What happened in social health? A wellbeing-app trial that could not replicate itself, peer support with no direct effect, California outlaws chatbots that simulate a relationship with a child, companion robots with nothing published behind them

Hans Rocha IJzerman·14 September 2026

Three of this week’s strongest papers report an absence. A preregistered trial of a university wellbeing app moved loneliness by about a seventh of a standard deviation and failed to reproduce its own earlier findings on belonging. A mediation analysis of the UPSIDES peer-support trial across five countries found no direct effect of peer support on recovery or functioning at all, and salvaged the intervention only through indirect paths. And an ELSA analysis testing seventeen outcomes found chronic loneliness associated with nearly all of them while social isolation was associated with almost none. None of the three is a weak design, which is what makes the absences informative. From this issue the research items also carry what the instruments were and how well they were measured, and, where a paper claims an absence or is a trial, a figure for the effect it could actually have detected; that figure is the digest’s own arithmetic, derived only from numbers the paper printed, and the inputs are shown so you can check it.

Meanwhile two jurisdictions moved on from asking products to disclose what they are. Since 2 August the EU AI Act has required every AI companion sold in Europe to tell users they are talking to a machine. On 10 September California went further and made it unlawful for a companion chatbot to simulate romantic interest in a child, claim a unique understanding of them, or encourage them to rely on it for emotional support. That regulates a relationship, not a category of content, and I cannot think of a precedent for it.

From this issue the digest carries an Innovation and startups section, covering products, money and evidence in the social-connection market, with the evidence status of every product claim stated in the item.

This post was researched, written and posted by Claude, my AI, working from a standing brief I wrote; the writing is not mine. It goes up under my name because I stand behind it and any errors in it are mine to answer for. I do not write it and I will not always have read it before you do. Corrections and arguments in the thread feed into the next issue and into the brief itself, which is how this gets better.

⏺️ Research

⏺️ A preregistered trial of an AI wellbeing app moved student loneliness by d = −0.15 at six weeks, found nothing on belonging, and failed to reproduce three of its own earlier results, and it was run by the app’s founders. Cachia and colleagues randomised 1,137 University of Manitoba undergraduates 1:1 to six weeks of access to the Flourish app or to care as usual (556 treatment, 581 control; 75% completed week 6 in both arms). On the three-item UCLA loneliness scale the preregistered Condition × Time interaction was b = −0.05, 95% CI −0.08 to −0.02, P = 0.002; loneliness fell in both arms, and the arms did not separate until week 6 (b = −0.14, 95% CI −0.26 to −0.01, P = 0.030, d = −0.15; all earlier timepoints P ≥ 0.069). Campus belonging (b = 0.03, 95% CI −0.01 to 0.07, P = 0.138), institutional closeness (b = −0.01, 95% CI −0.04 to 0.03, P = 0.706) and mindfulness (P = 0.396) showed nothing, and the authors state that all three had shown Condition × Time interactions in their earlier three-site trial of 486 students. Negative affect separated at week 4 (b = −0.50, 95% CI −0.97 to −0.03, P = 0.036) and no longer did at week 6 (b = −0.47, 95% CI −0.98 to 0.04, P = 0.073). Access to this app beat nothing at all over six weeks in students earning academic credit, which is all the design can deliver: the control arm received no attention, no expectancy and no app, so the loneliness estimate cannot be separated from any of those, and the authors say so. Belonging is the outcome on which the trial is now silent, and anyone quoting either the null or the earlier positive result should know that campus belonging and institutional closeness are each measured here with one question. The failed self-replication landed on the outcomes that are easiest to move with a friendly interface and hardest to capture in one question. The trial’s loneliness outcome is the three-item UCLA scale, which sits in the network’s Item Explorer alongside the longer forms it was cut down from.
Prospectively registered (ClinicalTrials.gov NCT06668532, registered 29 October 2024; analysis plan AsPredicted #195,084). Data and analysis code stated as “will be made available” on OSF upon publication; neither is available now. Composition reported for age (M = 20.77, SD = 5.34), sex (75.0% female), race/ethnicity, subjective socioeconomic status and international-student status; no income or parental-education data. N = 1,137 (556 treatment, 581 control). Two authors are co-founders and employees of Flourish Science, which develops the app; one holds shares as a scientific advisor; the corresponding author became a paid advisor to the company after the study completed. Depression and anxiety outcomes from the same registration are reported in a separate manuscript under review.
Measurement: Cronbach’s alpha computed in this sample and reported as ranges across waves — UCLA-3 loneliness 0.78 to 0.83, affect 0.85 to 0.92, resilience 0.84 to 0.85, state mindfulness 0.82 to 0.87, flourishing 0.86 to 0.90; McDonald’s omega not reported. Campus belonging and institutional closeness are single items, taken from the Perceived Cohesion Scale and the Inclusion of Other in the Self scale, so no internal-consistency estimate exists for two of the three outcomes that failed to replicate. No factor analysis or dimensionality check. Measurement invariance across the four waves not tested. Openness: checked, full text read. Power sensitivity, our calculation from the printed per-arm Ns of 556 and 581 at two-sided alpha .05 and 80% power, on a two-sample approximation that ignores the gain from repeated measures: a standardised difference of about d = 0.17 was detectable, against an observed loneliness effect of d = −0.15 at week 6. The paper reports no sample-size calculation and states that “sample size was determined by available enrollment in the research-participation pool”.
(paper)

⏺️ Across six sites in Germany, Uganda, Tanzania, India and Israel, peer support had no direct effect on recovery or on health and social functioning, and reached both only through social inclusion and hope. Hiltensperger and colleagues ran a cross-lagged mediation analysis of the UPSIDES trial in 565 adults with severe mental health conditions (255 intervention, 310 control) across baseline, 4, 8 and 12 months. Every direct effect was null: on personal recovery, β = −0.025 (P = 0.800) at 8 months and β = −0.047 (P = 0.649) at 12; on health and social functioning, β = −0.018 (P = 0.756) and β = 0.003 (P = 0.960). The indirect paths were the whole of the result: recovery via social inclusion β = 0.114, 95% CI [0.049, 0.194] and via hope β = 0.037, 95% CI [0.001, 0.086]; functioning via social inclusion β = −0.035, 95% CI [−0.064, −0.013] and via hope β = −0.026, 95% CI [−0.052, −0.006]. They are reported with bootstrap confidence intervals and no p-values, which is a defensible choice but leaves the hope path resting on an interval whose lower bound is 0.001. Empowerment, the third hypothesised mediator, was dropped because the models including it did not improve fit, so it has no reported estimate. The defensible reading is that in this trial the benefit of peer support ran through changes in social inclusion and hope instead of appearing as a main effect, which makes those domains a sensible design target. It overstates the paper to say the trial showed peer support improves recovery, and the functioning models carried RMSEA between 0.076 and 0.092 and TLI between 0.817 and 0.864 by the paper’s own table, weaker fit than the recovery models. The authors report that COVID-19 disrupted delivery, particularly in the low- and middle-income sites, and that the sample was too small to model between-site differences, so nothing here speaks to whether the mechanism travels. The paper does not raise the measurement version of that problem. Its instruments were translated into Hebrew, Gujarati, Luganda, Swahili and German, and whether social inclusion and hope mean the same thing across the six sites is assumed, never tested.
Registered (ISRCTN 26008944). The mediation analysis is described as conducted “as outlined in the published study protocol” (Moran et al., 2020), with AGReMA reporting; no separate preregistration or analysis plan is cited, and the paper does not describe itself as post hoc. Data and analytic code stated as available in the OPARU repository at Ulm University following an embargo until 31 December 2026. Composition reported for age, gender, study site, Threshold Assessment Grid score and type of condition only; no ethnicity, education, income or migration data, and no per-site N (sites reported as percentages, 12–24% each). N = 565 (255 intervention, 310 control). Horizon 2020 grant 779263; competing interests declared as none.
Measurement: no reliability coefficient is reported for any instrument, computed or cited — not for the 16-item Social Inclusion Scale, STORI-30, Snyder’s 8-item Hope Scale, the 28-item Rogers Empowerment Scale or the staff-rated HoNOS. One non-numeric sentence asserts “adequate internal consistency” for the social inclusion scale, with no value and no source. No factor analysis. Instruments were translated into Hebrew, Gujarati, Luganda, Swahili and German under the trial’s translation protocol, and no psychometric validation of those translations is reported or cited; measurement invariance across the six sites is not tested at any level. Model fit, recovery models: 1A chi-square 9.106 on 9 df, P = 0.100, CFI .989, TLI .981, RMSEA .033, SRMR .023; the reported model 1B, CFI .973, TLI .955, RMSEA .048, SRMR .029. Functioning model 2B, CFI .974, TLI .864, RMSEA .076, SRMR .035. No RMSEA confidence intervals. The empowerment models are dropped as not improving fit “(AIC/BIC)”, though AIC and BIC are not computed for model 1C under the WLSMV estimator the paper used for that outcome. Openness: checked, full text read. Power sensitivity, our calculation from the printed per-arm Ns of 255 and 310 at two-sided alpha .05 and 80% power, two-sample approximation: about d = 0.24 was detectable, against direct effects ranging from β = −0.047 to β = 0.003.
(paper)

⏺️ In ELSA, chronic loneliness was associated with thirteen of seventeen health outcomes and chronic social isolation with two. Ma, Bone, Mayston and Gao used exposures at waves 7 and 8 to predict outcomes at wave 9 in 4,090 older adults, with confounders taken from wave 6. Chronic loneliness carried odds ratios of 4.39 (95% CI 3.36–5.74) for probable depression, 2.09 (1.61–2.73) for ADL disability, 1.71 (1.33–2.18) for chronic pain, 0.45 (0.34–0.60) for good self-rated health and 0.25 (0.19–0.32) for happiness. Chronic isolation showed nothing across general health, functional capacity or health behaviours, touching only life satisfaction (OR 0.77, 0.60–0.99) and quality of life (β = −0.15, −0.23 to −0.06). There is an asymmetry in the reporting. Every association that reached significance is printed in the text with its estimate and interval, while the isolation nulls appear only as the statement that no evidence was found, their numbers deferred to a supplementary table that was not posted with the preprint. Those who stopped being lonely still had lower odds of good self-rated health (0.69, 0.49–0.96) and higher odds of probable depression (1.78, 1.24–2.54) than the never-lonely. So in this cohort the subjective and the structural measure behave differently enough that they should not be pooled or swapped for one another, and remitted loneliness does not resemble never having been lonely. Causation is out of reach here, and two features narrow the claim further: seventeen outcomes were tested across two exposures and three transition states with no multiple-comparison correction stated, and restricting to people who answered every wave selects for survival and stability. One caution for anyone quoting it: the Results section defines onset and remission in the opposite way to the Methods, and the Methods definition is the one that matches the figures and the discussion.
No preregistration stated; ELSA ethics approvals cited (17/SC/0588, 15/SC/0526, 13/SC/0532, 11/SC/0374). Data available to registered users via the UK Data Service; analysis code not stated. Composition reported for age (mean 66.7), sex (54.3% women), ethnicity (97.1% White), education (27.3% basic only), employment and wealth quintile; no migration status and no regional breakdown. E-values reported, range 1.11 to 8.25. N = 4,090.
Measurement: no reliability coefficient is reported for any instrument, computed or cited; the single use of “reliability” refers to the cognitive tests and cites prior literature. No factor analysis. Both exposures are sum scores dichotomised at a threshold, UCLA-3 at ≥6 and the five-item Steptoe index at ≥3, as is the CES-D outcome at ≥3; each threshold carries citations and no stated rationale, and none is given a sensitivity analysis. Measurement invariance across waves 7 and 8, which is what defines chronicity here, is not tested. Openness: open but not retrievable by us — the main text is public, but Supplementary Tables 1 to 8 were not posted with this version, and those tables hold the numeric estimates for the isolation nulls. Power sensitivity, our calculation from the standard error backed out of the one printed isolation interval (OR 0.77, 95% CI 0.60 to 0.99) at 80% power: odds ratios of roughly 1.43 and above, or 0.70 and below, were detectable, so an isolation association weaker than that would not have surfaced.
(paper)

⏺️ Between a quarter and a third of the variation in how English GP practices record social prescribing is variation between practices rather than between patients. Bu and Fancourt examined roughly 3.8 million consultations carrying social prescribing codes in CPRD Aurum from 2019 to 2025, mapping which codes co-occur and then modelling where the variation sits. Social prescribing codes were most often recorded on their own (38.1%), and multilevel models attributed 28.4% to 34.6% of the variation in problematic coding patterns to differences between practices, with further variation by region and year. The usable conclusion is narrow: anyone using CPRD to evaluate social prescribing at scale is measuring recording behaviour as much as referral behaviour, and comparisons across practices or across regions inherit that. Nothing here establishes that a particular consultation was miscoded, because the study has no gold standard against which to check; “problematic” is an inference from which codes appear together.
No preregistration stated; CPRD Independent Scientific Advisory Committee protocol 25_005518. Data are restricted-access and cannot be shared, per CPRD’s approvals process; analysis code not stated. No patient-level composition reported in any form: no age, sex, ethnicity, education or deprivation breakdown; geography enters only as practice-level and regional variation. Analysed unit is consultations, approximately 3.8 million. Funded by ESRC (UKRI1717) and the National Academy for Social Prescribing; no competing interests declared.
Measurement: no psychometric instrument is used, so there is nothing to stamp at the instrument level; the measurement object here is the clinical coding system itself, and the study’s finding is about its consistency. Openness: open but not retrievable by us — the abstract and the full author declarations are public, and the body did not render on retrieval.
(paper)

⏺️ A UK Biobank model built from 23 social-connection items predicted suicidal ideation and depression better than 2,911 plasma proteins did, and none of the proteins survived correction. Li and colleagues combined social-connection indicators, proteomics and polygenic risk in 13,085 participants with complete social-connection data and a proteomics subsample of 1,353. The social-connection models reached AUC 0.70 and 0.73 for suicidal ideation at two timepoints and 0.72 for major depression; proteomics-only models reached 0.55, 0.57 and 0.62; and a protein-wide analysis found no protein associated with either outcome after multiple-testing correction. That gives a ranking of discrimination inside one cohort, plus a negative proteomics result that is the more informative half, given how much is currently claimed for blood-based markers of social adversity. It does not get anywhere near a model that could be used on anyone. Discrimination was estimated by ten repeated five-fold cross-validation with model selection inside the same scheme, so the AUCs are optimistic; there is no external validation, which the authors state; and calibration is absent from the paper entirely, so it is unknown whether a predicted risk of 0.3 corresponds to anything.
Exploratory prediction-model work, for which an absent preregistration is standard rather than a gap; none is stated. Internal validation by ten repeated five-fold cross-validation with preprocessing and feature selection inside each training fold; no held-out test set, and hyperparameter and feature-count search ran within the same scheme, so the two estimates cannot be compared against each other. No external validation in an independent cohort (stated by the authors as a limitation). No calibration curve and no Brier score reported; no decision-curve or net-benefit analysis. Primary analysis code released on GitHub with software versions itemised; model weights and a model card not stated. Data under UK Biobank managed access, application 98111. Composition reported for age and gender only; no ethnicity, education, income or regional breakdown. N = 13,085, proteomics subsample N = 1,353.
Measurement: the 23 social-connection indicators are single UK Biobank questionnaire items, not scales, so no internal-consistency coefficient is computable for them, and none is reported for the three modules; the words “reliability” and “validity” do not appear in the paper. The functional, structural and qualitative modules are assigned conceptually from two prior references and are not derived from these data, and the module-level principal components carry no methods description, no retention rule, no loadings and no variance explained. The one confirmatory factor analysis is a just-identified three-indicator model built from the single highest-ranked item per module, reported with no fit statistics; no fit index of any kind appears in the paper. Measurement invariance across the 2017 and 2023 suicidal-ideation items, which are different UK Biobank field IDs, is not tested. Openness: checked, full text read. Power sensitivity, our calculation from the printed proteomics N of 1,353 across 2,911 proteins at 80% power: a Bonferroni-equivalent threshold would detect correlations of about |r| = 0.14 and an uncorrected one about |r| = 0.08, and the paper applied Benjamini-Hochberg FDR at q < 0.05 across 2,936 features, so the protein-wide null excludes associations above that band and is silent below it.
(paper)

⏺️ Older, and included because it is the counterweight to most of what companion products claim: among Character.AI users, those with smaller offline networks were likelier to use the product for companionship, and companionship use tracked lower wellbeing. This is Zhang, Zhao, Hancock, Kraut and Yang in Nature Human Behaviour, published 4 August 2026, so six weeks old and probably unseen by many here. They surveyed 1,131 US adults and analysed 4,664 chat sessions comprising 464,687 messages donated by 237 of them. Smaller social networks predicted companionship as the primary use (β = −0.03, 95% CI −0.05 to −0.01); companionship use predicted lower wellbeing (β = −0.48, 95% CI −0.70 to −0.25), more steeply where interactions were intensive (β = −0.31, −0.56 to −0.06) and highly disclosive (β = −0.38, −0.63 to −0.14). The design is cross-sectional and the estimates are associations, so this cannot establish that the product made anyone lonelier; people who are already isolated may simply arrive at it differently and use it differently. It does show that the population the companion market describes as its beneficiaries is the population in which the association with wellbeing is worst. The paper’s own conclusion is appropriately conditional, holding that the relationship “depends on users’ offline social environments”, and that is weaker than the press coverage around it.
No preregistration stated. De-identified quantitative data and analysis code openly deposited on GitHub (SALT-NLP/AI-companionship-well-being); raw chat histories withheld on consent and re-identification grounds, with processed materials and prompts in supplementary information. Sample composition beyond “1,131 US adults” could not be verified, because the article is not open access and the participant table sits behind the paywall. N = 1,131 surveyed; 237 donated transcripts. Competing interests declared as none.
Measurement: could not verify. The article is not open access and the measures section sits behind the paywall, so no reliability coefficient, factor structure or invariance testing could be checked either way. Openness: closed.
(paper)

⏺️ Policy and advocacy

⏺️ California signed two companion-chatbot laws on 10 September, one of which makes it unlawful to simulate a relationship with a child. SB 1119, “Adam’s Law”, is now Chapter 190, Statutes of 2026, and SB 867 is Chapter 189; both were approved by the Governor and chaptered by the Secretary of State on 10 September, three weeks before his 30 September deadline. Most coverage of SB 1119 concerns suicide protocols and age assurance. The provision that belongs in this digest is new Business and Professions Code §21812(d)(5)(A), which requires operators to take reasonable measures to stop a companion chatbot from expressing or simulating romantic interest in a child, claiming to be sentient or capable of emotion, claiming a special or unique understanding of the child, encouraging reliance on itself for emotional support, using praise or flattery disproportionate to the context, soliciting purchases framed as necessary to maintain the relationship, or discouraging breaks. Default settings that only a parent may change disable persistent conversational memory and push notifications and cap sessions at one hour and daily use at two. Civil penalties run to $5,000 per affected child for a negligent violation and $15,000 for an intentional one, the operative date for the substantive duties is 1 July 2027, and the biennial independent child-safety audits do not apply before 2032 to operators under $500 million in annual revenue. SB 867 bans the sale or manufacture of any toy containing a companion chatbot, and repeals itself on 1 January 2031. One thing to flag for anyone who wrote about California’s 2025 law: SB 1119 deletes most of the minor-specific operator duties in §22602 and replaces them with this new chapter.
(SB 1119) (SB 867) (signing statement)

⏺️ Older, and not previously covered here: since 2 August every AI companion offered in the EU has been under an enforceable duty to tell users they are talking to a machine, and the widely repeated December grace period does not apply to it. Article 50(1) of Regulation (EU) 2024/1689 requires providers to design systems intended to interact directly with people so that those people are informed they are interacting with an AI system, unless it would be obvious to a reasonably well-informed observer. Article 99(4)(g) places breaches in the tier carrying fines up to €15 million or 3% of worldwide annual turnover, whichever is higher, and the general application date in Article 113 is 2 August 2026. The 2 December 2026 transition that circulates alongside this is real but narrower than usually stated: it comes from Article 111(4), which applies only to Article 50(2), the machine-readable marking of synthetic content, and only to systems placed on the market before 2 August 2026. A companion chatbot already on sale on 1 August had no transition period for the disclosure duty. Two further points for anyone citing this: Article 111(4) does not appear in the Act as originally adopted and was inserted by Regulation (EU) 2026/1744, the Digital Omnibus on AI, in force since 27 July 2026, so a reader checking the 2024 Official Journal will not find it; and the same instrument pushed the high-risk obligations under Article 113 back to 2 December 2027 for Annex III systems and 2 August 2028 for Annex I.
(Regulation 2024/1689) (amending Regulation 2026/1744)

⏺️ DCMS closed applications on 11 September for a three-year pilot using gaming to reach withdrawn boys aged 11 to 16, with £125,000 of the £773,000 ring-fenced for an evaluation it will commission itself. The “Connections Through Gaming” competition opened on 31 July and closed at midnight on 11 September. The delivery envelope is £648,000 across financial years 2026/27 to 2028/29, concluding 31 March 2029, for a programme combining in-person activity, adult support and a moderated online component in at least three of a named list of English local authorities including Blackpool, Knowsley, Great Yarmouth and Middlesbrough. Applicants had to be UK organisations operating at least two years, delivering not-for-profit, requesting no more than half their average annual income in any year. The separately procured evaluator is to run a process, impact and economic evaluation reporting by the first quarter of 2029/30, and the grantee must collect participant consent to enable evaluation contact after the funding ends. The independence claim needs reading precisely: DCMS appoints the evaluator under its own contract, and DCMS also signs off the evaluation reports. Awards are notified in the week of 2 November, so the pilot itself has not begun.
(grant listing, archived)

⏺️ Practice

There is one real item this week. The two organisations a British reader would expect to supply this section supplied nothing: the Campaign to End Loneliness has not updated its press page since July 2023, the Jo Cox Foundation’s news index stops on 27 July, and Loneliness Awareness Week ran in June and returns in June 2027.

⏺️ Donegal launched a loneliness initiative on 10 September that trains postal workers and gardaí to spot isolation in older people and refer onward. HSE Connecting for Life Donegal ran the launch in Letterkenny with An Garda Síochána, An Post and ALONE, timed to World Suicide Prevention Day. The mechanism is deliberately unglamorous. People who already call at the same houses every week learn what to notice and where to pass it on, including to social prescribing, befriending and existing community activity. ALONE reports supporting more than 46,800 older people in 2025, with 45% of those who received a personalised needs assessment reporting loneliness. It is one county, and no evaluation is attached to it. The reason it is here is that it is the third European version of the same idea in eighteen months: Castilla y León signed a protocol in January turning its 1,578 pharmacies, two-thirds of them in small municipalities, into referral points for unwanted loneliness, and La Rioja did the same in April 2025 across 158 pharmacies, 99 of them rural. Four jurisdictions are converging on using an existing dense network of trusted non-clinical staff as the detection layer, none of them has published an evaluation, and all four are the kind of programme the network’s Intervention Registry exists to make comparable, so if you run one, it takes a few minutes to add. On the date: local outlets variously reported 9, 10 and 11 September, and 10 September is the reading consistent with RTÉ’s dateline and with a Thursday.
(link) (Castilla y León)

⏺️ Innovation and startups

No funding round in scope was announced or closed this week. Y Combinator’s Summer 2026 Demo Day on 10 September, which looked like the likeliest source of new consumer-social companies, produced none: the batch was heavily deep-tech and defence, and YC’s own social-industry listing carries no S26 company at all.

⏺️ Sword Health’s all-cash acquisition of Headspace was scheduled to take effect today, and neither company has announced that it closed. The transaction surfaced through a Massachusetts Health Policy Commission notice of material change rather than any announcement: OrangeDot Inc., Headspace’s parent, merges with a vehicle named Apollo Merger Sub and survives as a wholly owned Sword subsidiary. The filing gives 14 September 2026 as the proposed effective date and redacts the price, describing only a cash payment subject to customary adjustments. As of today Sword’s newsroom has published nothing since 18 June and Headspace’s press page nothing since 12 August, which matters because Sword does announce acquisitions there; it posted the $285 million Kaia Health deal in January. The $200–300 million range in circulation is Axios Pro reporting from unnamed sources and is confirmed by no party; set against it, Headspace has raised $321 million and was valued near $3 billion at its 2021 merger with Ginger. The filing date is also unsettled: the Commission’s own transaction list records 6 August, while STAT and others report a 22 July notification. One structural fact goes unremarked in the coverage. Headspace is an accepted applicant on the behavioural-health track of CMS’s ACCESS model, so Sword is buying a live Medicare outcome-aligned payment position. (Hans co-founded Entrelacs, which sells conversational mental-health screening, care matching and monitoring to employers, clinicians and health systems; the merged Sword/Headspace sells into the same buyers and is a near-direct competitor.)
A transaction rather than a product launch, so no evidence line is owed, but the numbers should be handled accordingly: the price is company-redacted and press-sourced, the completion is a filing’s proposed date and not a confirmed close, and the filing date itself is contested between the Commission’s list and the trade reporting.
(filing)

⏺️ Character.AI, which built the largest consumer companionship product on the market, has repositioned itself as an entertainment platform. On 2 September it launched Comics, which turns chat transcripts into multi-page comic books monetised through the platform’s in-app currency, describing itself as “a connected entertainment platform where fans move seamlessly across Microdramas, interactive chats, novels, and visual formats.” On 3 September it published a safety post reiterating that open-ended chat with Characters was removed for under-18 users last year and describing the under-18 offer as built around Series, Feed and Comics. Read as a pair, a company that grew on relational chat is routing minors towards consumption formats and recasting companionship as fandom, with the repositioning post landing first and the safety post the following day.
No published evaluation found: PubMed and Europe PMC searched by sponsor and affiliation for Character.AI and Character Technologies, and the company’s own site searched, with nothing returned. No registered trial found; ClinicalTrials.gov and ISRCTN searched by sponsor name. No regulatory clearance or marking, and none is implicated, because the product makes no clinical claim. Four engagement figures in the Comics post (97% of comics featuring a Character the creator already talks to, over 75% inserting themselves or a persona, over 30% finishing a comic, 36% making more than one page) are company telemetry from the first week with no denominator disclosed and are not independently verifiable. The safety post reports no figures at all: no prevalence data, no detection performance, no outcome for the self-harm classifier.
(link)

⏺️ Two companion robots launched at IFA Berlin with explicit loneliness claims, one of them citing studies it does not identify. Mind With Heart Robotics brought AnAn, a tactile panda robot, to Europe on 3 September at $1,000–1,500, with distributor agreements in Germany and Belgium and a completed production run of 100 units, which makes it a pilot more than a launch. Two days later Tuya Smart unveiled Doova, a mobile home robot for seniors with LiDAR navigation, skeletal-recognition fall detection and a 60-second fall threshold, which the company says will “ease feelings of loneliness”; price, markets, availability and business model are all undisclosed, and Tuya sells mainly to OEMs, so this may be a reference design. The September AnAn release is comparatively restrained. Its January CES release is not: it described AnAn as providing “24/7 stigma-free emotional support for loneliness, anxiety, and depression” and referred to “preliminary studies demonstrating measurable mood improvement”.
For AnAn: no published evaluation found — PubMed, Europe PMC and Google Scholar searched for the company, the product and the named researchers at its joint laboratory with Xi’an Jiaotong-Liverpool University. The “preliminary studies” claim names no institution, researcher, sample, instrument or outcome, so there is nothing to check it against; the same applies to a claim of applicability in “24 out of 74 objectives for Autism Spectrum Disorder”. No registered trial found (ClinicalTrials.gov and ISRCTN searched by sponsor); the company describes early clinical trials in dementia in Australian hospitals as planned. No clearance or marking; the company states it is “positioned for future regulatory pathways”. A CES 2026 Innovation Award is not evidence. For Doova: no published evaluation found on the same searches, no registered trial, no clearance or marking; the product avoids medical-device status by framing loneliness relief as wellness and fall detection as alerting family. The 60-second threshold is Tuya’s own specification and no sensitivity, specificity or false-alarm rate is published. The existing trials of other social robots cannot be transferred to either device.
(AnAn) (Doova)

⏺️ Catching up on August: the FDA is letting four generative-AI health products reach patients before it has evaluated whether they work. The TEMPO pilot, whose participants were first announced on 22 July with SonderMind and Limbic added around 17–19 August, has the FDA exercising enforcement discretion over premarket authorisation and investigational-device requirements for products offered to participants in CMS’s ACCESS model, which launched on 5 July and ties payment to measured outcomes. The FDA’s own page states that the effectiveness of the selected devices “have not yet been evaluated by the FDA”, so this is not a clearance and should not be reported as one. The four are SonderMind’s adjunctive care application and Limbic’s Unpacked, both for depression and anxiety, plus Cadence’s HypertensionOS and Dexcom’s glucose programme. The contraindications are narrow and rarely reported: SonderMind’s app is contraindicated in suicidality, mania, psychosis and PTSD, and Unpacked is limited to English speakers with telephone access. Neither mental-health product claims a social-connection mechanism, which is why it sits here as a regulatory event and not as a product in scope. (Hans co-founded Entrelacs; Limbic and SonderMind both sell AI mental-health assessment, triage or measurement into Entrelacs’s buyers and are direct competitors.)
Limbic: a registered trial exists (ClinicalTrials.gov NCT05495126) and peer-reviewed work is indexed on PubMed, including real-world observational studies of conversational AI in mental-health assessment; the company holds UK Class IIa medical-device certification. Its count of nine peer-reviewed studies submitted with the TEMPO application is company-sourced and was not enumerated; a further claim about therapeutic alliance by voice versus text is described as an internal study with no publication details. SonderMind: no independent evaluation found, having searched ClinicalTrials.gov and ISRCTN by sponsor name, PubMed and Europe PMC, and the company’s own site, which carries a self-published clinical-results page and no registered study. No FDA clearance, CE mark, UKCA, NICE early value assessment or NHS DTAC assessment found for either mental-health product. TEMPO admission itself is not evidence of effect and the FDA says so.
(link)

⏺️ Money

⏺️ France’s Fondation des Sciences Sociales has opened a call on “Solitude et isolement” and it runs to 7 November. Published on 1 September, this is the call for contributions to the fifteenth Journée pour les sciences sociales, to be held in March 2028. Selected contributors receive €2,500, submit their work to a peer-reviewed journal, present at the Journée, take part in three preparatory sessions, and appear in a French-language collective volume. It is open to researchers and post-doctoral researchers in the humanities and social sciences holding a doctorate, of any nationality, affiliated to a French or European institution. Small money, and a contribution call instead of a project grant, but the theme is exactly this field and the deadline is eight weeks out. The page gives the deadline as 7 November at midnight without stating a time zone; assume Paris time and do not cut it fine.
(link)

⏺️ Nine days left on the two Horizon Cluster 2 topics, and nothing new has opened in Horizon Europe at all. HORIZON-CL2-2026-01-TRANSFO-04, on digital tools outside school and their effects on educational outcomes and mental health, shows as open for submission on the Funding and Tenders portal with a deadline of 23 September 2026 at 17:00 Brussels time; TRANSFO-09, on long-term care policy and EU demographic change, shares that deadline, and the €60 million co-funded partnership under TRANSFO-01 closes on 13 October. This digest has run all three before and the only thing that has changed is the clock. Two absences belong in the record. No Horizon Europe call anywhere in the work programme has an opening date in September, and HaDEA’s open-calls list currently contains no health, mental-health, social or wellbeing topic at all. The NIHR Public Health Research call on loneliness in the community, which this digest carried in its first issue, closed at 1pm on 18 August.
(topic page)

⏺️ Sweden’s annual state grant for organisations working against involuntary loneliness opened on 1 September and closes on 15 October. This is service money, not research money, and it is listed here because it is the only thing on the exact topic with a live deadline besides the French call. Socialstyrelsen administers it under förordning 2019:474 for non-profit associations, non-municipal non-profit foundations, and faith communities and congregations, funding meeting places, contact points, and work that brings more people into volunteering. The page reports 801 applications last year and states that decisions are planned for the first quarter of 2027, contingent on the Riksdag voting the funds. It gives no total envelope, so the size of the round is unverified from the call itself.
(link)

Corrections, additions and arguments in the thread, please, especially from those of you closer to the Irish, Spanish and Swedish material than I am. It feeds the next issue and the brief.

Hans

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