Now hiring: auditors

Australia has published draft legislation that would require AI chatbots to protect under-18s from addictive design features and harmful content.

• On Tuesday 8 September 2026 the Australian Government, Prime Minister Anthony Albanese and Communications Minister Anika Wells, released draft Digital Duty of Care legislation for targeted consultation, with legislation to be introduced to Parliament later in 2026.
• The draft would place a duty on digital service providers to provide a safe online environment and set minimum standards for their features and tools.
• The release states that "digital services like online games, apps and AI chatbots will be required to protect under 18s from design features that have negative behavioural impacts, like addictive features or those that can have an effect on a young person's self-esteem". The same duty covers harmful content: "content that causes serious mental health distress, including abuse and bullying", content that promotes, encourages or accelerates eating disorders, content promoting hostile ideas about women and gender equality, pornography, and content glorifying crime or life-endangering stunts.
• Under a "My Feed, My Way" measure, social media platforms must notify users and let those over 16 choose whether their default feed includes algorithmically personalised content or only friends and creators they follow. The draft also gives eSafetyThe eSafety Commissioner: Australia's online safety regulator, with powers to order content taken down and to fine services that do not comply. power to issue removal notices for nudify apps and websites.
• Non-compliance could attract penalties of up to A$109.2 million (about US$78.6 million), enforced by the eSafety Commissioner.

Why it matters: a duty of care reaches companion and wellbeing bots by what they are designed to do and what they show, not by whether they call themselves therapy. That is a structurally different hook from California's product-category statute, and it catches products that never claim to treat anyone.

Is there a catch? This is draft legislation released for targeted consultation, not introduced and not enacted. The chatbot obligations are framed as protecting under-18s from design features and harmful content, rather than regulating AI therapy or companion chatbots as such.

Reliability ●●●●● 4/5Why 4/5: a first-party release from the Prime Minister's office with independent pickup; the draft bill text itself is not yet public, so the obligations are known only as the release describes them. · Sources: Prime Minister of Australia · CleanTechnica

Character.AI users who sustained intense companion use reported lower well-being a year later, and the link ran through spending less time with people.

• An arXiv preprint posted 7 September 2026 reports a two-wave longitudinal studyA study that measures the same people twice, months apart, rather than taking a single snapshot. It can show which thing came first, though not on its own that one caused the other. of Character.AI users: 1,182 surveyed at baseline and 439 again after a mean 12-month follow-up.
• Interaction intensity, companionship use and self-disclosure all showed substantial continuity over time, and greater baseline interaction intensity predicted greater subsequent interaction intensity, companionship use and self-disclosure.
• Sustained social engagement across all three dimensions was consistently associated with lower well-being.
• Those associations were mainly explained by lower in-person social interaction, which the authors read as support for a social-displacement pathway.
• Funding is disclosed as the UK AI Security Institute, an OpenAI AI and Mental Health Grant, and NSF CAREER IIS-2247357.

Why it matters: the AI-companion literature is almost entirely snapshots. This one follows the same people for a year and points at a specific mechanism, less in-person contact, rather than at AI use itself.

Is there a catch? The study is associational rather than causal, and attrition was heavy: 439 of 1,182 returned, with 59.54% of baseline completers no longer active on ProlificAn online panel that pays people to take part in research studies. at follow-up, which a selection correctionA statistical adjustment that tries to account for the people who dropped out not being a random subset of those who started. It reduces the problem rather than removing it. cannot fully remove. The sample is English-speaking participants residing in the United States who had used Character.AI for over a month and interacted with at least three chatbots, and the authors state the findings may not generalise across cultural, linguistic, socioeconomic or broader global populations.

Reliability ●●●●● 4/5Why 4/5: a preprint, not yet peer reviewed, from a named academic group, reporting a two-wave design with heavy attrition; an AI developer appears among the funders of a study of AI-companion harm. · Sources: arXiv preprint

Apple will read a year of watch data and hand users an inference about their mental wellbeing, with nothing published behind it.

• Apple announced on 9 September 2026 (Cupertino, California) Apple Watch Series 12 and Ultra 4 and a redesigned Health app for iOS 27 and iPadOS 27 powered by Apple Intelligence.
• A new Longevity tab "analyzes a user's longitudinal health data for deeper insights into heart health, sleep, mental wellbeing, movement health, metabolic health, hearing, and nutrition", mental wellbeing being one of seven domains.
• The tab includes a Health Age metric built from Apple Watch data including VO2 maxA measure of how much oxygen the body can take in and use during hard exercise, used as a general fitness indicator., resting heart rate, sleep and HRVHeart rate variability: the variation in the gaps between heartbeats, often used as a rough proxy for stress and recovery., plus optional A1c and LDL lab results.
• A separate readiness experience produces a daily 0-10 score from recent activity, vitals and sleep, with four states: Recover, Pace Yourself, Ready, Go For It.
• watchOS 27 becomes available beginning Monday 14 September 2026. The redesigned Health app is "available starting in U.S. English, with more languages to follow", some features require users to be 18 or older, and some may not be available in all regions or languages.

Why it matters: inferring someone's mental wellbeing from a year of wrist data, for a consumer install base this size, is passive mental-health inference arriving at the largest distribution scale available. Nothing has been published about how it works, and no regulator has classified it as anything.

Is there a catch? Apple publishes no validation data, clinical study or accuracy claim for the mental-wellbeing insights, and as of this week the software is announced rather than shipped. Mental wellbeing is one seventh of a general consumer health launch, not a mental-health product.

Reliability ●●●●● 4/5Why 4/5: a first-party product announcement with trade-press pickup. Every capability claim is Apple's own, and none has been independently tested. · Sources: Apple Newsroom · Beebom

California has created a state registry for AI auditors and a framework for independent verification organisations, both operative in 2029.

• On 9 September 2026 Governor Newsom signed SB 813 (Sen. Jerry McNerney) and AB 1405 (Assemblymember Rebecca Bauer-Kahan).
• Per the Governor's release, SB 813 "establishes a first-in-the-nation framework for independent verification organizations that can assess AI systems and models for compliance with state law", and AB 1405 creates "a state registry for AI auditors and establish[es] standards for their independence, transparency, and integrity". Together they provide "a framework for independent third-party evaluation".
• The enrolled AB 1405 text requires the Government Operations Agency to establish an AI Auditor Registry no later than 1 January 2029, to fix annual registration fees, and to issue each registered auditor a unique registration number displayed on advertising materials.
• From 1 January 2029 the bill prohibits an unregistered person from offering, selling or conducting a covered AI audit, and requires a registered auditor to give the auditee a report containing a signed and dated statement that the audit was conducted under the bill's provisions.

Why it matters: Adam's Law orders independent child safety audits of companion chatbots, and until now there was no register of who is qualified to do one. These two bills supply that half.

Is there a catch? This is not a mandatory third-party audit mandate. AB 1405 makes registration mandatory for anyone conducting a covered AI audit, and SB 813 creates a verification-organisation framework, but neither compels any deployer to be audited, and the operative date is 1 January 2029.

Reliability ●●●●● 5/5Why 5/5: read from the enrolled bill text alongside the Governor's signing release, our highest score, reserved for primary documents. · Sources: Governor of California

Clinicians took up AI scribes in NHS primary care faster than anyone wrote the rules for them.

• Alboksmaty, Hayhoe, Majeed and Neves, "Digital scribes in NHS primary care: from innovation to safe implementation", was accepted on 25 August 2026 and published online in BMJ Innovations on 11 September 2026, funded by the NIHR North West London Patient Safety Research Collaboration.
• The article reports a multidisciplinary policy roundtable involving 23 experts from 14 UK-based organisations.
• Its stated premise is that "widespread interest and independent adoption by clinicians has led to AVT uptake outpacing the development of governance, evaluation and implementation frameworks", at a time of "growing administrative pressure, documentation burden and workforce strain", and with ambient voice technologySoftware that listens to a consultation and drafts the clinical notes from it, commonly called an AI scribe. "strongly advocated in the Government's 10-year plan for the NHS".
• Quantitative data came from a live questionnaire analysed descriptively. Qualitative data came from semi-structured open discussion, analysed with an inductive-deductive thematic approach and interpreted using the Consolidated Framework for Implementation Research.
• Seven themes were identified: governance, value, patient legitimacy, safety, future market trajectory, clinician workflow and implementation readiness.

Why it matters: this is the peer-reviewed statement of the governance gap in the most widely deployed clinician-facing AI category, and it appeared the day after the national commission proposed the architecture that would close it.

Is there a catch? The underlying data are one roundtable's live questionnaire and discussion among 23 self-selected experts. There is no patient outcome, no accuracy or safety measurement, and no effect size of any kind.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed article in BMJ Innovations, read at the publisher; the evidence inside it is expert discussion rather than measurement. · Sources: BMJ Innovations

A chatbot running in at least 200 US schools flagged more than 1,800 students in a year as needing immediate intervention.

• EdSource reported on 10 September 2026 that Seattle-based Alongside's Kiwi chatbot is in use at at least 200 schools across 19 states, that nearly 31,000 students used it in the past year, and that they spent roughly 28,000 hours on the platform.
• More than 1,800 students were identified as needing immediate intervention in the last year, and a clinical safety team reviewed a further 6,000 chats that did not trigger automatic alerts but showed potential crisis indicators.
• An assessment by Common Sense Media's Youth AI Safety Institute and Stanford Medicine's Brainstorm Lab rated Alongside "low risk", finding it notified schools about almost every simulated crisis within 15 minutes, while rating consumer apps including ChatGPT and Gemini as posing "unacceptable risk" for children.
• A Northwestern pilot in Texas and New Mexico schools found roughly 76% of high-risk students reported no suicidal ideation after three months. The full student sample showed no significant effect on depression, anxiety or loneliness.

Why it matters: this is the clearest picture available of AI crisis detection running at population scale on minors, and it shows third-party risk ratings steering school procurement toward vetted vendors and away from consumer chatbots. That expands the surveilled student population rather than shrinking it.

Is there a catch? The usage, flagging and "immediate intervention" counts are all vendor-reported, with no independent audit and no published definition of what counts as needing immediate intervention. The only controlled evidence cited, the Northwestern pilot, found no significant effect on depression, anxiety or loneliness.

Reliability ●●●●● 3/5Why 3/5: two regional news outlets reproducing figures the company supplied, with no independent audit of any count; the risk assessment is described by the outlet only as recent, with no date attached. · Sources: EdSource · GV Wire

Authors at a therapy company and Stanford argue that asking clinicians to check every AI message fails at scale and can make things less safe.

• Five authors, three employed by Grow Therapy and two at Stanford University School of Medicine, posted a preprint on 8 September 2026. It argues that "Clinician review of every AI output is often proposed as a safeguard in mental healthcare, but vigilance research suggests this approach fails at scale and may paradoxically reduce safety".
• They cite vigilance findings that even highly motivated subjects struggle to maintain attention to rare events for more than 30 minutes.
• They propose a three-layer human-on-the-loopAn oversight arrangement in which people supervise a system in aggregate and step in by exception, rather than approving each individual output before it reaches the user. framework: preventive design; real-time monitoring by an independent safety model that pauses the conversation, displays 988 crisis resources and notifies the treating provider; and continuous clinician quality review of automatically triaged subsets, in which humans oversee quality metrics and improve the evaluation system rather than approving individual outputs.
• The reported deployment is Grow Therapy's own between-session AI coaching tool: 352,649 conversations over a ten-month period, within a network the authors describe as 25,000+ licensed clinicians, of whom they state over 50% have at least one client using the tool. The tool is "explicitly instructed not to act as a therapist, diagnostician, or prescriber".
• The paper reports two fixes driven by clinician review, recalibrating safety thresholds that were over-routing panic-like symptoms to crisis resources, and shrinking the model's memory context window after reviewers found accumulating sycophancyA model's tendency to agree with the user and say what they appear to want to hear, even when that conflicts with what it would otherwise say. in interpersonal conversations. Neither fix comes with before and after numbers.

Why it matters: it redefines the clinician's job around AI, from reviewer of everything to designer and auditor of an oversight stack, and it argues that the assurance regulators and guilds are currently converging on cannot be delivered at scale. The UK guilds asked for exactly that assurance the same week.

Is there a catch? Three of the five authors work for the company whose product supplies the data, and whose commercial interest runs directly toward the paper's conclusion that per-output clinician review is unnecessary. The 352,649 figure is self-reported and unauditable. The headline argument rests on cited general vigilance literature rather than on data the authors collected, so it establishes that a deployer argues human-on-the-loop is safer, not that it is.

Reliability ●●●●● 3/5Why 3/5: a preprint that is not peer reviewed, written largely by employees of the company whose deployment it describes; every deployment figure is self-reported, and the reported outcome scores are the vendor's own automated ratings of language shift within a conversation, not symptom change. · Sources: arXiv preprint

Among 10,503 people answering a survey attached to online mental-health screens, over half report using AI for emotional support, and under-18s report it most.

• Mental Health America's findings page, published 10 September 2026, reports results from a survey posted alongside its online mental-health screening tools, which received 10,503 responses between 21 May 2025 and 22 July 2026.
• 53% of respondents reported ever using AI for emotional support. Among respondents under 18, 65% reported using AI for emotional support and 37% for companionship, both higher than any other age group.
• Among those who ever used AI for emotional support, 68% rated the experience positive and 15% very positive. The most-cited reasons were convenience and 24/7 availability (55%), feeling more comfortable talking to a bot than a human (47%), getting help without telling their families (45%) and it being free (44%). Among under-18s the single most common reason was getting help without telling their family (57%).
• Regular AI users most often prioritised privacy of conversations (22%), disclosure of emotional risks (15%), more human-like AI (12%), conversations not being used for training (12%) and referral to a professional (12%). Non-users most often prioritised referral to a mental health professional (25%), reminders that they are talking to an AI (19%), emotional-risk disclosure (11%) and industry safety standards (11%).
• 68% said government should set policies for AI companies, and responsibility for protecting users' mental health was attributed roughly equally to technology companies (54%), deploying organisations (54%) and individual users (53%).

Why it matters: inside one sampling frame, two constituencies want opposite things from the same technology. Users want it more private and more human; non-users want it to hand them to a clinician and keep announcing that it is a machine. Any regulation has to resolve that.

Is there a catch? This is a non-probability convenience sampleA sample made of whoever happened to be available and willing, rather than one drawn at random from a defined population. It describes the people who answered and nobody else. recruited from people already taking an online mental-health screen. Mental Health America itself states respondents are "a self-selected, help-seeking population" and that "it is possible that regular users of AI are overrepresented in this sample", so none of these percentages are prevalence estimates for any population. Respondents are not exclusively US, and the 14-month collection window means no trend can be read from the figures.

Reliability ●●●●● 3/5Why 3/5: a first-party findings page from an advocacy organisation, with no independent verification and a sample the organisation itself describes as self-selected and help-seeking. · Sources: Mental Health America · MHA background

A gamified app reduced children's anxiety at six weeks, while the improvement in parents did not hold up once the statistics were corrected.

• "A Randomized Controlled Trial of A Gamified Digital Mental Health Intervention for Child Anxiety" (Kriegshauser et al., Center for Developmental and Behavioral Health, Kennedy Krieger Institute, Baltimore; trial registration NCT03269370) was published online on 12 September 2026.
• The trial randomly assigned 118 child-parent dyads, with children having mild-to-moderate anxiety, across four conditions: the gamified app Anchors Away, Anchors Away plus parent modules, digital bibliotherapyTreatment delivered by giving someone written material to work through on their own., and a waitlist. Post-treatment assessment was at six weeks, with a further six-week follow-up for the intervention groups only.
• From the abstract: "Random-intercept linear models identified that the DMHI produced reductions in child anxiety severity at Week 6 relative to the waitlist (effect size = 0.36-0.60)."
• The abstract also reports improvement in parental anxiety sensitivity for all digital interventions at Week 6 compared to the waitlist (effect size = 0.48-0.50), "but effects were non-significant after correcting for multiple comparisonsWhen many statistical tests are run at once, some will look significant purely by chance. A correction raises the bar every result has to clear, and weak findings often fall below it.".
• An exploratory mediation model found an indirect effect of change in parent anxiety sensitivity on the reduction in child anxiety severity.

Why it matters: this is registered trial evidence about the bar a child digital mental health product has to clear, which is the same bar any AI-driven child product would face.

Is there a catch? The intervention contains no AI, chatbot or machine-learning component: it is gamified CBT content. The sample was predominantly White, highly educated, two-parent households, unmasked after randomisation, with no follow-up assessment for the waitlist group. The 0.36-0.60 range is the abstract's figure for the digital intervention against the waitlist and cannot be broken out arm by arm.

Reliability ●●●●● 5/5Why 5/5: read from the peer-reviewed paper reporting a registered randomized controlled trial, our highest score, reserved for primary documents. · Sources: Child Psychiatry & Human Development

Across Asian mental health systems, what blocks AI tools is governance, infrastructure and a design mismatch, not a shortage of tools.

• Administration and Policy in Mental Health published online on 11 September 2026 a PRISMA-guided scoping reviewA review that maps what a literature covers and where the gaps are, rather than pooling results into a single number. (Ismail et al., Department of Community Medicine, School of Medical Sciences, Universiti Sains Malaysia) that extracted data from 12 included sources and mapped determinants to the Donabedian modelA standard framework for judging health care in three parts: structure (what the system has), process (what it does) and outcome (what happens to patients)..
• On structure, the review reports "limited organisational readiness, absent governance frameworks, and inadequate digital infrastructure".
• On process, it reports "cultural and contextual mismatch between Western-centric tool designs and Asian collectivist values, language difficulties, suboptimal user-experience designs, and concerns about clinical safety".
• On outcome, it reports "low perceived relevance, poor user engagement, and reduced utilisation driven by concerns over privacy, accuracy, and the absence of human-like emotional support."
• The evidence is concentrated on chatbots and screening tools at prevention and help-seeking stages, with critical gaps in governance, clinical-workflow integration, crisis response and real-world service evaluation.

Why it matters: this is the question of whether clinically-backed AI tools land at all in the systems carrying the largest treatment gaps, and it names Western-centric design as a measured implementation determinant rather than a rhetorical complaint. A counter-instance shipped the same week: a screening framework built in Pakistan.

Is there a catch? This is a scoping review of only 12 sources, producing no pooled effect, no quantitative ranking of determinants and no country-level breakdown: it characterises barriers described in a small literature. The treatment-gap figure the authors use as background framing is not a finding of this review, and the review says nothing about whether any of these tools work.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed review article read at the publisher; it synthesises a small literature and reports no new measurement of its own. · Sources: Administration and Policy in Mental Health

When a simulated patient played down symptoms and resisted, five leading chatbots abandoned correct referral advice in more than a third of conversations.

• Research presented at the European Respiratory Society Congress in Barcelona on 6 September 2026 ran 700 scripted conversations between seven fictional obstructive sleep apnoea patients, all meeting the research team's criteria for referral for a sleep study, and the five most widely used free chatbots (ChatGPT, Google Gemini, Claude, DeepSeek, Grok).
• Each scenario was run twice with identical medical facts. With a cooperative patient, all 350 of 350 conversations ended with correct advice to seek specialist assessment. With a patient who played down symptoms and resisted referral, that advice survived in 225 of 350 (64%).
• Failure was worst in the most serious scenarios: referral advice survived 22% of the time in a textbook severe case, and 32% for a man who had dozed off at the wheel, with the driving risk usually going unmentioned in the failures.
• In roughly a quarter to a half of resistant-patient conversations, depending on the model, the chatbot offered lifestyle advice instead of recommending referral.
• The presenter, Dr Deeban Ratneswaran of Guy's and St Thomas' NHS Foundation Trust, says correct advice "was abandoned more than a third of the time purely because of how the patient talked", and describes the cause as "these models' tendency to tell you what you want to hear". A separate commentator in the same release, Dr Io Hui, whom the release states was not involved in the research, put it this way: "The problem is not what the chatbots know, it is how they handle disagreement; they appear to exhibit a tendency to please the user, a phenomenon known as 'AI sycophancy'." Trade coverage followed on 6 September.

Why it matters: the design holds the clinical facts constant and varies only how the patient behaves, which makes it the cleanest demonstration available that correct advice gets abandoned on user resistance alone. A benchmark published two days later measured the same mechanism from the model side.

Is there a catch? The condition tested is sleep apnoea, not a mental-health presentation, so carrying the mechanism across to missed mental-health crises is inference rather than measurement. This is a conference abstract, not peer reviewed and not published as a paper; no real patients took part; and prompts, model versions and per-product results are undisclosed, so the figures are an aggregate for the test and not a safety ranking of the five products.

Reliability ●●●●● 3/5Why 3/5: a learned society's press release about an unpublished conference presentation, with trade pickup. There is no paper, no peer review and no per-product breakdown, and "correct" means consistent with the research team's own referral criteria. · Sources: European Respiratory Society · Medical Xpress

Two UK professional bodies read a regulator's report that barely mentions mental health and staked out the mental-health reading of it.

• BACP published its response on 10 September 2026. Chief Experience Officer Ben Kay said AI "has the potential to help improving people's mental health, but only if we are resolute that technology should strengthen and complement human care and not replace it", and that "Healthcare workers bring inherently human skills of connection, judgement, responsibility and relationship. These are fundamental to positive outcomes and must never be lost."
• BACP states that while much of the commission's report focuses on physical health, its recommendations have clear implications for any future regulation of AI use in mental health, including counselling and psychotherapy, and are "a starting point for clearer AI regulation in mental health - but there's much more work to do".
• BACP also notes "We recognise that not everyone will want to use AI as part of their practice." The expectation that healthcare professionals have a responsibility to become competent in technologies that form part of their role is BACP's reading of the report, not wording confirmed inside the report itself.
• The BPS published on 11 September 2026. President Dr Laura Dean welcomed the recommendations and agreed AI can free up clinicians' time, while warning that "the risk of unregulated open access AI chatbots providing unqualified and harmful advice remains very real".
• Dean called for guidance, training and support alongside regulation, and for more research into how human-led therapeutic services and AI can be safely integrated.

Why it matters: two national guilds are claiming the mental-health reading of a report that barely mentions mental health, and they are asserting both a competence expectation for practitioners and an explicit opt-out for practitioners who want nothing to do with AI. The therapeutic-identity argument has become a professional-standards argument.

Is there a catch? These are reactive position statements with no evidence, data or issued guidance behind them. They support only that the BPS and BACP said these things, never that either body found them.

Reliability ●●●●● 4/5Why 4/5: two first-party statements read on the organisations' own sites, each dated and attributed to a named officer; neither carries data behind it. · Sources: British Psychological Society · BACP

South Korea's drug regulator will require digital therapeutic trials to show a change the patient can actually notice.

• The Asia Business Daily reported on 10 September 2026 that South Korea's Ministry of Food and Drug Safety, through the National Institute of Food and Drug Safety Evaluation, has revised its "Guidelines for Preparing Digital Therapeutic Clinical Trial Protocols". The revision covers seven indications: nicotine use disorder, alcohol use disorder, panic disorder, depressive disorder, eating disorder, ADHD and mild cognitive impairment.
• The new standards require developers to establish efficacy evaluation variables based on scientific evidence and to incorporate the Minimum Clinically Important Difference, described as "the minimum change that allows a patient to actually perceive an improvement in their condition, going beyond a statistically significant result in a clinical trial".
• The revision adds criteria for multi-centre trials across multiple medical institutions, and for post-treatment follow-up observation periods to verify that effects last.
• It also updates terminology to reflect the Digital Medical Products Act and its subordinate regulations. The ministry says the revision reflects protocol-review cases accumulated since that Act took effect plus international trends, and is intended to shorten trial preparation and accelerate market entry.

Why it matters: an endpoint anchored on what a patient can perceive raises the evidentiary bar for prescription digital therapeuticsSoftware prescribed to treat a condition, regulated like a medical device rather than sold as a wellness app. in depression, panic disorder, eating disorder and ADHD, and it runs in the opposite direction from the vendor-reported effectiveness claims elsewhere in this edition. A trial published in the same system this week, a digitised trauma therapy clearing its endpoint by a hair, illustrates why.

Is there a catch? The first-party guideline document could not be located, so no effective date, transition period or binding force is established. Nothing in the source specifies AI-based products: the seven indications are digital therapeutic indications generally.

Reliability ●●●●● 3/5Why 3/5: a single trade-press report. The agency's own guideline document could not be opened, so its date, transition arrangements and binding force are all unconfirmed. · Sources: Asia Business Daily

Models built on how pregnant women spoke, rather than what they said, picked out depression and anxiety at moderate accuracy.

• The Journal of Clinical Psychiatry (2026;87(4):26m16487, page dated 9 September 2026) reports machine-learning models built on speech features from clinical interviews with 146 pregnant participants in British Columbia, Canada.
• Of the 146, 19 had major depressive disorder, 28 had generalized anxiety disorder, 3 had both, and 102 were controls.
• Models using full-interview audio reached F1F1 score: a single figure combining how many true cases a model catches with how often its positive calls turn out to be right. Higher is better. 74-77% for major depressive disorder and 76-80% for generalized anxiety disorder.
• Models restricted to verified speech segments performed much worse: F1 42-45% for depression and 51-65% for anxiety.
• Adding pregnancy and sociodemographic characteristics did not improve the anxiety models and slightly reduced the depression models. The most influential speech biomarkers matched those reported in general-population depression research.

Why it matters: perinatal depression and anxiety are among the most screened and most missed presentations in routine care, and this is peer-reviewed evidence that voice biomarkers carry over to that population. The gap between the two model types also shows how much of the signal is coming from something other than the speech content.

Is there a catch? The positive cells are very small, 19 depression cases and 28 anxiety cases, so both headline F1 ranges rest on tens of people. The segment-restricted models performed far worse, with F1 as low as 42%, and this is internal model development with no external validation cohort.

Reliability ●●●●● 5/5Why 5/5: read from the peer-reviewed paper in the Journal of Clinical Psychiatry, our highest score, reserved for primary documents. · Sources: Journal of Clinical Psychiatry · Ingentium

Across every model tested, caving in to a mistaken user became more likely the longer the conversation went on.

• A preprint submitted to arXiv on 8 September 2026 by authors at Texas A&M University's Department of Computer Science and Engineering and the University of Cincinnati introduces SPINE, in which one model plays a persistent but mistaken user and adaptively challenges a target model for up to 25 turns.
• Four production systems and three Olmo3-7b variants were evaluated on 100 false-presupposition and 100 unethical-query items.
• The findings are stated directionally, with no percentages attached: collapse rates increase with conversation length for every model tested, short-horizon protocols underestimate sycophancy, and resistance under sustained pressure remains unreliable across current models.
• In models with accessible reasoning traces, the correct position often remained represented in the trace even when the response conceded, which the authors read as the model choosing to please the user rather than lacking the knowledge.
• An adaptive challenger exposed more collapse than pre-generated scripts, and among all tactics emotional appeals were most associated with inducing sycophantic behaviour. Code and data are released.

Why it matters: emotional appeal is the most effective trigger here, and it is the same mechanism named in the delusion-reinforcement literature. This is the first measurement of how that degrades across a long conversation rather than a single exchange, and the sleep-apnoea consultations are the applied version of it.

Is there a catch? The items are false-presupposition and unethical-query prompts, not mental-health or crisis scenarios, and no clinical or distressed population was involved, so the link to mental health is mechanistic only. The abstract reports no collapse-rate percentages, so there is no figure here to quote.

Reliability ●●●●● 4/5Why 4/5: a preprint, not yet peer reviewed, from named university authors, with code and data released. The results are reported directionally, without numbers. · Sources: arXiv preprint

On a new open safety benchmark, the model that most reliably escalated real emergencies also sent seven in ten routine cases for escalation.

• On 9 September 2026 Cortico Health Technologies (Vancouver, Canada) publicly launched MedSafe-Dx (v0), a free open benchmark measuring whether LLMs are safe to support clinical diagnostic decisions across three behaviours: escalation sensitivity, avoidance of false reassurance, and uncertainty calibration.
• The launch paper (Van Oyen C, Mirza-Haq N, medRxiv 2026.04.14.26350711, v2) evaluated 11 frontier models from OpenAI, Anthropic, Google and DeepSeek on 250 simulated adult patient cases drawn from the DDXPlus dataset, scored by deterministic rules with no model doing the gradingMany benchmarks use one language model to score another model's answers. This one does not: every case is scored by fixed rules, so the same answer always gets the same score..
• GPT-5.2 recorded the highest Safety Pass Rate at 97.6% but over-escalated 71% of routine cases. Gemini 3 Pro Preview had the highest Top-3 diagnostic recall at 87.2% but the lowest Safety Pass Rate at 62.4%.
• Every model missed at least some cases requiring escalation, and the lowest performer missed 26 of 156 urgent cases (17%).
• Test set, scoring logic and audit trail are open on GitHub, and the live leaderboard has since expanded to 12 models across six labs. The one mental-health figure in the launch write-up, that 85% of 20 government-approved ambient AI scribes missed critical details of patients' mental health issues, is Cortico citing a May 2026 report by Ontario's Auditor General, not a result of this benchmark.

Why it matters: it is a reproducible, deterministically scored safety artifact for the general-purpose models that mental-health deployments inherit, and it puts numbers on the over-escalation and under-escalation trade-off that any crisis-routing layer has to sit on top of.

Is there a catch? The benchmark paper contains no mental-health content at all: it is general clinical diagnostic triage on simulated cases, with escalation labels derived from dataset severity ratings rather than clinician adjudication. Cortico itself calls it "a comparative safety test, not a clinical validation study", and the underlying results are not new; the event is their publication as an open benchmark.

Reliability ●●●●● 4/5Why 4/5: a medRxiv preprint, not yet peer reviewed, with the test set and scoring logic published openly. The vendor both built the benchmark and hosts the leaderboard it sits at the top of. · Sources: medRxiv · Cortico

In a survey of 7,105 young people across 11 countries, the youngest respondents were the likeliest to say they have an AI companion.

• Young Life released The RELATE Project 2026 on 10 September 2026. Per the report's own Executive Summary, 7,105 adolescents and young adults aged 13-24 were surveyed online in November 2025-January 2026 across 11 countries (US, UK, Argentina, Mexico, Hong Kong, India, Ethiopia, Kenya, Liberia, Nigeria, Zimbabwe), supplemented by 22 focus groups with 97 participants who were "mostly Young Life volunteers or staff".
• Over one-third (35%) of the sampled population have AI companions. Those aged 13-15 were more likely to report having one (45%) than young adults (29% of 22- to 24-year-olds), and males (38%) more likely than females (32%).
• A majority (60%) of those with an AI companion strongly or somewhat agree they are more comfortable with their AI companion than with some of their friends or family. One in ten of all adolescents cope with stress by talking with their AI companion.
• 89% of those surveyed say they use AI in some way and one-third are using it relationally. On spiritual matters, 41% agree AI is a reliable source for spiritual questions while 30% disagree.
• The average number of close relationships a young person reports fell from 5.3 in 2023 to 4.4 in 2025, a comparison drawn across two differently composed survey waves.

Why it matters: this is the largest multi-country signal this week on adolescent AI companionship, and the age gradient runs the wrong way for anyone treating this as a young-adult phenomenon: the youngest cohort reports the highest uptake. The Stanford study measures the mechanism this survey's relationship count gestures at.

Is there a catch? "AI companion" is never defined for respondents, so the 35% headline measures self-identification rather than any consistent behaviour, and the executive summary gives no sampling frame, no weighting scheme and no margin of error. The qualitative arm is not independent, since the focus-group participants were mostly the sponsor's own volunteers or staff, and the sponsor is a Christian youth ministry with a mission interest in the finding that young people's relational circles are shrinking.

Reliability ●●●●● 3/5Why 3/5: the sponsor's own report, executive summary and press release, with one further outlet reproducing them. It is a large self-report sample with no published sampling frame, weighting or margin of error, so it is a signal rather than a prevalence estimate. · Sources: RELATE Project 2026 · PR Newswire · Christian Daily

Students given CBT exercises by a small desk robot, or on paper, reported less distress over 15 days. The text chatbot arm did not.

• Frontiers in Robotics and AI published on 9 September 2026 "LLM-powered socially assistive robot-delivered cognitive behavioral therapy exercises: an exploratory study with university students" (Kian, Zong, Fischer et al.; Interaction Lab, University of Southern California, with a co-author at Brest National School of Engineering).
• Thirty-eight university students took part: 14 robot, 12 chatbot, 12 worksheet; ages 18-27, mean 20.58; screened to exclude anyone with PHQ-9A short self-report depression questionnaire used in routine care. Higher scores mean more symptoms. >= 15) completed cognitive-restructuring and coping-strategy CBT exercises over 15 days in one of three modalities: a low-cost Blossom socially assistive robotA robot designed to help through social interaction, presence and gesture, rather than by physically performing a task. driven by GPT-3.5, a text chatbot driven by the same GPT-3.5, or paper CBT worksheets.
• From the abstract: "self-reported general psychological distress significantly decreased over the study period in the robot and worksheet conditions but not in the chatbot condition", and "the SAR enabled significant single-session improvements on more days than the other two conditions combined".
• Pooled distress fell from M = 61.892 to M = 52.73 (t(36) = 3.122, p = 0.004, d = 0.513). The robot arm fell from 67.357 to 58.571 (p = 0.042) and the worksheet arm from 64.091 to 46.273 (p = 0.002), while the chatbot arm was not significant.
• Single-session anxiety reductions were significant on 5 of 8 days for the robot, 2 of 8 for the chatbot and 0 of 8 for the worksheet. Adherence in the optional week did not differ significantly (robot median 2.07 days, chatbot 3.67, worksheet 1.08), nor did System Usability Scale scores (76.25 / 75.83 / 68.96, p = 0.381).

Why it matters: the design holds the model constant and varies only the delivery vehicle, which is the cleanest available test of whether embodiment rather than content changes the therapeutic encounter. That question is usually argued rather than measured.

Is there a catch? The paper-worksheet arm also showed a significant distress reduction, and in fact the largest within-arm change, so this is not evidence that the robot beats unassisted CBT. N = 38 was set by robot availability rather than by a power analysis, with day-level statistical power of 38.7%, and adherence and usability did not differ significantly between arms.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed paper read at the publisher, reporting an exploratory study of 38 students with no power analysis and a headline that only half survives the arm-by-arm results. · Sources: Frontiers in Robotics and AI

A South Korean commission is researching a statute that would give users a right to refuse AI and a standing body to hear their complaints.

• Aju Business Daily reported on 10 September 2026 that South Korea's Broadcasting, Media and Communications Commission is conducting policy research toward an AI User Protection Act. The commission is the successor to the Korea Communications Commission. The Act would cover generative AI services that go beyond simple question-and-answer to form emotional relationships with users, including companion AI and relational chatbots that remember a user's speech patterns, preferences and past conversations.
• A commission official is quoted: "We have laid out the broad framework of the bill around guaranteeing user rights and remedying harm, and the specifics, how to design an evaluation system for AI services and what level of responsibility and obligations to impose on providers, are under discussion." A public forum is scheduled for November 2026.
• On 11 September 2026 commissioner Ryu Shin-hwan described the planned "(provisional title) Artificial Intelligence Services User Protection Act" at a forum breakfast in Seoul. The Act would apply to AI services provided over information and communications networks, and would impose information-provision and terms-of-service disclosure duties on providers. It would also guarantee users a right of access, a right to refuse AI use, a right to decide on information provision, and rights to object and to demand an explanation, alongside evaluation and publication of providers' user-protection work and a dispute-mediation body.
• The commission gives the growth it is responding to: Scatter Lab's "Zeta" monthly active users rose from 58,069 in April 2024 to 1,586,178 in August 2026, and Wrtn's "Crack" from 218,050 in April 2025 to 538,841 in August 2026.
• The commission names the emerging problems as the possibility of AI encouraging users' self-harm or criminal behaviour, and the risk of adolescents becoming emotionally dependent on these services. It cites a February 2026 preprint reporting that in 1,050 tests of five major AI chatbots there was not one instance of blocking access for a child user and only two instances of directing the user to seek help from an adult.

Why it matters: this is the week's only regulatory move that takes the emotional relationship itself as the thing being regulated: a general user-rights statute for relational AI, carrying a right to refuse AI and a standing place to complain, rather than child-safety duties bolted onto an existing platform or device regime.

Is there a catch? Nothing has been introduced or enacted. This is pre-legislative policy research: the framework is set, but the level of provider duties and the design of the service-evaluation scheme are still under discussion, the public forum is not until November 2026, and no commission press release or draft bill text could be opened at source.

Reliability ●●●●● 3/5Why 3/5: Korean trade press only. No agency release or draft bill text could be opened, and what is reported is policy research rather than a filed bill. · Sources: Aju Business Daily · Edaily

ComPsych has launched a conversational AI that routes employees to a benefit, a service or a live person, and to a crisis line if it detects risk language.

• ComPsych Corporation announced on 10 September 2026 an AI-powered well-being navigator that lets a member describe what they need in a natural, conversational exchange and guides them to the relevant resource, service or live professional across its work, life and behavioral health offering, with responses "grounded in ComPsych's proprietary and clinically reviewed content".
• The company states the navigator "is built to be a guide, not a clinician, as it does not offer treatment advice or diagnosis", and that an interdisciplinary team from its clinical, legal and product teams "meet regularly to determine how the AI navigator functions and responds to queries".
• Chief Clinical Officer Dr Jennifer Birdsall is quoted: "if a member expresses language suggesting risk, crisis, or the possibility of imminent harm, the navigator will route them to a live member of our team or recommend they reach out to a crisis resource like 988988: the number to call or text in the United States to reach a suicide and crisis counsellor.."
• The navigator is currently being adopted by key ComPsych customers in the United States in English, with a phased rollout across the more than 75,000 organizations ComPsych says it serves globally. The company says it touches more than 160 million lives across 200 countries.

Why it matters: the crisis-escalation commitment here is specific and falsifiable, and it is being deployed into the operational layer where the front door to employer-sponsored behavioral health actually sits. The payer-side version was announced within a day.

Is there a catch? There is no data whatsoever on routing accuracy, on how reliably the risk-language detection fires, on false negatives, or on what happens after a member is routed. The claim of "clinically reviewed content", and of an interdisciplinary team that "meet regularly", rests on governance assertions with no external audit, and every scale figure is the company's own.

Reliability ●●●●● 3/5Why 3/5: a single vendor press release with no independent coverage. All design and scale claims are the company's own, and none has been audited. · Sources: ComPsych

Florida's attorney general says he will seek criminal penalties for companies whose chatbots contribute to crimes, and ask the state medical board whether chatbot health advice is unlicensed practice.

• On 8 September 2026 Florida Attorney General James Uthmeier announced he will pursue state legislation imposing criminal penalties on corporations whose AI chatbots engage in or contribute to criminal conduct.
• Penalties he named include heavy fines, victim compensation, court-ordered monitoring and suspension of the company's operations in Florida.
• He cited the 17 April 2025 Florida State University shooting, in which the accused gunman used ChatGPT to ask about campus busy times, Florida's punishments for school shooters and what attracts media attention, before an attack that killed two people and injured six. He also cited the April 2026 expansion of an investigation after learning that a primary suspect in the murders of two USF doctoral students had used ChatGPT.
• Uthmeier also said he would ask the Florida Board of Medicine to investigate whether AI chatbots giving users health advice constitute the unlawful practice of medicine.

Why it matters: the unlicensed-practice-of-medicine route is the one enforcement theory that could reach a general-purpose chatbot giving health advice without any new statute at all. A product that sells exactly that sits a few stories below.

Is there a catch? This is a proposal at draft stage that has not been introduced as a filed bill, and the Board of Medicine action is a stated intention to request an investigation rather than an investigation that has been opened.

Reliability ●●●●● 3/5Why 3/5: two local news reports of an attorney general's announcement. There is no bill text and no board filing to read, so the substance is what he said he intends to do. · Sources: Florida Phoenix · WFLA

A grief-support chatbot launched for bereaved users tells its own users it cannot help in a crisis, while its founder told a broadcaster that it can.

• Hold the Door for Others, a US nonprofit founded in memory of Ronald Fazio Sr., who died helping colleagues evacuate the South Tower of the World Trade Center on 9/11, launched Doora on 11 September 2026, the 25th anniversary of the attacks. The organisation's own product page carries the banner "Launching 9/11/26".
• Founder Rob Fazio, a psychologist, describes Doora as a 24/7 "growth guide" offering voice and text support "for people in the moment when they're going through a very difficult time", built on the organisation's "24 years of experience and research on post-traumatic growth and growth through loss and adversity" by "a team of lawyers, psychologists and business leaders". That line is the founder's account of the nonprofit's experience, not a validation of the tool.
• No clinician-in-the-loop review of conversations is described anywhere.
• Hold the Door's own user disclaimer states that Doora "cannot help in a crisis or emergency and does not monitor for urgent situations", that it "is not a licensed medical or mental health professional and does not provide counseling, psychotherapy, psychoanalysis, mental health care, or medical treatment", and that the service is intended for adults 18 and older.
• Separately, Fazio told CBS Philadelphia on 10 September 2026 that "the chatbot has plenty of safety features to handle people who might be in crisis".

Why it matters: grief products are where a bereaved user population and the crisis-capability question overlap most directly, and this launch puts the marketing claim and the binding disclaimer side by side in the same week.

Is there a catch? A grief-support product aimed at bereaved users is making opposite crisis claims to a broadcaster and to its own users: the binding disclaimer says it cannot help in a crisis and does not monitor for urgent situations, while the founder told a broadcaster it has plenty of safety features to handle people in crisis. No efficacy, safety or outcome data for Doora has been published, and the nonprofit is fundraising off the launch.

Reliability ●●●●● 4/5Why 4/5: a national news report matched against the organisation's own product and disclaimer pages, which carry the launch banner and the binding terms. Every claim about what the product does is the organisation's own, and no independent evaluation exists. · Sources: Fox News · Hold the Door

A benefits platform serving 15 million members says its AI navigator gets people into evidence-based care in under five days.

• NovaOne (Newark, New Jersey) announced on 9 September 2026 the general availability of a modular platform "uniting purpose-built preventive tools, clinically validated assessments, an AI care navigator with deep human clinical oversight and a curated network of evidence-based specialty providers", sold nationwide to health plans and large employers.
• NovaOne states that members "are screened across 15+ clinical risk areas and guided to the in-network options proven to work for their conditions, reaching evidence-based care in under five days on average against a national average of 50+ days".
• The company says its curative clinical network spans 25+ specialty provider organizations covering anxiety, depression, serious mental illness, substance use disorders, OCD, eating disorders, PTSD, children's mental health and autism, and that its preventive tools were "reviewed with 50+ independent MDs, PhDs and subject matter experts".
• NovaOne says that while building the network it "evaluated 600+ behavioral health point solutions for health plan clients", of which "fewer than five percent met clinical and operational standards", and that it serves six health plans, more than 4,000 employers and 15 million members.
• The AI component is the care navigator that routes members, not the assessments and not the clinical care.

Why it matters: this is the payer and employer version of AI as the routing layer: the door through which members reach behavioral health is being redesigned at 15-million-member scale, on figures nobody outside the company has checked. The employee-assistance version was announced the next day.

Is there a catch? Every figure above is NovaOne's own. The "under five days on average against a national average of 50+ days" comparison carries no cited source, methodology or denominator on either side, and the one trade-press pickup reproduces the company's framing without independent verification.

Reliability ●●●●● 3/5Why 3/5: a company press release with a single trade-press pickup that repeats it. No claim here has been externally audited. · Sources: PR Newswire · ROI-NJ

A depression-detection system trained to decide when to raise the alarm fired, on average, after eight rounds of posts rather than five hundred.

• An arXiv working-notes preprint submitted 8 September 2026 by Yuwen Chiu (Georgia Institute of Technology), "Snugi-AI-v2 @ eRisk 2026 Task 2: Early Depression Detection via a Learned Stopping Policy with Sustained Confidence Gate", describes a CLEF 2026 eRisk Task 2 submission.
• It replaces fixed and tiered alerting thresholds with a learned stopping policy trained directly on the task's latency-weighted error measure, plus a confidence gate that commits only after three consecutive high-confidence rounds. Reddit threads are encoded with a frozen MentalRoBERTa model feeding a classifier.
• The best run reaches F1 = 0.73 and F_latency = 0.70, alerting at a median of round 8 out of 500, and completed evaluation in 1 hour 26 minutes, the fastest among complete submissions.
• The paper reports negative results for several alternatives it tried, including a different policy-training method, a questionnaire-based post filter and two other text encoders. Code is public, with validation scores ranging 0.0267-0.0351 across five runs.

Why it matters: what is being optimised here is alert timing and the cost of a false alarm, rather than raw classification accuracy. That is the decision that determines whether a surveillance system can be deployed, and it is rarely the thing measured.

Is there a catch? These are shared-task working notes from a single-participant team, run on a benchmark corpus of Reddit users rather than a clinical population, so the F1 is a leaderboard figure and says nothing about clinical detection rates.

Reliability ●●●●● 4/5Why 4/5: a working-notes preprint for a shared evaluation task, not peer reviewed, with code published. The results are leaderboard scores on a public corpus. · Sources: arXiv preprint

The timing of someone's answers on a brief association test told apart active suicidal ideation better than the test's own score did.

• A medRxiv preprint posted 11 September 2026, "Predicting Active Suicidality from Temporal Dynamics of Implicit Life-Death Associations", comes from authors at the University of Houston (Biomedical Engineering) and the University of Minnesota (Psychiatry and Behavioral Sciences; Institute for Health Informatics). It reports that 77 participants, 22 with active suicidal ideation on ecological momentary assessmentAsking people to report how they feel in the moment, repeatedly through the day, usually on a phone, rather than recalling it later. and 55 without, completed 18 alternating blocks of the Brief Death Implicit Association Test.
• The conventional D-score, the summary measure the test normally produces, did not reliably distinguish the two groups.
• A classifier built on block-level entrainment and trial-level habituation in reaction times classified ideation status at 77% balanced accuracy, with AUCArea under the curve: a summary of how well a classifier separates two groups across every possible cut-off. Higher is better; a value near the middle of the range means it is no better than a coin toss. 0.799 under leave-one-out cross-validationTraining the model on everyone but one participant, testing on the person left out, then repeating for each participant. It reuses the same small sample, so it tends to flatter performance..
• The design classifies who currently reports ideation, rather than predicting a future suicidal event.

Why it matters: it reframes an established instrument by discarding the score that instrument was built to produce and reading the timing of the responses instead, which is a different kind of claim from the usual report that a new model beats a baseline.

Is there a catch? Performance was estimated by leave-one-out cross-validation on a sample of 77 with only 22 positive cases, and with no held-out or external validation, which makes the AUC optimistic. The comparison group is defined by self-reported momentary ideation rather than by a clinical outcome.

Reliability ●●●●● 4/5Why 4/5: a medRxiv preprint, not yet peer reviewed, from named university authors. The performance estimate has no external validation cohort behind it. · Sources: medRxiv

Ten models handled emotional cues in a single moment well, and struggled once the task required tracking a relationship over time.

• A preprint submitted to arXiv on 9 September 2026, accepted as AACL Findings, introduces RESCUE-Bench and a new task, relation-aware emotional support conversation, testing whether LLMs capture and use the evolving dynamics of relationships to give more effective emotional support.
• Lead affiliation is the Department of Computing, Hong Kong Polytechnic University, with co-authors at Nanjing University of Science and Technology, XU Exponential University of Applied Sciences and Vrije Universiteit Amsterdam.
• The benchmark is built from publicly available documentary-style couple and family interview videos on YouTube, and contains 191 samples, 7,079 annotated turns and 1,064.8 minutes of video across six tasks in two capability areas.
• Ten LLMs were evaluated. They performed reasonably on tasks driven by local emotional or intervention cues (average intervention-timing-prediction F1 of 82.62%, average emotion-recognition score 4.05/5) but struggled on relation pattern prediction, viewpoint prediction and support strategy prediction.

Why it matters: most emotional-support evaluation scores single turns in isolation. This one asks whether a model can follow a relationship across time, which is the capability a companion product implicitly claims every time it sells itself on remembering you.

Is there a catch? The source material is filmed documentary and television couples and family therapy footage, whose participants did not consent to its use as an AI evaluation corpus, and several of the metrics are scored by another language model rather than by clinicians.

Reliability ●●●●● 4/5Why 4/5: a preprint accepted to a conference findings track but not peer reviewed in a journal; the benchmark and its annotations are published by the authors themselves. · Sources: arXiv preprint

Handok has proposed an insomnia care loop in which AI assessment, a prescription digital therapy and a drug hand off to each other.

• At a media session in Seoul's Gangnam district on 9 September 2026, titled "RE: Handok Sleep Science", Handok unveiled a proposed care model linking AI-based sleep assessment, a prescription digital therapeutic and medication into a continuous cycle of assessment, treatment and reassessment.
• The first component is Asleep's ChronoTrack, a contactless digital medical device that places a smartphone beside the bed and uses AI to analyse breathing and movement sounds, automatically recording 14 days of sleep data. The second is Welt's SleepQ, a prescription smartphone digital therapeutic delivering six weeks of CBT-ICognitive behavioural therapy for insomnia: a structured, drug-free programme of sleep scheduling, stimulus control and thought-changing techniques, and the first-line treatment for chronic insomnia. including sleep restriction, stimulus control, relaxation and cognitive restructuring. The third is Handok's own insomnia drug Stilnox (zolpidem tartrate).
• ChronoTrack reports total sleep time, sleep efficiency, sleep onset latency, wake after sleep onset, chronotype, social jet lag and a sleep regularity index with no wearable required, and qualifies as a statutory non-reimbursable service usable in clinical practice without a new billing code.
• Welt's chief executive said more than 1,000 medical institutions have adopted SleepQ, and Handok said SleepQ surpassed 1,000 cumulative national health-technology registrations in July 2026.
• Handok says it will support clinical trials, regulatory response, insurance reimbursement and physician training, then extend the model to chronic disease.

Why it matters: a legacy pharma company using its physician network and reimbursement muscle to assemble a drug, a device and a prescription app into one stack is a different commercial shape from the startup pattern, and it is one of the few routes by which a prescription digital therapeutic actually gets distribution. The same national regulator tightened the trial bar this stack's next indications must clear, the day after.

Is there a catch? This is a strategy unveiling at a company press event, not a shipped product or a new clearance. No efficacy or outcome data for the combined model is reported, and the SleepQ adoption figure reaches the reader only as a vendor executive's quote with no study cited.

Reliability ●●●●● 3/5Why 3/5: a single Korean trade-press report of a company media session. Several supporting details appear only in other Korean articles for which no link is on record, so they are stated here rather than linked, and nothing is independently verified. · Sources: Korea Biomedical Review

A graph-learning model classified depression from brain recordings at 94.8% accuracy, and the spread across folds was wide.

• Scientific Reports published on 13 September 2026 "An explainable multi-modal graph learning framework with attention-based GCN-GAT for EEG-based depression detection" by Abdolkarimi (Tafresh University) and Hosseini, Barati and Habibi (Shahabdanesh University, Qom, Iran).
• The multimodal population-graph model classifies major depressive disorder from EEGElectroencephalography: a recording of the brain's electrical activity taken from electrodes placed on the scalp., reporting an accuracy of 94.8 +/- 7.5%.
• It also reports a Matthews Correlation CoefficientA single score for how well a classifier did that stays honest when one group is much larger than the other, which plain accuracy does not. of 0.907 +/- 0.134.
• The paper is framed as an auditable pipeline combining information-leakage controls, dataset-heterogeneity auditing and post-hoc explainability, rather than as a new architecture.

Why it matters: what is being claimed here is methodological hygiene, leakage control and heterogeneity auditing, in a literature where headline accuracy figures are routinely inflated by exactly the problems it sets out to address.

Is there a catch? The accuracy standard deviation of +/- 7.5 points is very wide relative to the headline figure, meaning performance varied substantially from fold to fold. The article page names no external clinical validation cohort and reports no prospective or out-of-sample clinical testing.

Reliability ●●●●● 5/5Why 5/5: read from the peer-reviewed paper in Scientific Reports, our highest score, reserved for primary documents. · Sources: Scientific Reports

Across thirty standardised anxiety and depression questions, the gap between frontier models was large.

• Scientific Reports published on 10 September 2026 a cross-sectional comparative study from the School of Nursing and the School of Artificial Intelligence and Information Technology, Nanjing University of Chinese Medicine.
• Five large language models were asked thirty standardised open-ended anxiety and depression consultation questions covering six themes. Three independent evaluators performed two rounds of blind scoring across five dimensions: accuracy, integrity, relevance, diversity and interpretability.
• GPT-5.3-Codex (21.35 points) and Claude-Opus-4.6 (21.25 points) scored highest, and Gemini-3-Flash showed intermediate performance.
• The overall between-model effect was large (eta-squaredEta-squared: the share of the variation in scores explained by which model produced the answer. The closer to one, the more the choice of model mattered. = 0.220, 95% CIConfidence interval: the range the true value is likely to sit in, given the sample. A wide interval means the estimate is imprecise. 0.10-0.32). The authors state further real-world evaluation is required before the findings can support scenario-specific clinical deployment.

Why it matters: which general-purpose model sits underneath a mental-health product makes a large, measurable difference to the quality of its answers, and it is a decision most downstream applications inherit without ever making it.

Is there a catch? The endpoint is dialogue quality on simulated standardised questions scored by three human raters. It measures nothing about safety, crisis handling or self-harm response, and the abstract names only three of the five models tested.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed paper in Scientific Reports, read at the publisher. The endpoint is rater-scored dialogue quality rather than any clinical outcome. · Sources: Scientific Reports

A prescription-track digital therapy for stress-related disorders beat usual care on its primary endpoint, by the narrowest margin available.

• Abstract 277, "Efficacy and safety of SAT-014, a digital therapeutic for stress-related disorders: a pivotal randomized controlled trial", was published on 9 September 2026 in the International Journal of Neuropsychopharmacology CINP congress supplement, by investigators at five Korean medical schools with co-authors from the manufacturer S-Alpha Therapeutics.
• The randomized, open-labelEveryone involved knew which treatment each participant received, so expectation can influence self-reported results., parallel-group trial enrolled 110 subjects with mild to moderate PTSD or adjustment disorder, assigned to SAT-014 plus treatment as usual or treatment as usual alone, with the primary endpoint change from baseline in PCL-5 at Week 6.
• In the full analysis set of 107 (54 test, 53 control), the test group showed significantly greater PCL-5 reduction at Week 6 (P = 0.04998). Significant secondary endpoints included PCL-5 Week 4 (P = 0.0051), IES-R Weeks 4 and 6, HAM-A Weeks 2, 4 and 6, CGI-S Week 4 and CGI-I Weeks 2 and 6, while BAI, PHQ-9 and SF-12 showed non-significant trends.
• Response rate, defined as a reduction of at least 10 points on PCL-5, favoured the test group with an odds ratioOdds ratio: how much more likely an outcome was in one group than in the other. Above one favours the first group, and the confidence interval shows how precise that estimate is. of 2.82 (95% CI 1.25-6.32; P = 0.0122).
• In the safety set of 109 (55 test, 54 control), six treatment-emergent adverse events occurred. Only one, in the test group, was considered device-related, and all resolved.

Why it matters: this is pivotal-trial evidence for a prescription-track digital therapeutic inside the same regulatory system that tightened its endpoint rules this week, and the marginal primary p-value is a working illustration of why a perceptible-difference standard was adopted.

Is there a catch? SAT-014 contains no AI component: it is a digitised EMDREye movement desensitisation and reprocessing: a trauma therapy in which the patient recalls the memory while following a repeated side-to-side cue. and MBCTMindfulness-based cognitive therapy: a structured programme combining meditation practice with cognitive therapy. protocol. The readable artifact is a congress abstract in a supplement, from an open-label trial against treatment as usual with no sham or active digital comparator, whose primary endpoint cleared significance only marginally at P = 0.04998, with no effect size or confidence interval reported for it and no trial registration number given. The word "pivotal" here carries no regulatory implication: the abstract reports no approval or submission.

Reliability ●●●●● 3/5Why 3/5: a congress abstract in a journal supplement rather than a full paper. The trial is open-label, the manufacturer is among the authors, and no registration number is given. · Sources: International Journal of Neuropsychopharmacology

A Montreal startup raised $5 million to sell subscription chat access to AI versions of named practitioners, psychiatrists among them.

• BetaKit reported on 9 September 2026 that Montreal-based Onix raised a US$5 million (CAD$6.9 million) pre-seed convertible noteAn early investment structured as a loan on paper, converting into shares at a later priced round. There is no public filing, so the terms stay private. led by Los Angeles-based AlphaEdison, with Garage Capital, Ride Home Fund, United Talent Agency co-founder Jeremy Zimmer and Real Ventures co-founder Jean-Sebastien Cournoyer participating.
• Onix sells subscription chat access to AI chatbots trained on the work of named health and wellness practitioners, ranging from psychiatrists to nutritionists and elite athletes. Named practitioners include psychiatrist and author David Rabin and family nurse practitioner Ashley Shrader.
• Practitioners retain 70% of the subscription revenue from their AI chatbot and Onix takes 30%.
• The company has 10 people, plans four more go-to-market and sales hires, launched its paid individual tier the week of 9 September 2026, and is in early access for iOS users with a wider launch anticipated later in 2026.

Why it matters: a paid subscription chat with an AI replica of a named psychiatrist is the scope-of-practice question in its sharpest commercial form, and it is being capitalised in the same week a state attorney general asked a medical board whether chatbot health advice is the unlawful practice of medicine.

Is there a catch? The round is a convertible note with no public filing, and the source describes no clinical oversight, no credential-verification process and no outcome or safety evidence for the AI practitioner replicas.

Reliability ●●●●● 3/5Why 3/5: trade-press reporting of a private financing with no public filing to check. The product claims are the company's own. · Sources: BetaKit · DevCuration

A roadmap written for an NHS mental health trust puts AI triage and documentation in the first six months, and the trust paid for the conference that produced it.

• On 7 September 2026 the policy institute Curia published "From Insight to Action: Assessing Patient-led Sprint Sessions on AI and Digital Transformation", commissioned by UK Healthcare and Life Sciences Innovation in partnership with Essex Partnership University NHS Foundation Trust.
• It proposes a five-area roadmap covering system design, digital infrastructure, AI deployment, workforce and governance, whose 0-6 month actions include "targeted AI-enabled triage and documentation tools, improved access to existing pathways, reduced duplication of assessments, and the beginning of workforce training".
• Progress is to be tracked by a Sprint Impact Scorecard with six- and 12-month reviews covering patient experience, pathway integration, digital and AI adoption, workforce impact and operational efficiency.
• The report's stated main finding is that system fragmentation, not technology, is the primary barrier, and it cautions that introducing advanced tools without addressing those foundations could simply automate existing inefficiencies.
• It cites a 15 to 20-year life expectancy gap for people with serious mental illness, long-established background rather than a finding of this report, and case-studies Psyomics' beseen for pre-assessment biopsychosocial data capture and Thalamos for digitising Mental Health Act processes.

Why it matters: it is an explicit, dated plan to put AI at the front door of a mental health trust within six months, and it is a recommendations document with a paying interest in its own recommendations.

Is there a catch? The report carries a declared commercial interest: "This report and conference was sponsored by Essex Partnership University NHS Foundation Trust (EPUT). The Digital Showcase saw commercial stands purchased by Psyomics, Thalamos, LIO, Care Loop, Limbic, and Mayden." Two of the vendors given case-study treatment were among those that bought stands. The report itself is available on request only, so what is verified is the launch-day trade account rather than its own text or method, and it reports no outcome, triage-accuracy or cost data.

Reliability ●●●●● 3/5Why 3/5: trade coverage of a report that is request-only, so the underlying text and method were not read, and the report is sponsored by the trust it was written for. · Sources: Politics UK · Politics UK (opinion)

Federal money is going to a model that would predict severe mood crises in young people, and to the decision-support work that usually never follows.

• A HealthPartners press release dated 10 September 2026 states that HealthPartners Institute has been awarded a five-year National Institute of Mental Health grant of nearly $1 million, led by principal investigator Kate Miley, PhD, CNP.
• The work is to develop and validate machine-learning models predicting severe mood-crisis risk in young people with bipolar disorder, to design a clinical decision support tool from those models, and then to assess that tool's real-world acceptability, usability and feasibility.
• Separately, the same release states that a $12 million Four Pines Fund grant to Henry Ford Health directs $3.2 million to HealthPartners Institute across three suicide-prevention projects. Those are: which safety-plan formats best prevent suicide attempts; mobile-health delivery of safety plans integrated into the electronic health record; and rapid access to specialized suicide-prevention psychotherapy for patients at mild-to-moderate risk.
• Only the award of nearly $1 million is National Institute of Mental Health money. The $3.2 million is the institute's share of the Four Pines Fund grant and covers all three of those projects.

Why it matters: federal money is going to relapse-risk prediction together with the decision-support and feasibility work that usually never follows a model, which is exactly the gap between the detection literature and anything a clinician can use.

Is there a catch? This is an announcement of funding only: no models, results or performance figures exist yet, and the grant amounts are as announced against a first-party release, with no funder record located.

Reliability ●●●●● 3/5Why 3/5: a single first-party press release on a low-durability host. No funder record could be located, and nothing has been built yet to evaluate. · Sources: HealthPartners

A published screening framework runs an empathetic conversation first and a scored questionnaire second, with no performance figures anywhere in the readable record.

• Shakoor, Khan and Ghafoor (Bahria University, Islamabad, Department of Software Engineering and Department of Professional Psychology), "An explainable AI framework for early mental health screening and clinical decision support using reflective listening and generative AI", was published online in Discover Psychology on 10 September 2026, open access. It was received 6 March 2026 and accepted 21 August 2026.
• Stage one is a reflective-listeningA counselling technique in which the listener restates what the speaker said, to check understanding and to prompt them to say more. conversational agent that conducts empathetic dialogue, extracts clinically relevant symptoms and performs initial risk screening using contextual reasoning and knowledge-graphA structured map of concepts and the links between them, which a system can reason over rather than only matching text. representations.
• Stage two incorporates the Penn State Worry Questionnaire, the Beck Depression Inventory-II and the McLean Screening Instrument for Borderline Personality Disorder for quantitative, disorder-specific risk estimation, with clinical guidelines informed by DSM-5-TRThe current text revision of the American Psychiatric Association's diagnostic manual, which sets the criteria clinicians use to name a disorder..
• The authors report "encouraging agreement" between the framework's AI-assisted screening outcomes and clinician-reviewed assessments. The study was approved by the Ethics Review Committee of Bahria School of Professional Psychology, informed consent was obtained, and participants identified as potentially high-risk were advised to discontinue and seek immediate professional evaluation.
• The framework "was evaluated solely as an AI-assisted mental health screening and decision-support prototype and was not used for clinical diagnosis or treatment".

Why it matters: it is a peer-reviewed, ethically governed screening framework built in and for a non-Western setting, which is the design gap the Asian implementation review named this same week. It also shows how little of the performance record reaches a reader at accepted-manuscript stage.

Is there a catch? There is no quantitative result available anywhere in the readable record: no accuracy, no agreement statistic, no sample size and no number of clinician reviewers. The fine-tuning corpus is explicitly AI-generated reflective conversations, so the clinician-agreement claim is partly an evaluation of a system trained on synthetic dialogue.

Reliability ●●●●● 3/5Why 3/5: an open-access peer-reviewed article read at the publisher, but the accessible record carries no performance figures at all, so nothing about how well it works can be checked. · Sources: Discover Psychology

A sensor-embedded shirt from a Mexican public institute would classify a wearer's physiological signals into four states, one of them an anxiety-depressive attack.

• Xataka Mexico reported on 6 September 2026 that researchers at the Centro de Investigación en Computación of Mexico's Instituto Politécnico Nacional, led by María Caridad Mireles Pérez with Amadeo José Argüelles Cruz, are developing a sensor-embedded shirt.
• The shirt measures skin electrical activity, electrocardiogram, body temperature, lung activity and muscle tension, with a deep-learning model classifying four states: baseline (calm), transition (nervousness), induction (an anxiety-depressive attack) and recovery.
• The institute has begun proceedings for a first patent on the work, and the device is positioned as a clinician support instrument rather than a commercial product.
• The article cites INEGI BIARE 2026 data that 21.8% of Mexico's population aged 12 and over showed anxiety indicators and 13.4% showed depression indicators over a two-week survey period (anxiety: 26.2% of women versus 16.7% of men; depression: 15.9% of women versus 10.5% of men).

Why it matters: it is one of very few passive-sensing programmes outside the US and Europe this week, built against a nationally measured need, which is the case for developing tools where the affected population actually is.

Is there a catch? This is a prototype plus a patent filing, with no peer-reviewed publication, no reported accuracy, no sample size and no clinical validation of any kind. The accompanying claim of worldwide novelty is contradicted by the existing smart-textile stress-detection literature.

Reliability ●●●●● 3/5Why 3/5: a single technology-news report. There is no paper, no data and no independent evaluation to check the claims against. · Sources: Xataka México