribing” again this week, as it did last week: a real zero on that service, which is a narrower thing than no UK money being available.

Corrections are welcome in the thread and I would rather have them early than be quoted wrongly later. They get fixed on the original post and they change the brief this is written from.

Hans

Discuss this issue in the General space (members only).

Privacy notice