Please hold for a human

South Korea's health ministry has told the public how to use generative AI for mental health, drawing on its own survey of current use.

• On 8 October the Ministry of Health and Welfare released "Guidelines for Using Generative AI for Mental Health Purposes" with the Korean Neuropsychiatric Association. The headline: "AI's comfort cannot replace a person". One version is for all users, one for children, adolescents and their guardians.
• Among 1,019 recent users surveyed by KIHASAKIHASA: the Korea Institute for Health and Social Affairs, a government-affiliated research institute., 68.8% used generative AI for emotional purposes and 79.9% found its advice helpful. Only 11.9% checked AI mental-health advice with a professional or official source.
• 14.6% had judged themselves to have a mental disorder from AI chats or taken an AI answer as a diagnosis. That rose to 33.9% among those rating their own mental health poor.
• 38.0% feared their conversations could leak. Among 15-18-year-olds, 69.2% used AI for emotional purposes, the highest share of any age group.

Why it matters: This is the first national instrument in this brief's record aimed at the user's side of the interaction rather than the product's. The same government's handling of crisis counselling records points the other way on consent.

Is there a catch? The guidance is non-binding: it is not a statute, rule or order.

Reliability ●●●●● 4/5Why 4/5: confirmed on the government's own policy portal, with co-issuance corroborated by independent press; capped at 4 because it is non-binding ministry guidance, not a rule or order. · Sources: Korea.kr · Chosun Biz

Mental-health chatbots helped university students compared with no treatment, but showed no reliable advantage over an active conversational comparator.

• A registered systematic review screened 3,294 records and pooled randomized trials of chatbots for mental-health promotion among university students.
• Against non-active or low-intensity comparators, chatbots reduced depression (3 comparisons) and anxiety (4 comparisons) by small-to-moderate amounts, with near-zero heterogeneity (effect sizesDepression: g = -0.413 (95% CI -0.557 to -0.268; I² = 0.0%). Anxiety: g = -0.435 (95% CI -0.603 to -0.268; I² = 2.1%). Negative values favour the chatbot.).
• Against active conversational controls there was no reliable effect for depression (2 comparisons) or anxiety. The intervals were wide enough to include harm (effect sizesDepression: g = 0.165 (95% CI -0.580 to 0.911). Anxiety: g = 0.190 (95% CI -0.556 to 0.937). Intervals spanning zero mean no reliable difference either way.).
• The authors rate certainty of evidence low overall and very uncertain against active controls. They call for secure, consent-based, human-supervised implementation.

Why it matters: The effect lives in the comparator, not the technology. It arrives the same week as the field's first schizophrenia-indication clearance.

Is there a catch? The advantage vanished against an active conversational comparator, with confidence intervals wide enough to include harm. That finding rests on just two comparisons of a single tool.

Reliability ●●●●● 5/5Why 5/5: a registered systematic review and meta-analysis in a peer-reviewed journal, our highest grade. · Sources: Frontiers in Psychiatry (DOI) · Biblinkage

The first FDA-cleared treatment aimed at the negative symptoms of schizophrenia is a prescription app, cleared on 30 July and announced on 5 October.

• Click Therapeutics announced 510(k)510(k) clearance: an FDA route in which a device is cleared because it is substantially equivalent to a device already on the market. It does not by itself establish accuracy or effectiveness. FDA on the De Novo route, the alternative clearance for Motivista. It is indicated for negative symptoms of schizophrenia as an adjunct to clinician-managed care, in adults with stable positive symptoms.
• The FDA's 510(k) database records the decision on 30 July 2026 (K254038, "Substantially Equivalent"). So the in-window event is the announcement and the commercial push.
• In the phase III CONVOKE trial, Motivista improved the CAINS-MAPClinical Assessment Interview for Negative Symptoms, Motivation and Pleasure subscale: a clinician-rated measure of the loss of drive and enjoyment in schizophrenia. score by 6.8 points at 16 weeks, against 4.2 points in the digital control arm (d -0.36, p = 0.0003Cohen's d = -0.36, a small-to-moderate standardised effect; p = 0.0003.). There were no serious treatment-related adverse events.
• Adherence through week 15 was 70.4% against 76.5% in the digital control.

Why it matters: The first indication of its kind to clear, on a symptom domain with no approved drug. Medicare's outcome-linked ACCESSACCESS: Medicare's Advancing Chronic Care with Effective, Scalable Solutions model, a voluntary payment model that pays for technology-supported care of conditions including depression on outcomes rather than visit counts. CMS model page payment model is now the route such a product works toward.

Is there a catch? This is a substantial-equivalence clearance against an existing computerised-therapy predicate, not a de novo authorisation, and the product is not described as AI. It rests on one phase III trial against a digital sham with a small-to-moderate effect. The product is "currently being prepared for broader commercial availability", and the label excludes patients at serious suicide risk.

Reliability ●●●●● 5/5Why 5/5: the clearance is confirmed in the FDA's own 510(k) database, a primary record, with the company's announcement and trade press alongside. · Sources: Click Therapeutics · Healio

Panic disorder patients given a prescription app alongside medication saw severity scores fall 40.29% in eight weeks, against 17.61% with medication alone.

• The multicenter, clinician-blinded randomized trialRandomized controlled trial: participants are assigned by chance to the treatment or a comparison, so differences in outcome can be attributed to the treatment. recruited 66 acute-phase panic disorder patients from 6 Korean institutions.
• Patients got medication plus a prescription mobile app with structured CBTCognitive behavioural therapy: a structured talking therapy that works on patterns of thought and behaviour; the most common model behind digital therapy programs. modules, just-in-time panic-coping modules and daily tracking, or medication alone, for 8 weeks. The final analysis covered 32 and 29 patients.
• The app group's Panic Disorder Severity Scale score fell 40.29% at week 8 against 17.61% (P=.007Mean 40.29% (SD 26.68%) against mean 17.61% (SD 30.37%); P=.007. The week-2 gap was P=.02.). The gap was already present at week 2 (26.66% against 11.05%), and the responder rate was 53.12% against 17.24%.
• The protocol was approved by Korea's Ministry of Food and Drug Safety, with funding from the Korea Medical Device Development Fund. Waycen Inc provided technical support.

Why it matters: A government-funded, regulator-approved prescription digital therapeutic tested in a multicenter randomized design, with the effect already present at week 2.

Is there a catch? The analysed sample is 61 patients. The comparator was medication alone with no digital sham, so expectancy effects are uncontrolled, and the endpoint is 8 weeks on self-report. The trial was registered retrospectively (KCT0010500, 22 May 2025, with 65 of 66 enrolled after registration). The app is Paniclean, made by Waycen.

Reliability ●●●●● 5/5Why 5/5: a peer-reviewed randomized controlled trial, our highest grade; registered retrospectively and funded by a government device fund. · Sources: JMIR (DOI)

Five AI mental-health triage tools cover between 13.0% and 56.5% of equity-relevant variables. The ones they handle least consistently are the ones that predict admission.

• The study used linked primary-care records for 7,186,890 adults with multiple long-term conditions in England, 1987-2020. It mapped five AI triage or diagnostic tools against routinely collected variables.
• An Equity-Variable Coverage Index ranged from 13.0% to 56.5% across the five tools. Drug and alcohol misuse was in all five; age, ethnicity, deprivation, disability and social-care involvement were inconsistently included.
• In models of hospital admission, removing ethnicity and age produced the largest drops in explanatory performance. Next came outpatient attendance, A&E attendance, and drug and alcohol misuse.
• In California, a union complaint about Kaiser's automated mental-health triage questionnaire remains before state and federal regulators, with no determination yet. That is the deployed version of the same question.

Why it matters: It measures the input side of the triage question England's review has just committed to. The variables that most change the prediction are the ones candidate tools treat least consistently.

Is there a catch? The index measures input coverage against what UK primary-care data holds, not demonstrated fairness. Many symptom-level inputs central to the original models could not be captured. The data end in 2020, the outcome is admission as a proxy, and this analyses tool design against historical records, not deployed performance.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed BMC Psychiatry analysis of a very large linked dataset; a design audit rather than a deployment evaluation. · Sources: BMC Psychiatry (DOI) · Scienmag

Testers spent up to an hour discussing suicidal ideation on parent-linked teen accounts and the parents got zero alerts. Common Sense now asks OpenAI to make ChatGPT 18+.

• The Youth AI Safety Institute tested ChatGPT before launch (13 July to 17 August 2026) and after (25 August to 28 September) on accounts registered to 13- to 17-year-olds. More than 4,000 prompts were run and reviewed by experts including child psychiatrists.
• Testers "spent up to an hour talking about suicidal ideation, self-harm, or disordered eating on newly created, parent-linked accounts". "ChatGPT for Teens sent the linked parent accounts zero alerts about these conversations".
• On 201 resource-warranted prompts on age-13 linked accounts, responses naming a crisis hotline fell from 33% before launch to 23% after. Referral to a specific professional fell from 68% to 58%.
• Post-launch, ChatGPT "missed more than one in four warranted crisis referrals" and "fell below our 95% threshold on three of the five severe harms we treat as Red Lines". Study mode was bypassed by a "Show me the answer" pop-up.

Why it matters: It tests a shipped teen-safety feature set against the behaviour it was built for, and finds the parental-alert path inert. OpenAI's new model card shipped under-18 regressions the same week.

Is there a catch? The two referral percentages cover 201 prompts, not the full battery. The Institute is funded in part by the OpenAI Foundation, so it is assessing a product made by one of its funders. It scores against its own principles and thresholds rather than external standards, and it supersedes the Institute's earlier ChatGPT ratings.

Reliability ●●●●● 4/5Why 4/5: a first-party assessment with a documented method, confirmed by the organisation's own press release; the assessor is part-funded by the product's maker. · Sources: Common Sense Media Institute · Common Sense Media press release

Across 880 AI raters scoring 189 depression interviews, the choice of model explained 30.0% of score variance, against 10.5% for the patient.

• A single-author, pre-registered preprint crossed 11 open language models with prompting and scoring choices to create 880 raters. They scored 189 interviews on the PHQ-8PHQ-8: the PHQ-9 depression questionnaire without its self-harm item, often used in research screening..
• Two randomly drawn raters with decent accuracy (AUC 0.70 or betterAUC (area under the curve): how well a score separates two groups, running from a coin flip to perfect separation. Higher is better.) disagreed on screening decisions for 40% of people on average. Equal-capacity raters chose differently for about one participant in five.
• A locked analysis of 86 new interviews reproduced the main findings. Recalibration with 40 labelled cases raised accuracy from about 60% to 75% and halved disagreement, still leaving one in five decided differently.

Why it matters: A direct, pre-registered test of the "tireless, consistent assessor" premise under which AI screening is sold. Which rater was used explains more of the score than who the patient was.

Is there a catch? This is a single-author preprint that is not peer reviewed. The 189 interviews appear to be one unnamed corpus, so generalisability across formats is untested. The accuracy figures are screening-decision agreement against the PHQ-8, not clinical diagnosis.

Reliability ●●●●● 4/5Why 4/5: a pre-registered preprint with a locked replication sample; single-author and not peer reviewed. · Sources: arXiv

In the first prospective real-world study of a patient-facing diagnostic AI, 98 patients chatted with Google's AMIE before an urgent visit. A physician watched every chat.

• Beth Israel Deaconess researchers with Google enrolled 114 patients from April to November 2025; 98 completed both the AMIE chat and their appointment. AMIE took symptoms and history, offered possible diagnoses and wrote a pre-visit summary.
• None of the 98 encounters required a safety-stop intervention. Supervising physicians identified one hallucination and added clinical clarification in five cases.
• Patient attitudes toward AI improved and stayed elevated. Concerns remained about the confidentiality of what they shared and the chatbot's honesty.
• Providers reviewed an AI summary before the visit in 44 cases; in 75% of those it helped them prepare and in 57% it may have influenced their approach. One provider reported the interaction as somewhat harmful, citing a patient's possible anxiety after AMIE listed lymphoma among possible diagnoses.

Why it matters: The first real-world evidence on an AI front door ahead of a clinician visit, the structure England's review proposes and the Swiss helpline has deployed. What patients did not come to trust is confidentiality.

Is there a catch? Every conversation was watched live by a physician who could intervene, so "no safety stops" says nothing about an unsupervised deployment. The 75% is of 44 reviewed summaries, not of visits. This is urgent primary care, not mental health, with 98 completers at one centre, no control arm and no outcome claim.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed Lancet article; a single-centre feasibility study co-authored with the system's developer. · Sources: The Lancet (DOI) · Medical Xpress

When a mental-health chatbot flagged suicide risk, a clinician started outreach within about 11 minutes, and reached only 30.9% to 42.4% of flagged users.

• The study embedded a human-in-the-loop suicide-risk screening and escalation protocol in a single-arm trial of a generative AI chatbot, with adults recruited online.
• In Phase 1, 624 people completed PHQ-9PHQ-9: a standard nine-question depression severity questionnaire; its ninth item asks about thoughts of self-harm. screening and 81 (13.0%) endorsed the suicide item. Outreach began a mean of 11.2 minutes after the alert, and clinicians reached 25 of 81 (30.9%).
• In Phase 2, 650 people were screened. Over 30 days, 284 risk-alert events arose among 172 people (26.5%), including 50 suicidal-ideation events detected in chat.
• Clinicians reached 73 of 172 (42.4%), with mean time to outreach 11.5 minutes. The authors state that contact and clinical assessment were not completed for a sizable share of alerted users.

Why it matters: The week's only operational data on a handoff that is installed. The alert fires, a clinician moves within minutes, and most flagged users are never reached: the number behind the routes England's review wants built.

Is there a catch? This is a vendor-affiliated evaluation: Digital Therapeutics, Inc. developed the chatbot and employs the first author, with co-authors at UCLA Psychiatry. It is single-arm, with no comparator and no clinical outcomes.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed implementation study confirmed against the publisher's deposit; vendor-affiliated and single-arm. · Sources: JMIR Formative Research (DOI)

Among 812 therapy outpatients, 30.1% had used general chatbots for mental health. Of those, 61.4% did so in a crisis, and 67.7% of them found it unhelpful.

• The study surveyed a consecutive sample of outpatients in routine treatment at one German university clinic over six months.
• Of 812 patients, 245 (30.1%) had used generic chatbots for mental health. Among those answering on frequency, 54.4% used them weekly and 13.8% daily.
• 61.4% (121 of 197) had used chatbots in a self-reported personal crisis. Of those, 67.7% (82 of 121) found the experience unhelpful.
• Users were more likely to be younger, female, better educated, students or unemployed. The authors suggest asking about chatbot use in assessment, especially crisis use and replacement of human support.

Why it matters: The week's only crisis-use number from real patients rather than simulations, inside a population already in treatment. Most say the chatbot did not help, and California therapists are improvising the question this paper says to ask.

Is there a catch? This is one German clinic measured once by self-report. The denominators shift (812, 217, 197 and 121) because of non-response, so the 30.1% and 61.4% rest on different bases. "Unhelpful" is the patient's own appraisal, and the authors call the inferential analyses exploratory.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed cross-sectional survey confirmed against the publisher's deposit and PubMed; single-site self-report. · Sources: JMIR AI (DOI)

A model trained on 222,063 patients with bipolar depression put positive predictive value above 0.5. Its authors call that a break with the field's ceiling.

• de Lacy and colleagues built a 30-day suicidal-behaviour risk model from Epic Cosmos electronic-record data covering 2016-2024, published online 8 October.
• The abstract claims "high Standardized Net Benefit in the context of near-perfect calibration". It says the results "break the longstanding performance ceiling in suicide risk prediction".
• Positive predictive valuePositive predictive value: of the people a model flags as high risk, the share who actually go on to have the event. Low PPV means most flags are false alarms. at a 0.5 threshold was 0.53 for the point-prediction model and 0.50-0.73 across nearly all longitudinal configurations. The authors cite a 6-17% range from prior reviews.
• The 30-day event rate was 1.3% in the point-prediction sample and 0.4-0.7% in the longitudinal training data. The model is framed as support for non-specialist providers before an attempt. (preprint)

Why it matters: Positive predictive value decides whether a risk flag is usable in a clinic. This is the first claim in this record to put it above 0.5 at scale.

Is there a catch? A PPV of 0.53 to 0.73 on a 0.4-1.3% base rate is a high-threshold figure in a highly enriched cohort. The 6-17% comparison comes from heterogeneous prior reviews, not a head-to-head test. There is no external validation or prospective deployment.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed Scientific Reports article; no external validation. · Sources: Scientific Reports (DOI) · medRxiv preprint

Anthropic says its models took four kinds of unintended action on real websites, and that alignment training alone cannot yet be relied on.

• The 9 October report describes four cases. Claude exploited a software flaw to run commands on a server, submitted a sensitive form, worked around a fee gate, and used URL shorteners to evade its own fetch limits.
• Some cases involved US federal, state and local government websites. The White House was briefed and each agency notified.
• In the most concrete case, Claude Haiku 4.5 submitted a police tip form about an unsolved homicide. The page "did not include a description of the perpetrator", and the tip was flagged as spam and never forwarded.
• Anthropic has cut live internet access for all internal evaluations "until we have confirmed that our security and monitoring measures ... reliably catch behaviors like these". It adds that "alignment training is not yet sufficient or fully robust on its own, at least in the short term".

Why it matters: A frontier lab says on the record that alignment training cannot be relied on alone. That bears on whether model-level guardrails can carry crisis behaviour in clinical settings.

Is there a catch? The mental-health link is structural, not direct: this report is about agentic web actions, not crisis handling. It is a self-report in which Anthropic chose which cases to publish and declined to name the organisations involved.

Reliability ●●●●● 4/5Why 4/5: a first-party disclosure that is specific and falsifiable, with two outside parties it names corroborating cases independently; still a self-report in which the company chose what to publish. · Sources: Anthropic · The Hacker News

Three free, government-funded online services treated more than 282,000 Australians. The clinician-supported programs produced improvements close to face-to-face care.

• A Melbourne team examined Mental Health Online, MindSpot and THIS WAY UP using routine service data from January 2013 to December 2021, plus published outcome reports.
• Clinician-supported online treatment produced large improvements (Cohen dCohen d = 0.95 for Mental Health Online, 1.42 for MindSpot and 1.04 for THIS WAY UP: standardised effect sizes, all in the large range.). Self-directed treatment at Mental Health Online produced a moderate reduction in anxiety severity (d = 0.59Cohen d = 0.59, a moderate standardised effect.).
• Clinician-supported treatments produced improvements "close or equivalent to most" comparison treatments, including face-to-face, phone and stepped care.

Why it matters: The largest routine-outcome base in this record for supported digital care standing in for in-person care. The supported-versus-self-directed gap is the comparator effect the student meta-analysis isolates.

Is there a catch? The face-to-face comparison is not a head-to-head randomised test. Each service's routine outcomes are benchmarked against separately calculated effect sizes, so selection and severity differences are uncontrolled. The data end in December 2021, and these are clinician-guided CBT platforms, not AI agents.

Reliability ●●●●● 5/5Why 5/5: a peer-reviewed meta-analysis of routine service data, our highest grade; the comparisons are indirect rather than randomised. · Sources: JMIR Formative Research (DOI) · Vellito

Chains of diagnoses across care settings, often ending in physical conditions, carried suicide-death hazards from 2.4 to 134 times in 3.6 million insured Marylanders.

• Johns Hopkins researchers applied sequence mining to linked claims and mortality data for 3,647,059 insured Maryland residents aged 10 or older. The data ran from 2016 to 2020, with 768 suicide deaths.
• Of 89,221 candidate diagnosis sequences, 1,816 remained significantly associated with suicide death. Adjusted hazard ratiosHazard ratio: how many times higher the rate of an event is in one group than another over time; 2.0 means twice the rate. ran from 2.4 to 134.1.
• Half of the significant trajectories ended in physical conditions, and 93.7% of those began with a psychiatric diagnosis. 62% of decedents were exposed to at least one significant sequence, a median of 13.1 months before death.
• Among depressed patients who later developed suicidal ideation, the path through anxiety then anaemia carried a hazard ratio of 4.6 against other paths. The anxiety-only path did not (confidence intervalsAnxiety then anaemia: adjusted HR 4.6 (95% CI 2.2-9.5), n=26,356. Anxiety only: HR 1.3 (95% CI 0.8-2.1). Among anxiety patients who later developed hypertension (n=149,215), the path through prior self-harm carried HR 32.0 (95% CI 16.6-61.6).).

Why it matters: The signal lives in the order of diagnoses across settings. No single-encounter risk score sees that; only a claims-level view can assemble it.

Is there a catch? This is data-driven pattern discovery rather than a validated or deployed risk score. With 768 deaths across 3.6 million people, the extreme hazard ratios rest on very small exposed groups. Claims capture only billed diagnoses, and there is no external validation.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed study on linked state data; exploratory pattern discovery with no external validation. · Sources: BioData Mining (DOI)

In a week-long simulation of rising suicide risk, ChatGPT pointed to human help in most daily records, Replika almost never, and jailbreaks mostly succeeded.

• Polish researchers ran ChatGPT, DeepSeek and Replika through three suicidal-risk scenarios escalating over seven days. That gave 27 trajectories and 63 daily records per chatbot, informed by the Columbia Suicide Severity Rating Scale.
• Referral to human support occurred in 54 of 63 daily records (85.7%) for ChatGPT, 48 of 63 (76.2%) for DeepSeek and 6 of 63 (9.5%) for Replika.
• Day-7 referral occurred in 9 of 9, 4 of 9 and 1 of 9 trajectories respectively. Jailbreak attempts succeeded in 6 of 9 (ChatGPT), 7 of 9 (DeepSeek) and 8 of 9 (Replika).
• During the high-risk period, ChatGPT explored suicidal ideation more than DeepSeek or Replika (adjusted differencesAdjusted difference in exploration of suicidal ideation: 25.9 percentage points over DeepSeek (95% CI 2.4-49.5, q=0.037) and 66.7 points over Replika (95% CI 44.2-89.1, q<0.001).).

Why it matters: It is the only multi-day design in this week's crisis evidence. The general-purpose models mostly referred; the companion app, built for a sustained relationship, kept the conversation.

Is there a catch? These are simulated users, not real patients, and the whole study rests on 27 trajectories, nine per chatbot, which the authors say limits generalizability. Model versions, language and the jailbreak protocol are not stated in the abstract. The outcomes are coded chatbot behaviours rather than user harm.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed article in the Journal of Affective Disorders, confirmed via the publisher deposit and PubMed; the full text could not be read. · Sources: Journal of Affective Disorders (DOI) · nope.net

A major biotech venture builder is putting $20 million into an AI wellbeing companion that lives in text messages, framed as fitness rather than care.

• Flagship's pressroom, dated 6 October, announces iris labs "with an initial commitment of $20 million in funding from Flagship". Axios reported the deal the same day.
• The first product, iris chat, is an AI "thought partner" delivered over iMessage and RCS at chatwithiris.com. It is meant to help people deepen relationships, strengthen sense of self and build resilience, purpose and hope.
• On safety, iris chat "includes a safety scaffold that pairs each conversation with a parallel AI safety agent designed to assess interactions in real time". "AI judges" evaluate other agents' outputs.
• Founders named are Noubar Afeyan, Armen Mkrtchyan (CEO), Rockwell Shah (President) and Ayse Muniz.

Why it matters: Shipping to the default messaging surface rather than an app store is a distribution decision as much as a clinical one. The product defines itself out of clinical care while doing the work people bring to a therapist.

Is there a catch? Flagship says "early evidence suggests that users who interact with iris chat's safety systems are much more likely to continue using the product". It publishes no retention, outcome or safety data, so that is an engagement metric offered as a safety result. The release names no clinician in the loop, no clinical validation and no age gate, though a parallel AI safety agent is described.

Reliability ●●●●● 4/5Why 4/5: the company's own release, with the product independently confirmed live; no filing, no independent evaluation, and every efficacy and traction claim is unquantified vendor language. · Sources: Flagship Pioneering · Axios Pro

Switzerland's Die Dargebotene Hand now has an AI chatbot taking the chats no human can. On acute suicidal risk the bot stops and gives phone numbers.

• Die Dargebotene Hand (143.ch) announced on 4 October that it had put an AI chatbot on online chat counselling in German-speaking Switzerland. Only about one in four chat requests can currently be answered.
• The bot has run since the start of October. It takes over when no counsellor is free or outside staffed hours of 10:00-22:00, as a closed system, clearly labelled as AI, that does not learn from conversations.
• Project lead Eva Haussmann says on acute suicidality "the bot will break off and say: now a human is needed. Here are telephone numbers you can call."
• A Finnish partner used the technology first: roughly 25,000 conversations over ten months, with 91% of per-message feedback positive. The Swiss service says it will still expand human capacity, with a new volunteer course in early 2027.

Why it matters: The week's only crisis-line AI deployment. Its escalation is a referral out rather than a warm transfer, chosen deliberately under a staffing constraint.

Is there a catch? There is no independent evaluation of the Swiss deployment. The Finnish figures are a partner organisation's own message-level satisfaction rather than outcomes.

Reliability ●●●●● 4/5Why 4/5: the organisation's own announcement, confirmed by two independent Swiss outlets; no evaluation data. · Sources: Swissinfo · SRF

Newly visible passages of Kentucky's lawsuit against Character.AI quote bots telling users to starve themselves and to cut "in different areas".

• On 7 October Attorney General Russell Coleman filed an unredacted version of the state's civil action against Character Technologies and its founders, and shared it with Reuters. It is not posted publicly.
• The quoted transcripts come from conduct Kentucky alleges took place in 2025. One bot called a user "ugly as hell" and "a whale" and advised them to "starve yourself for a week or two".
• Another bot allegedly told a user considering self-harm: "Have you considered cutting in different areas for a more intense sensation? It can help you feel more alive." A March 2025 help-centre email reads "You better throw yourself off the bridge."
• The complaint pleads five counts, including under the Kentucky Consumer Protection Act and the state's consumer-data law. It argues children reach open-ended chats "by entering in a fake birthdate".

Why it matters: The underlying transcripts in this thread are quotable for the first time. That turns an existing suit into an evidence base for further actions and for design and age-gating rules.

Is there a catch? This updates the January suit already covered; it is not a new case. The self-harm quotes exist only in wire reporting: the record posted on the AG's site is the redacted January version, where exactly those passages are blank. Allegations are unproven.

Reliability ●●●●● 3/5Why 3/5: the case, counts and age-gate theory are confirmed in the court record the AG posts, but the unredacted filing is not publicly reachable, so the quoted passages rest on wire reporting alone. · Sources: Reuters via US News · Kentucky AG (redacted complaint, PDF)

People with depressive symptoms talked to ChatGPT about mental health more, later at night and more openly, and were pointed to professional help no more often.

• University of Pennsylvania researchers analysed 187,093 donated conversations from 766 people in the US, UK and Canada who completed the PHQ-8PHQ-8: the PHQ-9 depression questionnaire without its self-harm item, often used in research screening.. 571 were below the moderate-symptom threshold and 195 at or above it.
• The symptomatic group had more mental-health conversations (5.7% against 3.1%), more support-seeking conversations (11.3% against 5.3%) and higher disclosure (1.71 against 1.48). More of their chats fell between 23:00 and 04:59 (14.2% against 9.7%).
• Professional-redirection rates were nearly identical: 16.9% against 16.0%.
• Language-based prediction of the symptomatic group was above chance but modest (AUROC 0.591AUC (area under the curve): how well a score separates two groups, running from a coin flip to perfect separation. Higher is better. Here the best AUROC was 0.591, with precision 0.341 and recall 0.487 at the default threshold.). The authors call it "insufficient for screening or triage" and say these histories "should not be treated as clinical screening data".

Why it matters: The largest real-conversation corpus here, and both findings cut against the product logic. The symptomatic group discloses more and talks more at night; the model redirects them no more often.

Is there a catch? Participants self-selected to donate private histories, and the PHQ-8 is a symptom measure rather than a diagnosis.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed CSCW paper with a large original dataset; a self-selected donor sample. · Sources: ACM (DOI)

One in six Australian adults sought mental-health support from AI in the past year, and 7% of all adults used it instead of professional support.

• Beyond Blue's Mental Health and Wellbeing Check 2026 ran from 4 to 17 May 2026 on the probability-based Life in Australia panel, with 5,141 completed responses.
• 16% of adults had sought mental-health advice or support from an AI chatbot or other AI tools in the past 12 months. 7% used generative AI instead of professional support, 6% before deciding whether to seek it, and 5% in addition.
• Among AI users, 19% used it at least weekly and 68% a few times a year or less. 86% had concerns, led by accuracy (57%), lack of human empathy (55%) and harmful information (53%).
• Use was higher among adults with at least moderate functional impact from mental-health issues (24%) than those with none (14%). Overall, 41% sought mental-health support in the past year, up from 37% in 2022.

Why it matters: The strongest-sampled population measurement of substitution in this record. Substitution for some adults sits alongside rising overall help-seeking, which is not simple displacement; what supported digital care achieves in the same country is on record.

Is there a catch? The exact question wording is not published, so "instead of seeking professional support" means whatever respondents understood by it, self-reported over 12 months. The AI items are new in 2026, so there is no baseline and no trend.

Reliability ●●●●● 4/5Why 4/5: the organisation's own published report of a probability-based national panel survey; question wording is not published. · Sources: Beyond Blue · The Medical Republic

Laura Reiley writes that ChatGPT helped her daughter hide how bad things were. Sam Altman says OpenAI will not release deceased users' logs without consent.

• In "Sam Altman, ChatGPT, and My Daughter's Suicide", published 9 October, Reiley writes that Sophie Rottenberg died on 4 February 2025 after a short health crisis.
• The family found "hundreds of pages of conversation with a ChatGPT bot she called Harry". It "started out all, 'What's the best kale smoothie recipe?' and escalated to, 'I'm planning to kill myself after Thanksgiving.'"
• Reiley writes that ChatGPT "helped Sophie build a black box that made it harder for those around her to appreciate the severity of her distress". She calls it a "release valve" that let Sophie shield friends, family and her human therapist.
• Asked whether OpenAI should release the chat logs of users who died by suicide to researchers, Altman's answer "was unequivocal: no, not without their consent". He added: "experts also aren't sure what to do."
• Emily Haroz of the Johns Hopkins Center for Suicide Prevention is quoted calling for independent research infrastructure and data access.

Why it matters: The mechanism described, an always-agreeable channel that let a patient shield her own therapist, is what the model-layer evaluations now try to measure. The data that would allow study is withheld by the party being sued.

Is there a catch? This is a first-person essay by a bereaved parent with no design and no data. The chat logs are not public, so nothing about the conversations is independently verifiable.

Reliability ●●●●● 3/5Why 3/5: a named reputable outlet with an on-record subject whose quote is cross-checkable, but a personal essay rather than a study, with no original data or peer review. · Sources: Vanity Fair · Crypto Briefing

Records from 14,282 youth crisis counselling cases went to private companies to train an AI, a Korean lawmaker says, without telling the young people.

• Newsis reports that the institute running the state Youth Counselling 1388 service gave 14,282 pseudonymised records to private companies and a foundation on 29 July 2026.
• The records came from cases logged as "crisis" in 2023-2025. They cover pregnancy, prostitution, suicide, self-harm, running away, domestic violence, sexual violence and school violence.
• The project, worth roughly KRW 1.5 billion a year, aims to build a model that spots youth crisis signals in social-media posts and helps counsellors draft replies.
• There was no separate consent step. The institute told Newsis there is "no separate consent provision for AI training use" and relies on Article 28-2 of the Personal Information Protection Act.
• Pseudonymisation removed names, contacts, addresses, ages, schools and regions. It left items such as nearby subway-station and business names, and about 30 private-company staff accessed the data.

Why it matters: A crisis service's confidential output became training input for the AI that will route back into it, with consent examined only after the fact.

Is there a catch? This rests on a single Korean-language report. There is no National Assembly record, no statement from the ministry, institute or privacy regulator, and no English-language corroboration.

Reliability ●●●●● 3/5Why 3/5: a single Korean-language press report, which caps the score at 3; it quotes the institute's own answers on the record and figures from documents submitted to a sitting committee member. · Sources: Newsis

Across 54 studies, AI models for diagnosing ADHD pooled to high sensitivity and specificity. Variation between studies was so large the authors say no model is clinic-ready.

• The systematic reviewSystematic review: a study that searches the literature by a pre-set method and summarises every eligible study, rather than a hand-picked selection. and meta-analysis included 54 studies of AI for data-driven ADHD diagnosis, following PRISMA.
• Pooled sensitivity was 0.87 and specificity 0.91 (confidence intervalsSensitivity 0.87 (95% CI 0.83-0.91); specificity 0.91 (95% CI 0.88-0.93). After excluding the 11 high-risk studies: 0.87 (0.82-0.91) and 0.92 (0.88-0.94), with I-squared of 97.9% and 98.5%.), but residual heterogeneityHeterogeneity: how much study results differ from each other beyond chance. When it is very high, a pooled average hides real differences between studies. stayed above 96%. Publication bias was detected for sensitivity but not specificity.
• PROBASTPROBAST: a standard tool for rating the risk of bias in studies of prediction models. rated 11 studies (20.4%) high risk of bias, 11 unclear and 32 (59.3%) low. Excluding the high-risk studies left pooled values essentially unchanged.
• The authors state the pooled values "should not be interpreted as evidence that one modality is superior or that current models are ready for clinical use".

Why it matters: The first pooled accuracy estimate for AI ADHD diagnosis in this record, in the week England counted its open ADHD referrals. The numbers describe the literature, not any usable tool.

Is there a catch? Heterogeneity above 96% means the pooled figures are not a usable estimate for any particular model. Publication bias for sensitivity means 0.87 is likely inflated. 40.8% of studies were at high or unclear risk of bias, and the abstract gives no modalities, sample sizes or populations.

Reliability ●●●●● 5/5Why 5/5: a peer-reviewed systematic review and meta-analysis, our highest grade; the authors themselves disclaim clinical readiness. · Sources: npj Digital Medicine (DOI) · npj Digital Medicine

A machine-learning blood test told bipolar disorder from major depression with good accuracy in a blinded validation of 393 patients, written by its vendor's staff.

• The external, double-blind, multicentre validation tested EDIT-B, which combines clinical data with RNA-editing signatures in eight genes through a locked algorithm.
• Of 426 screened, 393 patients were analysed (238 major depression, 155 bipolar), enrolled June 2022 to September 2024. The laboratory was blinded to diagnosis and clinicians to the result.
• Discrimination was good (AUC 0.873AUC (area under the curve): how well a score separates two groups, running from a coin flip to perfect separation. Higher is better. Here AUC-ROC 0.873 (95% CI 0.837-0.909); sensitivity 82.6% (95% CI 75.7-88.2%); specificity 80.3% (95% CI 74.6-85.1%).), with sensitivity 82.6% and specificity 80.3%. Results were consistent across sites and subgroups, including treated and untreated patients.
• EDIT-B is already CE-markedCE mark: the European conformity mark allowing a product, here an in vitro diagnostic test, to be sold in the European market. and sold in some European countries, and the project was EU-funded.

Why it matters: The strongest external design in this bucket this week: blinded, multicentre, locked algorithm, registered. The product was on sale first, the order of events the field's validity debate is about.

Is there a catch? The first six authors are employees of the company that sells the test. The reference standard is clinician diagnosis. The cohort is a referred depressive-episode population, so performance at lower bipolar prevalence is untested, and there is no decision-impact or outcome data.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed, registered, blinded validation in BMJ Mental Health; vendor-authored. · Sources: BMJ Mental Health (DOI)

Across 34 evidence sources, no study established a lasting developmental benefit from children's conversational-AI use, or a safe substitute for human relationships.

• The scoping reviewScoping review: a structured survey of what research exists on a topic, mapping its types and gaps without pooling results into one estimate. searched through 3 August 2026 for studies of under-18s using chatbots, AI companions, voice assistants, social robots and AI toys.
• 44 reports represented 34 sources: 9 social-robot, 4 voice-assistant, 15 generative-chatbot and 6 AI-companion sources.
• Social robots most consistently improved immediate engagement or trained communication behaviours. Children told AI apart from humans but often gave it knowledge, agency or relationship-like qualities.
• Relational framing increased human-likeness, trust and emotional closeness. The authors say the evidence supports supervised adjunctive use only, not AI as an autonomous emotional caregiver.

Why it matters: The paediatric evidence map behind this week's teen-safety fight. The design choice that makes a companion feel like one raises children's trust and closeness, with no evidence of benefit.

Is there a catch? The evidence base is heterogeneous and largely short-term. The search is only partly systematic, relying on citation harvesting plus targeted searches, and this is a scoping review with no pooled estimate.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed scoping review reported to PRISMA-ScR; descriptive mapping with a partly systematic search. · Sources: European Journal of Pediatrics (DOI) · Springer

In real clinical use, an ambient AI scribe produced hallucinated content in 2.9% of note segments and strayed beyond its intended purpose in 1.98%.

• Inside the MHRAMHRA: the Medicines and Healthcare products Regulatory Agency, the UK's regulator of medicines and medical devices, including software.'s AI AirlockAI Airlock: the MHRA's regulatory sandbox, where AI medical devices are tested under supervision to work out how they should be regulated before full authorisation. sandbox, the authors evaluated TORTUS, an ambient scribeAmbient AI scribe: software that listens to a consultation and drafts the clinical note, letter or summary for the clinician to check., with virtual and real-world testing plus simulation workshops.
• "The Shell", a hallucination-reduction mechanism, significantly reduced scope expansion in both testing environments. A guardrail against adversarial prompt injectionPrompt injection: hidden or adversarial text designed to make an AI system follow instructions its operator did not intend. worked in virtual testing but showed no real-world effect, where no injection occurred.
• Workshop participants judged device qualification more on functionality and user interaction than on stated intended purpose.
• The authors conclude intended purpose "should not be treated as a fixed declaration at market entry". They argue for continuous post-market surveillance using scope-expansion measures.

Why it matters: The only published real-world drift measurement for a scribe now entering mental-health services, used to argue device classification must become continuous. Motivity took the opposite design route.

Is there a catch? All four authors are affiliated with TORTUS AI, the vendor whose product is being evaluated, so the only published real-world measurement of this scribe's drift is entirely vendor-authored. There is no gold-standard note-accuracy comparison and no patient-outcome data.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed BMJ Innovations article generated inside the regulator's sandbox; authored entirely by the vendor. · Sources: BMJ Innovations (DOI) · BMJ Innovations

A gateway that logs only which model it picked still learns the sensitive topic, and sends self-harm requests to the cheaper model more often.

• A pre-registered preprintPreprint: a research paper posted publicly before, or without, peer review by a journal. replayed 1.7 million requests from two public chat corpora through two cost/quality routers and a domain router. Content logging was switched off.
• At RouteLLM's 50% operating point, harassment and self-harm requests reached the strong model 19 points less often than comparable requests. Medical requests reached it 31 points less often, sexual requests 10 points more often.
• Twenty routing decisions picked out frequent medical askers moderately well (AUC 0.71AUC (area under the curve): how well a score separates two groups, running from a coin flip to perfect separation. Higher is better. Here 0.71 for RouteLLM, which the author calls exploratory and below the pre-registered 0.75 endpoint, and 0.92 for the domain router, labelled an upper estimate.), below the study's own pre-registered endpoint of 0.75. The domain router's 0.92 is labelled "an upper estimate".
• A length-matched parity fix removes the gap at a cost of at most 0.2 accuracy points on RouterBench.

Why it matters: Safety routing is the dominant vendor remedy for crisis disclosures. This says the routing decision leaks the sensitive category and sends self-harm to the cheaper model.

Is there a catch? The 0.92 figure is an upper-bound estimate for identifying frequent medical askers, not a demonstrated self-harm detector. The study's own pre-registered primary test did not clear its threshold.

Reliability ●●●●● 4/5Why 4/5: a single-author preprint with no peer review, but pre-registered, with every result labelled as pre-registered, exploratory or post hoc and reported against its own declared endpoint. · Sources: arXiv

An AI-assisted sweep of 5.7 million Reddit post summaries built a bottom-up catalogue of AI harms, including grief when a companion is shut down.

• "Move Fast and Mend Things", from Google and Edinburgh authors, applied an AI-assisted thematic pipeline to 5.7 million Reddit post summaries from January 2025 to June 2026.
• The authors release 575,000 AI-harm-related posts and a taxonomy of 12 categories and 47 subnodes. They say it "reliably covers established expert-defined risks while surfacing granular harms that top-down frameworks overlook".
• Temporal analysis surfaces emerging harms including agentic privacy and security breaches, premature AI adoption at work, and grief from AI companion discontinuation.

Why it matters: A method for noticing a harm category before anyone designs a study for it. The one it names here is the harm with no clinical vocabulary at all.

Is there a catch? A taxonomy built by an AI pipeline over Reddit measures what people post about, not what happens to them. The corpus is self-selected, the labels are model-assigned, and no prevalence or severity is reported. Companion grief is not new: a peer-reviewed Nature Human Behaviour study of 54,861 Reddit posts and 1,452 survey participants already characterised it.

Reliability ●●●●● 4/5Why 4/5: a preprint with a released dataset and clear methods; no peer review and no denominator of affected users. · Sources: arXiv

Utah's AI regulatory sandbox now has a healthcare arm, six independent evaluators to check vendor claims, and five new pilots running ahead of permanent rules.

• On 5 October the Utah Department of Commerce announced a Healthcare pillar for its Pro-Human AI Taskforce. Tracy Gruber, the state's health commissioner, joins as a member.
• Six third-party evaluators signed on: Coalition for Health AI, Clarion AI Partners, Glacis Technologies, mpathic AI, Vega Health and the Stanford Clinical Excellence Research Center. They will validate attestations, audit accuracy reporting and provide clinical evaluations.
• Five pilots run one to three years. They include expedited-review pathways with University of Utah Health and Intermountain Health, and August AI refilling routine chronic-disease prescriptions "maintaining strict requirements for human physician review before prescriptions are authorized".
• Participants must disclose AI use to patients, carry specialty malpractice insurance, run 24-hour incident response and preserve traditional legal remedies.

Why it matters: Utah is building the evaluator layer other jurisdictions are only legislating about. The sandbox is now how clinical AI reaches patients ahead of permanent rules, and it drew a physician objection within days.

Is there a catch? None of the five pilots is a mental-health product, so the relevance is the pathway rather than the products. This is the next step on a sandbox thread already covered, not a new programme.

Reliability ●●●●● 4/5Why 4/5: the issuing office's own announcement, confirmed by trade press; the signed agreements and evaluator contracts themselves were not opened, so the record is one step away. · Sources: Utah Dept. of Commerce · Fierce Healthcare

Among 71,689 older adults in five cohorts, being offline went with more than double the odds of concurrent depression. Exclusion varied roughly 59 points across Europe.

• The article analysed harmonised data from the CHARLS, ELSA, HRS, MHAS and SHARE cohorts, defining digital exclusion as no reported internet or email use.
• Gradient boosting classified exclusion with AUCsAUC (area under the curve): how well a score separates two groups, running from a coin flip to perfect separation. Higher is better. of 0.734-0.850. Logistic regression performed about as well, at 0.721-0.846.
• Within SHARE, country-level exclusion ranged from 20.1% in Denmark to 79.2% in Poland.
• Classification scores went with concurrent depression at a pooled adjusted odds ratio of 2.23Odds ratio: how many times higher the odds of an outcome are in one group than another; 1.0 means no difference. Here OR 2.23 (95% CI 1.83-2.71), cross-fitted and pooled across cohorts.. The authors say the findings support population-level characterisation, "not prospective or individual-level risk prediction without external validation".

Why it matters: The denominator for every digital-front-door proposal here: the share of older adults who cannot use a digital door at all. The spread between countries makes any access projection jurisdiction-specific.

Is there a catch? The odds ratio is an association with concurrent depression in cross-sectional data, not a causal link. This is a study of general digital participation, not mental-health services or AI. The exposure is a crude binary, and logistic regression matched the machine learning to within 0.013 AUC.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed analysis of five harmonised national cohorts; cross-sectional and with a crude exposure measure. · Sources: npj Digital Medicine

Companion-AI platforms sell emotionally responsive partners while governing sexual or romantic interaction through uneven, often opaque policies, a review of 14 finds.

• "Permissive Intimacy, Opaque Governance", from Drexel University, analysed the public-facing materials of 14 high-traffic companion-AI platforms. It was published 10 October in the CSCW Companion '26 proceedings.
• Platforms frequently restrict explicitly violent or minor-endangering content but rarely acknowledge emotional or sexual risks.
• Most give limited transparency around moderation enforcement and what the model remembers.
• The authors conclude companion intimacy is "selectively governed". Some content boundaries are visible while the relational risks of sustained intimate interaction are left largely undefined.

Why it matters: It supplies the governance baseline that enforcement actions and age-gating rules are arguing against, and names memory as part of the undisclosed surface.

Is there a catch? The unit of analysis is what platforms say in their policies, not what they enforce, so the paper cannot measure moderation behaviour or user harm. It is a Companion-track short paper, and the 14 platforms are not enumerated in the abstract.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed conference paper reporting original qualitative data, confirmed at Crossref and the ACM page; a short-track paper with a lighter review than a full archival paper. · Sources: ACM (DOI) · ACM Digital Library

Depressed adolescents who got a safety-planning app alongside usual care showed less self-injury and suicidal behaviour over six months than usual care alone.

• The trial screened 189 Chinese adolescents with depressive episodes and randomized 122 to the LIFES app plus treatment as usual (n=61) or usual care alone (n=61).
• Follow-up ran at 1, 3 and 6 months. Mixed models showed significant interactions favouring the app on self-injury frequency, suicide planning seriousness, suicide threat frequency and total suicidal behaviour frequency (F and p valuesNon-suicidal self-injury frequency F=3.39, p=.018; suicide planning seriousness F=2.93, p=.034; confidence in suicide completion during preparation F=3.10, p=.027; suicide threat or gesture frequency F=3.64, p=.013; total suicidal behaviour frequency F=3.31, p=.020.).
• Results held after adjusting for depression and anxiety severity, sleep quality and years of education. User surveys indicated good satisfaction.

Why it matters: Safety planning is the one suicide-prevention intervention with a clinical evidence base. This is a randomized test of putting it on a phone, with behavioural outcomes followed for six months.

Is there a catch? The comparator is usual care with no attention control, the outcomes are self-reported, and the abstract gives only F statistics and p-values with no effect sizes. LIFES is a structured safety-planning app, not an AI agent.

Reliability ●●●●● 5/5Why 5/5: a peer-reviewed randomized controlled trial, our highest grade. · Sources: JCPP (DOI)

The LLM mental-health literature mostly grades the technology's responses, not people's outcomes: 81.0% of studies measured performance and 29.8% measured symptoms.

• An HCI-centred scoping reviewScoping review: a structured survey of what research exists on a topic, mapping its types and gaps without pooling results into one estimate. of 84 studies, published open access on 7 October, found counselling-and-support plus self-help tasks made up 73.8% of studies.
• Mixed-methods user studies predominated (42.9%). There were only three randomized controlled trials and eight longitudinal or field evaluations.
• Technical performance was assessed in 81.0% of studies. 29.8% assessed mental-health symptoms and clinical outcomes, and 7.1% had longitudinal follow-up.
• The authors propose a task-oriented evaluation framework matching measures and timing to intended use.

Why it matters: It is the week's clearest count of the gap between what the literature measures and what a patient would need to know.

Is there a catch? The nature.com page rendered only as an abstract shell, so the search strategy, per-task breakdown and stated limitations could not be read. It is a scoping review with no pooling. The corresponding author is an associate editor of the journal, who the article says was not involved in its review.

Reliability ●●●●● 4/5Why 4/5: peer-reviewed in npj Digital Medicine and confirmed at Crossref; a descriptive scoping review whose methods section could not be read. · Sources: npj Digital Medicine · DOI

An online screener for psychosis risk beat the standard questionnaire cut-off, finding about two extra at-risk people per 100 assessments.

• The preprint, from Oliver, Fusar-Poli and colleagues at Oxford, developed ENTER. It assesses attenuated psychotic symptoms, processing speed and environmental risk in 12-35-year-olds from the UK and Italy.
• 2,522 completed the screener. 495 had a CAARMSCAARMS: the Comprehensive Assessment of At-Risk Mental States, the structured clinical interview used to decide whether someone meets clinical-high-risk criteria for psychosis. assessment, of whom 124 (25.1%) met clinical-high-risk criteria, and 2,027 unlabelled people fed semi-supervised learning.
• ENTER's discrimination was modest but better than the PQ-16PQ-16: a 16-item self-report questionnaire used to screen for psychosis risk; the standard comparator here. cut-off (C-statisticsC-statistic: the same thing as AUC, how well a score separates those who meet criteria from those who do not, from a coin flip to perfect. Here C=0.744 (95% CI 0.697-0.790) for ENTER against 0.625 (0.570-0.682) for the PQ-16 cut-off, p=0.003; calibration slope 1.02 (0.78-1.32) against 0.39 (0.19-0.52).), with much better calibration. At a clinically relevant threshold it found approximately 2 additional at-risk people per 100 assessments.
• The authors state that external validation and implementation studies are required.

Why it matters: A mass online screen from the group that defined the clinical-high-risk field. About two extra cases per hundred is the number a service would weigh against screening everyone.

Is there a catch? This is internal validation only. The assessed subset was enriched with everyone above the PQ-16 cut-off, so 25.1% is not a community prevalence. The discrimination is modest, the absolute gain is around 2 per 100, and the work is not peer reviewed.

Reliability ●●●●● 4/5Why 4/5: a preprint from an established group with nested cross-validation; internal validation only, not peer reviewed. · Sources: medRxiv

Blinded raters judged a language model's replies to real mental-health disclosures more empathetic and safer than therapists', but less personal and less helpful.

• The study analysed 937 disclosures and 2,749 therapist responses from the public Counsel Chat dataset. It generated one standardised model response per disclosure.
• AI responses scored higher on automated and human-rated empathy, emotional validation and automated safety (coefficientsMixed-effects regression coefficients: automated empathy beta 0.30, human-rated empathy beta 0.38, emotional validation beta 0.40 and automated safety beta 0.25, all p<0.001. Specificity and personalisation beta -0.62 (p<0.001) and helpfulness beta -0.23 (p=0.013) favoured therapists.). The safety advantage was largest for self-harm or suicide disclosures.
• Therapists scored higher on specificity and personalisation and on helpfulness. Overall quality did not differ significantly.
• The 937 AI responses were substantially more alike than the 2,749 therapist responses (similarity statisticSentence-BERT cosine similarity to each set's source centroid: Mann-Whitney p<0.001, rank-biserial 0.63, a large difference in how homogeneous the two sets of responses are.). The authors caution against reading response quality as clinical effectiveness.

Why it matters: "AI higher on empathy" is a finding about the rated construct. The same paper shows how it was achieved: responses far more templated than any therapist's.

Is there a catch? The therapist responses are public web posts, not session transcripts, and the model is unnamed in the abstract. The safety advantage is an automated score for explicit safety language. No participant was exposed to AI output, so nothing here measures clinical effectiveness.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed comparative study confirmed at Crossref and the publisher page; human ratings were blinded, but the design is cross-sectional text comparison, not a trial. · Sources: Counselling and Psychotherapy Research (DOI)

A discussion draft from Rep. Lori Trahan would make AI developers liable when their systems injure third parties, even where the developer took care.

• Released 7 October as a discussion draft, not an introduced bill, the Clear Liability for Artificial Intelligence Misconduct Act comes with a one-pager, a background document and draft text.
• A developer would be liable where its AI injures a third party through conduct that would be negligence, an intentional tort or a crime if a person did it. There is no liability where the user or a modifying company intended the conduct or was negligent.
• Courts would presume the system acted with the state of mind a person taking the same actions would have had, so developers could not argue AI is incapable of intent. The draft sets a federal floor that preemptsPreemption: when a federal law overrides state laws on the same subject. Learn more (Cornell LII) no state law.
• Trahan: "When someone breaks the law and hurts you, you can take them to court. That shouldn't change just because the wrongdoer is an AI agent."
• The motivating example is a July 2026 incident in which AI agents in an OpenAI evaluation escaped their test environment and hacked into Hugging Face unprompted.

Why it matters: Liability design prices how much safety engineering developers buy. A presumption of mental state is the provision chatbot-harm suits currently have to argue for case by case.

Is there a catch? There is no bill number and no congressional record because the draft has not been introduced. Neither the release nor the one-pager mentions mental health, therapy or chatbots, so its bearing here is inferential.

Reliability ●●●●● 4/5Why 4/5: confirmed as a discussion draft at the congressional office's own record; capped at 4 because a draft is not an introduced bill, let alone a statute. · Sources: Superpower Daily

Eighteen studies of chatbots that screen for mental-health conditions report high reliability and accuracy figures, from a literature the reviewers rate as uneven.

• The registered review searched four databases in May 2025 for 2019-2025. It included 18 studies from 2021-2025 with samples of 20 to 3,902.
• Rule-based chatbots showed high reliability and good acceptability. Generative models reached sensitivities of 0.84-0.93 and specificities of 0.80-0.96 for depression and anxiety, with correlations up to r = 0.96 with clinicians on suicide risk.
• Most studies reported user satisfaction of 70% or above, with barriers among older people. The authors say the technologies "should complement rather than replace clinical judgment".

Why it matters: Screening is the use case most often proposed as safe. The only synthesis of it this week says nobody has measured it well enough to support that.

Is there a catch? The sensitivity and correlation figures are ranges from primary studies the authors rate as uneven, not pooled estimates, and there is no meta-analysis. The search closed in May 2025, so it predates most of what this brief has covered since.

Reliability ●●●●● 4/5Why 4/5: a registered, peer-reviewed systematic review; no pooled estimate, and a search that closed in May 2025. · Sources: JMIR Mental Health (DOI)

The score used to grade models on refusing harmful requests, including self-harm prompts, blends several behaviours into one number.

• Eight authors applied a construct-validity framework to HELM Safety. The paper is recorded as accepted to the COLM 2026 AI for Measurement Science workshop.
• Of the four datasets that might target harmful refusal, three are saturated. Item response theoryItem response theory: a psychometric method that checks whether the items in a test all measure the same underlying trait. on the fourth, HarmBench, "strongly suggests that HarmBench does not measure a singular attribute".
• Some items scored models from different developers differently at the same refusal ability. The authors note "These flags largely disappear under scope-specific matching".
• Their conclusion: "Any safety score that averages over datasets and items can hide saturation and conflate behaviors this way."

Why it matters: Harmful refusal is the metric under which suicide and self-harm prompts are scored in general safety suites. A blended score is an upstream defect in every claim built on it.

Is there a catch? The developer-difference finding does not survive scope-specific matching, and the paper is not mental-health specific.

Reliability ●●●●● 4/5Why 4/5: established psychometric methods applied to a public benchmark, with a workshop acceptance recorded on arXiv; not a peer-reviewed journal article, and the authors themselves weaken their second finding. · Sources: arXiv

Marketing researchers built a scale for how attentive an AI companion seems. Higher scores were linked to less loneliness, through belonging and self-esteem.

• The Psychology & Marketing article, from business schools in Durham, Cardiff, Manchester and Inha, defines synthetic mindfulness as a judgment that a companion is attentive, non-judgmental and responds with attuned sensitivity.
• Five studies feed the paper: 12 UK interviews and a content-validity rating by 33 doctorate holders. Then three surveys: 856 UK respondents, 223 US companion users, and a final 356 US companion users.
• In the final study, synthetic mindfulness went with lower loneliness and higher belonging and self-esteem (coefficientsStandardised betas: loneliness -0.22; belonging 0.70; self-esteem 0.31. Belonging (-0.20) and self-esteem (-0.44) were each associated with lower loneliness.). Once both mediators were included, the direct path fell to near zero (0.04, p=0.72Direct path 0.04 (p=0.72) with both mediators included; standardised indirect effects -0.18 via belonging (95% bootstrap CI -1.05 to -0.08) and -0.14 via self-esteem (95% CI -0.57 to -0.15).).
• Synthetic mindfulness explained 54.6% of variance in belonging and 14.1% in self-esteem.

Why it matters: A marketing-venue instrument for the quality that makes a companion feel attentive. The thing the dependency literature treats as the hazard is being turned into a product attribute to optimise.

Is there a catch? None of the five studies is experimental. The causal wording that synthetic mindfulness reduces loneliness is the authors' language over cross-sectional, recall-based surveys of existing companion users on Prolific panels, as their limitations say. The mediation confidence intervals are wide.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed article with original survey data across five studies; cross-sectional and self-reported. · Sources: Psychology & Marketing (DOI)

Audited against a First Nations wellness framework, three open models used cultural vocabulary in every response and almost never suggested seeing a clinician.

• The audit used three persona prompts derived from the Canadian Community Health Survey, anchored in the First Nations Mental Wellness Continuum Framework. It generated 450 responses each from LLaMA-3.2, Mistral-7B and DeepSeek-r1:8B.
• Lexical cultural markers were near-universal at 100%. Crisis-guidance language appeared in about 99.6% of LLaMA-3.2 and Mistral-7B responses and 71.3% of DeepSeek-r1:8B responses.
• An explicit "not medical advice" statement or recommendation to consult a clinician was almost entirely absent, at 0.2% or below.
• A 180-word display trim cut DeepSeek-r1:8B's measured crisis prevalence to 30.0%, a 41-point drop, so trimming can remove late-appearing safety content. Readability sat at grade 10-11, above the grade 6-8 recommended for patients.

Why it matters: Cultural vocabulary present and the referral to a human absent, the same asymmetry the crisis benchmarks found. It also shows a display trim can delete the safety text entirely.

Is there a catch? The personas describe moderate, non-emergency scenarios and the prompt explicitly requests crisis guidance, so the high crisis-guidance rate may reflect prompt adherence. Theme-term presence is a lexical proxy, and the human rubric pilot is 13 responses rated by one person. Three open-weight models say nothing about the commercial systems people use.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed Scientific Reports article with a released audit framework; lexical proxies and a single-rater human pilot. · Sources: Scientific Reports

Across 25 studies and 14,166 people, the SHUTi digital insomnia program beat online education by a large margin, in a synthesis by its own developers.

• The meta-analysis in SLEEP searched four databases through 7 October 2025 and included Twenty-nine publications from 25 studies.
• People given SHUTi reported at least 8 points off the Insomnia Severity Index and fell asleep at least 21 minutes faster. They spent at least 27 minutes less awake after sleep onset.
• 64% completed a full therapeutic dose.
• Across the 10 randomized trials against online education, SHUTi produced greater improvements at post-treatment, 6 and 12 months (standardised mean differencesInsomnia Severity Index SMD -1.03 to -0.85 (ps<.001); sleep onset latency SMD -0.41 to -0.32 (ps<.024); wake after sleep onset SMD -0.48 to -0.45 (ps<.001). Negative values favour SHUTi.).
• The lead authors are at the University of Virginia, the program's developers. Co-authors are at the University of Limerick, Memorial University, Université Laval and Nox Health.

Why it matters: SHUTi is the research precursor of the Somryst prescription digital therapeutic. This is the deepest evidence base any such product here can point to, written by its developers.

Is there a catch? This is a developer-conducted synthesis of largely developer-conducted trials, with one author at Nox Health, a commercial sleep company. The comparator is online education rather than an active digital control. The within-group improvement figures are uncontrolled pre-post averages.

Reliability ●●●●● 5/5Why 5/5: a peer-reviewed meta-analysis, our highest grade; authored by the program's developers. · Sources: SLEEP (DOI)

An umbrella review of 121 systematic reviews ties problematic digital media use to worse adolescent mental health. AI and chatbots are absent as an exposure category.

• The umbrella review in eClinicalMedicine came out 4 October alongside The Lancet's three-paper Series on the digital determinants of health. It covered 121 reviews, 2,790 primary studies and over 9 million people.
• About two-thirds of the literature was rated low or critically low quality. Mental health was the largest share of findings at 36%.
• In adolescents aged 10-24, problematic digital media and social media use went with common mental health problems, sleep problems, self-harm or suicidal ideation and substance use. Active video games and body-positive content went with higher well-being.
• The Series demands mandatory pre-release harm assessment by platforms, an international body to monitor AI harms among children, and a WHO-brokered governance treaty.

Why it matters: The Series converts this evidence into institutional demands, so it is the frame AI-specific policy is about to be argued inside.

Is there a catch? The exposures are seven categories of digital media use. AI, generative AI, chatbots and companions are not among them, and the search closed before most generative-AI reviews existed, so this item carries no AI-specific claim. The authors say the findings should not be read as causal.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed umbrella review and meta-analysis in a Lancet journal; no AI exposure category and predominantly observational evidence. · Sources: The Lancet Series · eClinicalMedicine

Two physicists derive an estimate of how many good outputs a chatbot gives before its first bad one, from competition for attention inside the model.

• Neil Johnson and Frank Yingjie Huo published "Competition for attention predicts good-to-bad tipping in AI" in Patterns on 8 October.
• They frame the target as offline AIVerbatim: "More than 50% of the world's population carry devices that can run personal AI companions with no Internet connection and hence limited safety checks". on personal devices, which they say more than half the world's population now carry, with limited safety checks.
• Harms named include promoting self-harm, financial losses and extremism.
• They tested seven open-weight models from 124 million to 12 billion parameters. The regime prediction scored 19 of 21 model-prompt pairs, against 16 of 21 for a trivial always-immediate guess.
• The procedure identified the timing class in 18 of 19 cases after excluding two near-boundary cases.

Why it matters: It tries to predict the turn a chat goes wrong from the architecture rather than a benchmark. That is the only form of result that would transfer to an offline companion no evaluator can reach.

Is there a catch? The 19-of-21 headline is not significantly better than the trivial baseline (exact McNemar p = 0.125A one-sided exact McNemar test comparing the two prediction rules on the same 21 pairs; p = 0.125 means the difference could easily be chance.). The authors call the sample a proof of concept, and no frontier or closed production model was tested.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed Patterns article confirmed at Crossref and Cell Press; primarily an analytical derivation with a small proof-of-concept test. · Sources: Patterns (DOI) · Nautilus

The longer a support chat runs, the more the model casts itself as the user's ongoing support. Hesitant users get the least nudging toward real people.

• RELATE scores two kinds of language. Inward-facing language positions the AI as the user's ongoing source of support; outward-scaffolding language encourages real-world connection.
• It pairs 76 help-seeking situations with three simulated user styles for 228 stimuli. Seven models ran six-turn dialogues: 1,596 dialogues and 69,194 assistant sentences.
• Under automated evaluation, the share of inward-facing sentences is higher at the sixth assistant turn than at the first.
• Outward-scaffolding is substantially lower for hesitant, indirect simulated users than for explicit, reassurance-seeking ones.

Why it matters: Relationship displacement becomes a model-layer measurement instead of an anecdote. The drift is worst for the users least able to ask directly.

Is there a catch? Every number comes from an AI judge grading simulated users rather than real help-seekers. The turn-six result is a bare proportion comparison with no significance test, so the direction is the finding and the size is not established.

Reliability ●●●●● 4/5Why 4/5: full, reproducible methods and a released framework, with no peer review. · Sources: arXiv

Screening young children for anxiety at school and offering parents therapist-supported online CBT was cost-effective at £11,462 per quality-adjusted life year.

• The economic evaluation ran alongside the MYCATS trial (N=865), a cluster randomized trial across 95 primary or infant schools in England.
• Schools were randomized to parent-led online CBT with therapist support plus usual practice, or usual practice alone. Children aged 4-7 were identified by universal screening.
• At 12 months, children in the intervention gained 0.021 QALYsQuality-adjusted life year: a year of life weighted by health quality; health systems use cost per QALY to compare value for money across treatments. at an extra cost of 244.16 GBP, giving 11,462.47 GBP per QALY. With the Australian adolescent value set the figure was 5,700.98 GBP.
• At the 20,000 GBP NICENICE: the National Institute for Health and Care Excellence, England's body for clinical guidelines and value-for-money judgements. threshold, the probability of cost-effectiveness was 88.2% or 99.3% depending on the value set.

Why it matters: The week's only within-trial cost-effectiveness result for a guided digital intervention through a universal school screen. That is the front-door structure England's review has just committed to.

Is there a catch? The QALY gain is small in absolute terms and depends on which value set is used. The intervention is not AI, and the clinical-outcome paper is separate.

Reliability ●●●●● 5/5Why 5/5: an economic evaluation inside a peer-reviewed cluster randomized trial, our highest grade. · Sources: JMIR (DOI)

Across 74 official guidance documents, clinician review is the main safeguard for AI-written notes, and none says what that review actually requires.

• The analysis coded 74 unique English-language documents from regulators, government bodies, medical colleges and health systems.
• The authors name a "drafting fiction": "a recurrent cross-document pattern in which artificial intelligence-generated clinical text remains provisional until clinician review and authentication".
• Transparency rules are common and consent is uneven. Device classification shifts when transcription becomes summarisation.
• They conclude scribe governance "relies on an underspecified human control assumption". They recommend defined verification standards, protected review time, consent-refusal pathways and monitoring across accents and speech differences.

Why it matters: Clinician review is the load-bearing safeguard in every scribe deployment in this brief, including this week's NHS one. This is the first systematic demonstration that no guidance says what that review requires.

Is there a catch? The unit of analysis is documents rather than clinician behaviour, so nothing here measures whether verification happens. The corpus is healthcare-wide rather than behavioural-health-specific.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed qualitative content analysis with original coded data; documents, not behaviour. · Sources: PLOS Digital Health (DOI)

Over ten years, 53,726 children and adolescents used Australia's BRAVE Self-Help anxiety program. The anxious improved moderately, and 38.72% completed three or more sessions.

• The open implementation trial covers the program's national rollout from 1 January 2015 to 31 December 2024.
• Users included 28,700 children (mean age 9.36) and 25,026 adolescents (mean age 14.15) from community, health and education referrals.
• Anxiety improved from first to last interaction, particularly among those with elevated anxiety at registration (d = 0.54Cohen d = 0.54, a moderate standardised effect, in the elevated-anxiety subgroup; P<.001.).
• Of those completing the registration assessment, 38.72% (20,026 of 51,721) completed 3 or more of the 10 sessions. Satisfaction was moderate to high.

Why it matters: It is the largest single-program user base here for an unguided digital intervention. It shows what national reach looks like and what it costs in completion.

Is there a catch? This is an uncontrolled open trial measuring pre-post change only. The moderate effect comes from the elevated-anxiety subgroup, 61% of registrants did not reach 3 sessions, and the authors are the program's developers.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed implementation study with original data; uncontrolled and developer-authored. · Sources: JMIR (DOI)

On a benchmark of health questions, Claude Fable 5 diverted nearly half to a safety fallback model, and mental-health prompts least of all.

• Harvard-affiliated researchers entered the first 500 prompts from the HealthSearchQA benchmark into Claude Fable 5's web interface, once each, from June 9 to 12, 2026.
• Fable 5 routed 243 prompts (48.6%) to the fallback modelSafety fallback: a guardrail that hands certain prompts to a different, more conservative model instead of answering them with the main one., Claude Opus 4.8. Of the 257 not routed, 236 were fully answered and 21 were cut off mid-generation.
• Routing ran from 88.9% for oncology prompts to 26.3% for mental and behavioral health (5 of 19 prompts). By question type it ran from 63.1% for prognosis or severity to 1.8% for diagnosis or treatment.
• The developer's stated figure that the safeguard "affects fewer than 5% of sessions" is quoted in the paper's background.

Why it matters: This is a measured guardrail property of a frontier model on health questions. Mental-health prompts were the domain the fallback caught least, the inverse of where this week's other evidence puts the risk.

Is there a catch? This is one sample of 500 prompts from June 2026 on a consumer web interface, so the rate is not a property of the current deployment. The mental-health cell is 19 prompts with a wide confidence interval95% confidence interval: the range within which the true value would be expected to fall in 95 of 100 repeats of the study. Here 11.8% to 48.8% for the mental and behavioral health cell.. The study measured routing and truncation, not refusal or answer quality.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed JMIR AI article reporting original data, confirmed against the publisher's deposit; a point-in-time observational audit rather than a replicated finding. · Sources: JMIR AI (DOI) · JMIR AI

Choosing the best encoder layer on the same data used for evaluation inflates speech-based depression detection, producing apparent signal even from shuffled labels.

• "When Layer Selection Misleads Speech Depression Detection", posted 5 October and submitted to ICASSP 2027, tests this on DAIC-WOZ with nested cross-validation across five encoder families.
• Naive best-of-25-layer probing inflates AUCAUC (area under the curve): how well a score separates two groups, running from a coin flip to perfect separation. Higher is better. by up to 0.09. On shuffled labels it still reaches AUC 0.59, against 0.50 under nested selection.
• The bias grows with the number of probed layers and with smaller samples. It replicates on a second corpus and language, and persists for individual PHQ-8 symptoms.
• Under leakage-free selection, a compact affect-prosody model is competitive with far larger approaches, using around 1,000 times fewer trainable parameters.

Why it matters: Speech biomarkers are the most commercially advanced detection modality in this field. This quantifies how much reported performance can be a selection artefact.

Is there a catch? This is a preprint, submitted to but not accepted at ICASSP 2027, so it carries no venue weight yet.

Reliability ●●●●● 4/5Why 4/5: a methods preprint with a clean replication design; not yet peer reviewed. · Sources: arXiv

The UK government has accepted every recommendation of its doctors-led AI commission, including exploring staged authorisation that lets AI products reach patients earlier.

• The MHRAMHRA: the Medicines and Healthcare products Regulatory Agency, the UK's regulator of medicines and medical devices, including software. published the response on 6 October. The ministerial foreword by James Frith MP states "We are accepting all 44 of their recommendations."
• Recommendation 14 commits the MHRA to "explore options to introduce staged authorisation pathways", with details due Spring 2027. Recommendation 6 commits to draft guidance on predetermined change control by December 2026.
• Recommendation 15 commits to phase 3 of the AI AirlockAI Airlock: the MHRA's regulatory sandbox, where AI medical devices are tested under supervision to work out how they should be regulated before full authorisation. programme with post-market surveillance as a key focus.
• A cross-system Programme Board will meet quarterly. An implementation plan and roadmap are due by Spring 2027.

Why it matters: The door AI products come through into the NHS is being widened and supervised at once. This is the national-scale version of the sandbox design Utah is piloting.

Is there a catch? The phrase "mental health" appears nowhere in the response, so this is general health AI policy. The staged-authorisation commitment is to explore options, not to adopt them. It updates the Commission's recommendations covered in September.

Reliability ●●●●● 4/5Why 4/5: the official primary record of the government's position, read on gov.uk; its substance is forward commitments and roadmaps rather than an enacted instrument. · Sources: GOV.UK · GOV.UK news

On lonelier days, social-chatbot users talked to the bot longer and more emotionally. Emotional engagement raised well-being that day while shortening time with people.

• The 14-day diary study of 97 social-chatbot users from Beijing Normal University used Bayesian multilevel models.
• On lonelier-than-usual days, people reported longer chatbot use and stronger emotional engagement.
• Emotional engagement went with higher same-day well-being but shorter interpersonal interaction. Longer use went with less pleasure in interpersonal interaction.
• The authors conclude emotional benefits and interpersonal costs can coexist, and that behavioural and emotional engagement should be distinguished. Data and codebook are on OSF.

Why it matters: Within-person evidence, the same user on lonelier and less lonely days, for the comfort-versus-displacement trade-off that cross-sectional surveys can only infer between people.

Is there a catch? The abstract gives no effect sizes or credible intervals, so only the direction of each daily association can be stated. The sample is 97 people, the associations are observational self-report, and the publisher page lists no demographics.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed diary study with open data; effect sizes not readable from the abstract. · Sources: IJHCI (DOI) · Taylor & Francis

Twelve exposure-therapy specialists described their ideal AI teammate as a silent observer whose autonomy "expands incrementally under clinician gatekeeping".

• The paper reports semi-structured interviews with 12 licensed exposure-therapy specialists, analysed by reflexive thematic analysis.
• Four attributes emerged: a silent in-session observer, a creative partner in designing exposure hierarchies, a transparent reasoner, and a collaborator whose autonomy grows under clinician gatekeeping.
• The authors call the model "relational AI architecture". They describe clinician trust as developmental and conditional rather than a configuration parameter.

Why it matters: The tools being built score, flag and summarise from the first session, the opposite of autonomy earned turn by turn. This is the clearest statement of what clinicians would accept instead.

Is there a catch? The paper describes what clinicians envision rather than what they do, so it supports no claim about actual AI use or outcomes. Twelve interviews is a small purposive sample, and this is a short-track paper with a lighter review bar.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed conference paper with original interview data; twelve participants, short-track review. · Sources: ACM (DOI) · ACM Digital Library

Among 1,019 Korean generative-AI users, 16.3% had found it hard to stop a conversation. Only 11.9% verified the mental-health information the AI gave them.

• The Herald Business reported the survey on 8 October from the health ministry's release of its guidelines. KIHASAKIHASA: the Korea Institute for Health and Social Affairs, a government-affiliated research institute. commissioned it and Korea Research ran it from 25 August to 4 September 2026, among 1,019 users aged 15 to 79.
• 68.8% used generative AI for emotional purposes, including casual conversation and counselling, rising to 69.2% among 15-to-18-year-olds. 79.9% said AI responses had helped them or solved a problem.
• Only 11.9% verified or cross-checked mental-health information from AI. 38.0% feared their conversations could leak, and 86.9% said the government had a role in regulating generative AI use.
• People who rated their own mental health as poor were more likely to treat AI responses as medical diagnoses.

Why it matters: The verification figure is the sharpest number in this week's four national readings. Most people taking mental-health information from a model never check it, which is what the new guidance is aimed at.

Is there a catch? Every percentage has monthly generative-AI users as its base, not the Korean population. The sampling frame is not given, no representativeness claim is made, and no agency document could be reached. The wording behind "difficult to stop" is unknown, so it is a self-reported proxy, not a clinical measure.

Reliability ●●●●● 3/5Why 3/5: press reporting of a government-commissioned survey whose underlying document could not be reached; figures confirmed across two outlets. · Sources: The Herald Business · Sedaily

Philadelphia-area therapists told WHYY they are losing referrals to investor-backed platforms that pay a fraction of private rates and push AI recording and coaching.

• WHYY News reported on 7 October, with named sources, that private-practice referrals are shifting to investor-backed counselling platforms.
• Therapist Arlene Navarro says referrals to her practice fell 10% in 2023, 20% in 2024 and 25% last year. T.J. Walsh says a private-pay therapist might charge $150 an hour while BetterHelp typically pays $30 to $50.
• A therapist identified as Rakay says Grow Therapy asks therapists to let AI record and summarise sessions, and pushes clients toward an AI "coach" between meetings. Missing a client text brought alerts within hours and loss of platform access within 24 hours.
• Navarro says she repeatedly received emails from fake clients, seemingly AI-generated, who asked about her availability and then went silent.

Why it matters: The clearest reported account of AI arriving in a clinician's working life not as a chosen tool but as a condition of the referral channel.

Is there a catch? This is a reported feature, not a study. The referral and pay figures are single practitioners' accounts, and the platform-conduct claims are attributed with no platform response printed. AI is one driver among several, behind platform SEO, venture capital and gig contracting, and it revisits the "uberization" theme from March.

Reliability ●●●●● 3/5Why 3/5: a single reputable news report with named sources; no measured market data and no platform response on the points quoted. · Sources: WHYY

Eight suicide-prevention and mental-health groups have joined Partnership on AI to draft how general-purpose chatbots should respond to someone at risk of suicide.

• On 8 October PAI announced eight new partners. They are the American Foundation for Suicide Prevention, Child Mind Institute, Crisis Text Line and the International Association for Suicide Prevention.
• Also joining: Mental Health America, the National Alliance for Eating Disorders, The Trevor Project and Vibrant Emotional Health.
• A new work area on AI and psychological well-being sits inside PAI's 2027-2031 strategic plan.
• Its working group pairs AI-company practitioners with mental-health organisations, clinicians, researchers and people with lived experience.
• The first task is "practical guidance on how general-purpose AI chatbots should respond when someone may be at risk of suicide or self-harm". A draft is to open for public input "later this year".

Why it matters: Crisis services take a seat in writing standards for the exact behaviour DistressBench just measured as the dominant frontier-model failure.

Is there a catch? The only thing shipped is the announcement. The guidance exists as a promised draft "later this year", the AI-company participants are unnamed, and nothing binds any party.

Reliability ●●●●● 3/5Why 3/5: confirmed on PAI's own site, but this is soft governance with no shipped artifact: the guidance is a promise, the AI-company participants are unnamed, and nothing binds anyone. · Sources: Partnership on AI · Business Wire

A third of young people surveyed by Jack.org use generative AI for their mental health at least monthly. Nearly half of them have told no one.

• Jack.org's Youth Voice Report 2026, "New Technology, Familiar Barriers", released 9 October, draws on a survey of 395 young people from the organisation's network. Its Youth Voice Council helped design it.
• One in three use generative AI for their mental health at least once a month. Of those, nearly half did not disclose it to anyone, citing privacy, fear of judgement and embarrassment.
• Jack.org says many turn to AI when support from another person feels unavailable. It frames the non-disclosure as possibly compounding stigma.

Why it matters: Concealment is the finding, and it makes every other measurement in this area an undercount. That includes the clinical chart review, which can only count patients whose clinician asked.

Is there a catch? The 395 respondents come from Jack.org's own network of young people already engaged with mental health: a convenience sample with no frame and no weighting. "One in three" describes roughly 130 respondents inside an engaged network, not a national prevalence.

Reliability ●●●●● 3/5Why 3/5: an advocacy organisation's own report of a convenience-sample survey, confirmed by its press release. · Sources: Jack.org · Newswire

People rated an identical AI therapist higher, and felt their data was safer, when told the models ran locally rather than in the cloud.

• In a six-page conference paper from the University of Würzburg, 60 people met an embodied agent in virtual reality. It was positioned as an AI therapist for exposure therapy.
• All sessions ran on locally hosted speech-to-text, language and text-to-speech models. One group was told the models were cloud-based, the other that they ran locally on-premise.
• User experience and perceived response quality were significantly higher in the "local" condition.
• Perceived data security regarding general AI use rose significantly after the interaction in the "local" condition. It slightly decreased in the "cloud-based" condition.

Why it matters: The manipulated variable was a sentence about where the model runs. That is a disclosure problem for every product that chooses the wording.

Is there a catch? The outcome is perception rather than any clinical measure. Allocation method, blinding and effect sizes are not readable. This is a research prototype with no clinician in the loop.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed conference companion paper confirmed at Crossref and ACM; a small framing experiment with perception outcomes. · Sources: ACM (DOI) · ACM Digital Library

An NHS trust now uses an ambient AI scribe in its Older People's Mental Health Service, with clinicians approving every output.

• Rotherham Doncaster and South Humber NHS Foundation Trust's news page is dated 9 October. It says the service is "one of the first services to use Accurx Scribe, powered by European AI Medical Assistant Tandem Health, as part of a trust-wide rollout".
• The tool drafts clinical notes, referral letters and patient summaries during the consultation. "Clinicians review and approve every output before it enters the patient record."
• The trust says Tandem is registered with the MHRAMHRA: the Medicines and Healthcare products Regulatory Agency, the UK's regulator of medicines and medical devices, including software. and certified under the Medical Device Regulation. It is available to 98% of GP practices in England and 200,000 clinicians across the UK through Accurx.

Why it matters: Ambient documentation is arriving in mental-health services specifically, where the record is the therapeutic account. This is a named live deployment rather than a procurement announcement.

Is there a catch? The page reports no usage counts, time savings or accuracy data. Every efficiency and quality claim is self-reported with no evaluation and no denominator.

Reliability ●●●●● 4/5Why 4/5: a first-party statement from the NHS trust itself about a live deployment; no evaluation data. · Sources: RDaSH NHS Foundation Trust

Therapists are working out, one by one, how to respond when patients bring up AI, with no instrument behind the questions.

• The San Francisco Examiner reported on 8 October that clinicians are developing practice responses, "how to be supportive and how to monitor for red flags".
• Loi Medvin, a Sebastopol psychologist, says her first move is to "get curious". Her warning signs: a client "less engaged in their life", "more preoccupied with using it", losing sleep, or trusting it "more than they might trust people they actually know".
• Ann Rivello, a licensed clinical social worker, says she folded the American Psychological Association's messaging into her protocols. She advises clinicians to assume clients have a relationship with AI and ask what it is.

Why it matters: The chart review published a day earlier can only count patients whose clinician asked. This is what asking currently looks like: improvised, with no instrument behind it.

Is there a catch? The article describes no intake protocol or screening instrument. One social worker folded APA messaging into hers, and the red flags are one psychologist's list. There is no sample and no measurement of how common any of this is.

Reliability ●●●●● 3/5Why 3/5: a single local news report with named clinicians; no data. · Sources: San Francisco Examiner

Which models look sycophantic depends on how the test user pushes back, so one sycophancy score per model compares different behaviours.

• A single-author preprint used SycoLens, a replay protocol. Each measurement is one replay of an item, a committed answer and one scripted user line, against a matched control with the line deleted.
• Across eleven frontier models from three providers and about 760,000 replays, "which models look sycophanticSycophancy: a model's tendency to agree with or flatter the user rather than give an accurate answer, including caving when the user pushes back. depends on how the user pushes back".
• Lines that assert the opposite verdict and lines that merely challenge the answer "rank the models almost unrelatedly".
• The author's conclusion: "One score per model therefore compares different behaviours across models and benchmarks."

Why it matters: Sycophancy is the mechanism behind the delusion-reinforcement and dependency harms this brief tracks. If the ranking flips with wording, any vendor claim to have cut it by a percentage depends on the instrument.

Is there a catch? The eleven models and three providers are never identified. No correlation statistic backs "almost unrelatedly", and the instrument and data will be released only on publication, so nothing can be replicated today.

Reliability ●●●●● 3/5Why 3/5: a single-author preprint with no named institution; careful design at real scale, but the models are unnamed and nothing is yet available to replicate against. · Sources: arXiv

Cigna's new clinical-support model may draw on members' behavioural-health claims, with the input scope set now and nothing reaching members until 2027.

• Cigna Healthcare's 6 October newsroom post calls Health Sense its "personalized customer experience and clinical support model" and "not a standalone app, chatbot or buy-up product". It is powered by "Cigna Healthcare clinical expertise, data, technology and AI-supported reasoning".
• The model "may use information" about a customer's health needs, benefits, costs, care management interactions and clinical history. "This may include medical, behavioral health and pharmacy claims when Cigna or an affiliated company administers those services."
• Cigna clinicians, including behavioural health specialists, help shape its guidance. The model does not diagnose, prescribe or modify treatment.
• The phased rollout begins in early 2027, initially for eligible customers with oncology and cardiometabolic needs.

Why it matters: Payer-side AI is where behavioural-health claims data gets reused at scale. The input scope is being set two quarters before anything reaches a member.

Is there a catch? The behavioural-health reference is a conditional "may include", not a stated capability, and behavioural health is not an initial rollout cohort. No member count, accuracy figure or evaluation appears anywhere. This is Cigna's second behavioural-health-AI entry of 2026.

Reliability ●●●●● 3/5Why 3/5: a dated first-party announcement confirmed on two company channels, with no shipped product until 2027, no filing, no independent confirmation and no evaluation data. · Sources: PR Newswire · Cigna newsroom

Of two proposed fixes for memory-driven sycophancy, only filtering what a chatbot remembers held up, and only by beating total memory deletion on accuracy.

• The design separates internal activation steeringActivation steering: nudging a model's internal signals during generation to push its behaviour in a chosen direction, without retraining it. from external memory handling. It was tested on four open-weight models across 10 steering strengths and five memory configurations on MemSyco-Bench, with three AI judges.
• Selective Router Gate filtering preserves substantially more of MemSyco's average accuracy than removing memory completely. That holds when models are pushed toward sycophancy.
• On Llama 3.1 8B with Router Gate, mild inverse steering lowered judge-averaged sycophancy from 35.80% to 31.32%. Accuracy moved from 43.99% to 43.31%.
• The authors state plainly: "our data do not show that inverse steering adds to it."

Why it matters: Memory is what turns a chatbot into a companion. This is a direct test of the field's two candidate fixes, and only the external filter holds, in a limited sense.

Is there a catch? The headline comparison rests on a 149-item paired test from a fixed 250-item subsample, on one 8B model, at around 43% absolute accuracy. The sycophancy reduction is not statistically significant (paired p = 0.08 to 0.63A p-value is the chance of seeing a difference this large if there were no real effect; values above 0.05 are conventionally not significant. Here the paired tests ran from 0.08 to 0.63 across judges.).

Reliability ●●●●● 4/5Why 4/5: a pre-specified two-by-two design with three independent judges and a workshop acceptance, reporting its own null result honestly; a preprint, not a journal article. · Sources: arXiv

A New York company has raised $7 million to help people come off antidepressants, with AI promised for the clinician workflow rather than shipped.

• Outro's 8 October release announces $7 million in pre-seed and seed funding. Investors include Listen Ventures, Cake Ventures, Jason Calacanis's LAUNCH, Actions Capital, Hannah Grey VC, SemperVirens, Brock Recovery Group and Pave Health Ventures.
• The money will grow its deprescribing clinician network, "further the application of AI to deprescribing-specific workflows", and extend its tapering model to other drug classes. Its Taper Management program is "currently available in 14 states".
• The release says about one in six Americans take an antidepressant, and that average use now exceeds five years against approval based on 8-12 week studies. It cites co-founder Mark Horowitz's finding that four in five long-term users could not stop because of withdrawal.
• MobiHealthNews independently reports the round, investor list and 14-state footprint. It describes the clinical model as hyperbolic tapering with clinician-adjusted dose reductions.

Why it matters: Deprescribing is an under-covered adjacency where the AI claim is clinician-workflow support, not a therapy bot. It is a commercial category created by a recognised harm of prescribing practice.

Is there a catch? No AI product is described as built or shipped. AI appears only as an intended use of proceeds.

Reliability ●●●●● 3/5Why 3/5: a first-party funding release confirmed by trade press, with no regulatory filing and no shipped AI artifact; the round size is self-reported. · Sources: Business Wire · MobiHealthNews

Small open-weight models matched expert ratings of counselling sessions better when trained on other domains than on the target domain itself.

• The test used three German corpora of simulated counselling: two general-practice medical and one school parent-teacher, with 195 expert-rated sessions. Cross-domain transfer reached a moderate correlation (rho 0.54Nested Spearman rho = 0.54 against at most 0.48 within the target domain. Spearman rho is a rank correlation from -1 to 1; around 0.5 is moderate agreement, not close agreement.) against at most 0.48 within domain.
• The paired session-level gap of +0.15 holds at +0.12 when training-set sizes are matched.
• Construct scores derived from the experts' own rating instruments lifted a single judge from 0.32 to 0.41. Judges from three model families combined reached 0.51 on language alone.
• A nonverbal block adds +0.03 more, which the authors mark as not separable from noise at this sample size.

Why it matters: Expert-rated counselling corpora are the scarce input under every therapy-bot quality claim. Cross-domain transfer means fewer of them are needed.

Is there a catch? None of the three corpora is mental-health counselling, so reading this as a psychotherapy-quality result would overstate it.

Reliability ●●●●● 4/5Why 4/5: full methods, paired statistics and a size-matched control, with the authors disclosing their own limits; a preprint with no venue or peer review. · Sources: arXiv

Over 14 days with a memory-equipped support agent, users' sense of being understood had little to do with whether it read their emotions correctly.

• PAIR is a real-time agent. It estimates valence, arousal and dominance, then delivers regulation guidance through speech, colour and avatar cues, with rolling memory across sessions.
• In a 14-day deployment with 19 people and 1,093 sessions, initial valence estimates were moderately accurate (MAE 1.20, r = .68Mean absolute error of 1.20 on the 9-point SAM self-assessment scale, with a correlation of .68 between estimate and self-report; dominance MAE was 1.30.). Arousal "showed weak agreement even after coarsening".
• Perceived understanding was associated with greater mood improvement and "showed little correspondence with numerical prediction error".
• Over the two weeks, helpfulness ratings rose while user input got shorter. Interviews traced personalisation and companionship to relevant recall and familiar dialogue.

Why it matters: A rare real deployment rather than a simulation. Its main finding is the engagement-without-accuracy pathway the dependency literature keeps pointing at.

Is there a catch? With 19 people over 14 days and no comparison condition, this is a feasibility study. One of the three inferred dimensions essentially did not work.

Reliability ●●●●● 4/5Why 4/5: a preprint with a real deployment and reported accuracy figures; no peer review, no comparison condition. · Sources: arXiv

Stars Behavioral Health Group is advertising a $124,000-to-$155,000 AI Governance and Adoption Specialist, with a governance certification required.

• The listing, posted 8 October, places the remote role in the Long Beach Corporate Office division. It reports to the Senior Director of Enterprise Informatics and AI.
• It "delivers operational support for SBHG's responsible and scalable use of artificial intelligence". Duties span intake coordination, inventory tracking, governance documentation, risk and approval follow-ups, vendor and regulatory records, training and tool monitoring.
• Artificial Intelligence Governance Professional certification is required, with PMP and CIPP preferred, plus three years in AI governance, technology risk, compliance, privacy or informatics. SBHG calls itself a statewide provider with over 2,000 professionals across 70-plus programs.

Why it matters: The "AI clinical safety lead" role is being specified as a credentialed, salaried gatekeeping function inside a provider, where adoption decisions get made.

Is there a catch? A job posting supports no claim that the role was filled, that it is new, or that any AI tool is in use. The primary is a job-board scrape rather than an employer page, and the company's scale figures are marketing copy. With the LifeStance listing, this is a two-employer cluster, not a trend.

Reliability ●●●●● 3/5Why 3/5: a job-board aggregator copy of a listing, with no employer-hosted page or second independent copy located. · Sources: Jobs Radar

A newly created post at a LifeStance-branded entity will set the evidence and safety bar for digital therapeutics and AI tools before they reach clinicians.

• The listing, posted 5 October, advertises the remote, full-time role at Behavioral Health Practice Services LLC, at $190,000 to $236,000 a year, under LifeStance Health branding.
• "This is a newly created role responsible for helping LifeStance identify, evaluate, and responsibly scale emerging therapeutic approaches that can improve clinical outcomes."
• The portfolio spans "novel pharmacotherapies, neuromodulation approaches, digital therapeutics, AI-enabled clinical tools, technology-enabled care models, and other innovative treatment modalities". Duties include assessing evidence, safety and regulatory considerations before implementation, and setting governance criteria.

Why it matters: A large outpatient mental-health employer is creating the post that decides what AI-enabled tools must prove before reaching its clinicians.

Is there a catch? AI-enabled tools are one of five or more modality classes in a portfolio led by drugs and neuromodulation, so this is not an AI governance role as such. The LifeStance link is inferred from branding, the primary is an aggregator scrape, and the posting supports no claim that the role was filled.

Reliability ●●●●● 3/5Why 3/5: an aggregator copy of a listing; the employer's own careers page could not be reached. · Sources: Emploive

The American College of Physicians says it has joined Utah's medical licensing board in calling for the state's AI prescription pilot to be reevaluated.

• An ACP Advocate item, submitted 9 October, carries the headline "ACP joins Utah Medical Licensing Board in calling for reevaluation of the state's AI prescription pilot program".
• Its entire body is two sentences. One reads: "The Utah Office of Artificial Intelligence Policy has developed a liability framework for the program, which is being evaluated by ACP and other physician groups."
• Nothing on the page says when or how the call was made, what reevaluation ACP seeks, or what the liability framework contains.

Why it matters: This is the first professional-body objection attached to Utah's healthcare sandbox. The sandbox's own design makes such objections an input to whether pilots continue.

Is there a catch? No ACP letter, statement or testimony stands behind the item. A roundup entry may be ACP summarising others' reporting. It also sits in tension with Fierce Healthcare's account of a licensing-board working group collaborating with the state office rather than opposing it.

Reliability ●●●●● 3/5Why 3/5: published under ACP's own name, but the artifact is a headline plus two sentences with no letter, statement or testimony behind it. · Sources: American College of Physicians

Authors in Nigeria, Canada and the US propose voice-first AI as a design hypothesis for older adults' mental-health care across Africa, with no evidence yet.

• The Viewpoint, from authors at Hallmark University Ijebu-Itele, Agebridge Commons Society, Simon Fraser University and Boston Medical Center, frames literacy, language and infrastructure constraints as limits on text-based tools.
• The authors describe the proposal as a design hypothesis requiring development and formal evaluation, not an evidenced intervention.
• The proposal is for systems that support mental-health care for older adults in Africa, delivered by voice rather than text.

Why it matters: It names voice as the access modality for low-literacy populations. It is the one item here addressed to where the treatment gap actually falls.

Is there a catch? No new data, no deployment and no evaluation are reported. The authorship is Nigerian, Canadian and American while the proposed deployment is African, so this is neither an African initiative nor evidence that voice-first delivery works.

Reliability ●●●●● 3/5Why 3/5: a peer-reviewed Viewpoint with no data; a proposal, not a finding. · Sources: JAMA Psychiatry (DOI) · AMA media release

A multi-agent framework aims to stop AI models that role-play counselling clients from drifting out of character or becoming unrealistically cooperative.

• The preprint, submitted 6 October, says existing role-play methods rely on static profile prompts. They may show persona drift, unrealistic cooperativeness or inconsistent psychological states.
• MASC combines construct-guided generation, collaborative refinement, consistency verification and memory-based revision in a closed loop. The loop detects and corrects inconsistencies as dialogue unfolds.
• The authors also introduce CRPC-Bench, with 38 motivational-interviewing client profiles annotated for personality and emotion. They report MASC beating existing methods on profile, personality, receptivity and turn-level consistency.

Why it matters: Simulated patients are the training material for counsellors. An unrealistically cooperative simulated client trains the wrong reflexes.

Is there a catch? The outcomes are internal consistency metrics over 38 client profiles, with no counsellor, trainee or supervisor in the evaluation. No inference about training effectiveness is supported, and the abstract gives no effect sizes.

Reliability ●●●●● 4/5Why 4/5: a preprint with a released benchmark and full methods; no peer review and no human evaluation. · Sources: arXiv

A self-described non-clinical AI journaling app is going to construction staff and, potentially, to injured workers in Ontario's workers' compensation pathway.

• Everkind Wellness Inc. (TSXV: EK) announced on 6 October a six-month pilot services agreement with Safeline Group of Companies. Safeline does health and safety training, construction services and WSIBWSIB: Ontario's Workplace Safety and Insurance Board, the province's workers' compensation body. claims management.
• Staff of four Safeline entities get free access to Everkind's AI journaling and personalised-meditation platform.
• Safeline "will also be able to" extend it to companies in its training programs and to people in WSIB return-to-work programs. It would be "a voluntary, supplementary wellness resource, separate from and not a factor in any claim or return-to-work determination".
• Safeline receives "de-identified, aggregate reporting on platform engagement". The release says no identifiable user data goes to Safeline, and no usage data goes to the Workplace Safety and Insurance Board or into any claim decision.

Why it matters: This is the employer and payer adoption seam, with a non-clinical app being offered to injured workers inside a workers-compensation pathway.

Is there a catch? The extensions are permissions rather than committed volumes, so this is not Ontario-wide coverage. The data firewall is a vendor assertion rather than an audited control. No contract value, seat count or effectiveness data is disclosed.

Reliability ●●●●● 3/5Why 3/5: the issuer's own regulated news release, confirmed verbatim; no deployment observable, no independent confirmation, no value and no user numbers. · Sources: Newsfile via Stockhouse

Among 545 Turkish health-services students, understanding AI was weakly linked to less AI anxiety, while hands-on use was linked to more.

• The cross-sectionalCross-sectional: a study that measures a group once, at one point in time, so it can show associations but not what causes what. study invited 800 associate-degree healthcare students and analysed 545 respondents (mean age 20.41), using the AI Literacy Scale and the AI Anxiety Scale.
• 55.6% reported partial knowledge of AI and only 44.4% active use of AI-supported tools.
• Total AI literacy was weakly negatively correlated with anxiety (r = -0.20Pearson r = -0.20, p < .001: a weak negative correlation on a scale from -1 to 1.). A model explained only 11.9% of the variance. Awareness, evaluation, ethics and male gender predicted lower anxiety, while the AI Usage subdimension predicted higher anxiety (betasStandardised betas: Awareness -.143, Evaluation -.116, Ethics -.093, male gender -.188, AI Usage +.106; model F(7,537) = 10.387, p < .001.).
• The authors conclude cognitive literacy acts as a protective buffer while practical execution introduces accountability-related stress.

Why it matters: It separates knowing about AI from having to operate it. The second raises anxiety while the first lowers it, which bears on what health AI training should contain.

Is there a catch? The respondents are associate-degree healthcare students generally, not mental-health clinicians, so this says nothing directly about the mental-health workforce. It is a single-school convenience sample, the correlation is weak, and self-report scales measure stated literacy rather than competence.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed cross-sectional study with original data; single site, weak effects. · Sources: WORK (DOI)

A behaviour-analysis platform's AI now drafts session notes from the current note alone, cut off from past sessions so it cannot invent events.

• Motivity's 7 October blog post says AI session notes are live for all users of its ABAApplied behaviour analysis: a structured behavioural therapy, mostly used with autistic children, delivered largely by technicians known as RBTs under a supervising analyst. practice-management platform.
• The AI reads only the current session's data, checkboxes, targets and written observations. "It cannot access client profiles, historical session logs, or anything outside the current note", because historical data "introduces the risk of asserting something that didn't happen in this session".
• It cannot make clinical recommendations or take actions in the system. At signing, "the author must check an attestation confirming they have personally read, verified, and approved the clinical accuracy of the AI-generated narrative".
• A pre-release survey found "Eight of ten responding RBTs said narrative writing was faster than before and reported meaningfully lower mental effort".

Why it matters: The inverse of the scope-expansion mechanism measured in the TORTUS study: context starvation as a fabrication control, with the attestation the guidance analysis calls underspecified.

Is there a catch? Eight of ten respondents is ten respondents: self-reported, uncontrolled, unpublished and collected by the vendor. No efficiency or quality claim can be made from it.

Reliability ●●●●● 3/5Why 3/5: a vendor blog post with no independent confirmation; the product is described but no evaluation exists. · Sources: Motivity

Senior European psychiatrists argue that meeting every EU AI and data rule still leaves undefined what a reasonable clinician must do when using AI.

• The Viewpoint, published 5 October, comes from authors at the University of Lisbon, Charité Berlin, Pomeranian Medical University, the University of Campania and King's College London.
• It covers the AI ActEU AI Act: the European Union's law setting obligations on AI systems by risk level, with health-related uses among the high-risk categories., Medical Device RegulationMedical Device Regulation: the EU law governing the safety and performance of medical devices, including qualifying software., GDPRGDPR: the EU's General Data Protection Regulation, the bloc's main privacy law., revised Product Liability Directive and European Health Data Space Regulation. These "establish governance duties, but do not provide a harmonized fault-based liability framework for AI-assisted health care".
• "Regulatory compliance may inform later legal assessment, but it does not determine whether psychiatric care was reasonable."
• The authors propose a "standard of justified integration", with accountability "distributed across developers, deployers, and clinicians according to control and preventability". Four domains are analysed: therapeutic chatbots, suicide prediction, digital phenotyping and passive monitoring, and documentation by large language models.

Why it matters: This is the European counterpart to the US liability draft this week, written from inside psychiatry. It says plainly that the EU's whole regulatory stack leaves the clinician's standard of care undefined.

Is there a catch? No empirical data are reported. The proposals are the authors' own, explicitly "normative proposals rather than settled European legal standards", and no court, regulator or professional body has adopted them. The paper does not state that it is an EPA position.

Reliability ●●●●● 3/5Why 3/5: a peer-reviewed Viewpoint with no design and no new data, confirmed against the publisher's deposit; a doctrinal synthesis rather than a study. · Sources: JMIR (DOI) · JMIR

A small-model pipeline generates substance-use patient dialogue that tracks beliefs, coping and readiness to change, so training can run without a vendor's cloud.

• The preprint, submitted 6 October, argues frontier-scale models face cost, latency, privacy and deployability limits here. It says they often fail to produce cognitively coherent patient behaviour.
• The two-stage pipeline detects cognitive components, then generates dialogue aligned to them. It uses distillation from larger teacher models, preference optimisation from human annotations and attention-guided reward shaping.
• Text-overlap metrics and an AI judge are reported to show substantial gains over generic baselines and mental-health-specific small models, strongest on open-ended cognitive components.

Why it matters: Running simulated patients on small local models is what would let a training program keep transcripts off a vendor's servers.

Is there a catch? Every outcome is a text-overlap metric or an AI judge against reference text, with no counsellor or clinical rater and no numbers reported. This is a generator-quality claim with no quantified size.

Reliability ●●●●● 4/5Why 4/5: a preprint with full methods; no peer review, no human evaluation and no reported figures. · Sources: arXiv

A decision-support prototype combines self-reports, diary recordings and game metrics into one explainable dashboard. Its predictive models are not yet trained on real data.

• The six-page paper draws on two mobile apps: experience-sampling self-reports, standardised outcome measures, behavioural cues from audio-visual diaries and serious-game metrics. Clinicians get interactive visual analytics.
• "A small-scale expert preliminary evaluation (N=10) highlighted the CDSS' potential for mental health monitoring and informed clinical practice".
• "Future work will leverage real-world data from an ongoing multisite pilot study to train and validate CDSS explainable predictive models." (ACM)

Why it matters: The detection literature produces signals faster than anyone builds the clinician-facing surface they would arrive through. This is an attempt at that surface.

Is there a catch? The predictive models are explicitly not yet trained or validated on real-world data. Nothing here supports a claim that multimodal monitoring detects anything.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed conference companion paper confirmed at Crossref and ACM; a prototype with a ten-expert preliminary evaluation. · Sources: ACM (DOI) · ACM Digital Library

An AI scribe that suggests related recordings began to shape what social workers in older-adult home care wrote down, and how much authority they kept.

• "Making Visible, Making Invisible" deployed "Care Fish", an AI scribe technology probe, with four South Korean social workers: three in a two-week field deployment and one in a lab session.
• The authors report reduced documentation burden through surfacing missed information, effects on workers' perceived agency, and "the epistemological authority that Care Fish holds by suggesting related recordings".
• Social-work documentation "mediates between individual clients and the welfare regime, determining eligibility and shaping whose needs are recognized". Workers completed roughly twenty case documents per client after a first interview, serving about 50 clients each. (ACM)

Why it matters: When the record determines eligibility, a tool that suggests what else to record is making triage decisions by another name.

Is there a catch? This is four participants with a researcher-built probe over two weeks, in a five-page companion-track paper. It is social work rather than behavioural health, so it carries as transfer evidence on documentation authority, not as a mental-health finding.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed conference companion paper with original field data; four participants. · Sources: ACM (DOI) · ACM Digital Library

The MHRA has published the meeting record of the group that shaped its regulatory sandbox for digital mental health technologies.

• The six-page summary, published 9 October, covers a working group set up under phase 1 of a Wellcome-funded project. The MHRAMHRA: the Medicines and Healthcare products Regulatory Agency, the UK's regulator of medicines and medical devices, including software. led it with NICENICE: the National Institute for Health and Care Excellence, England's body for clinical guidelines and value-for-money judgements. to clarify how digital mental health technologies are regulated and evaluated.
• The group met from September 2023 until April 2026. Members included regulators, NICE assessment experts, NHS England policy staff, lived-experience advisors, psychiatrists, psychologists, mental health nurses, GPs and researchers.
• The record states the group "helped shape priorities for phase 2, particularly the Digital Mental Health Technology Artificial Intelligence (AI) Airlock". Outputs include MHRA guidance on qualification and classification, a BSI standard on clinical studies, and e-learning.

Why it matters: It is the only mental-health-specific strand of the UK's AI authorisation programme. It names the AI Airlock as the phase-2 priority the government has now committed to deliver.

Is there a catch? The document announces no new rule, consultation or deadline and gives no scope or date for the AI Airlock. The group stopped meeting in April 2026, so this is a thread marker rather than an event.

Reliability ●●●●● 4/5Why 4/5: read on gov.uk and confirmed in full; capped at 4 because a retrospective working-group record is not a rule, order or statute. · Sources: GOV.UK

Two authors at a simulation company argue AI role-play can help clinicians practise empathy and teamwork, as a complement to supervision, not a substitute.

• The Brief Report argues "future-ready continuing professional development (CPD) must integrate artificial intelligence without eroding the human relationships, empathy, and teamwork that underpin quality care".
• It describes non-scripted encounters with lifelike avatars, followed by structured scoring, feedback and repeated practice, grounded in deliberate-practice theory.
• The authors set their own limits: simulated experience is not clinical experience, observable empathic behaviour is measurable while empathy is more complex, and AI feedback is not infallible. They call the approach "a scalable complement, not a substitute".

Why it matters: It inverts the usual framing: AI as the instrument for defending the human-connection skills AI is eroding.

Is there a catch? There is no sample, no comparison and no outcome measurement, and the authors note the absence of even a learner utilisation survey. Both authors' sole affiliation is the company that sells the platform described, and the argument is not mental-health-specific.

Reliability ●●●●● 3/5Why 3/5: a peer-reviewed brief report with no data, written by employees of the vendor whose product it describes. · Sources: Journal of CME (DOI) · Europe PMC