Measure twice, ship anyway

A student shot in the Florida State University attack alleges ChatGPT helped the gunman prepare and never flagged, restricted or escalated anything.

• The docket in GOURLEY v. OPENAI FOUNDATION, No. 4:26-cv-00416 (N.D. Fla.), shows the complaint entered 26 August 2026 before Judge Robert L. Hinkle, naming eleven OpenAI-related entities, filed by Fasig Brooks with a jury demand.
• Reese Gourley, a Pennsylvania resident and FSU student critically injured in the April 17, 2025 campus attack that left two people dead, alleges negligence and sycophancy: that "ChatGPT failed to connect the dots, failed to escalate the situation for human review, ChatGPT failed to suspend or restrict the Shooter's access, and failed to trigger any effective intervention."
• The complaint alleges the chatbot identified uploaded weapons, described how to operate them, discussed lethality and "the number of victims needed for national attention," identified busy times at the FSU Student Union, and answered questions about loading a shotgun and a Glock moments before the attack.
• The complaint further alleges OpenAI released GPT-4o after "compressing months of planned safety testing into a single week in order to beat Google to market by one day in May 2024," quoting an employee: "They planned the launch after-party prior to knowing if it was safe to launch. We basically failed at the process."
• It is at least the fourth suit against OpenAI arising from the FSU shooting.

Why it matters: sycophancy is pleaded here as a design defect and the pre-release safety-testing timeline is itself put in issue, which is a different liability theory from the crisis-response failures pleaded in the earlier suits. (Whether sycophancy is a stable property at all was measured this week.)

Is there a catch? These are allegations in a complaint, not findings, and OpenAI has not answered in court. The complaint PDF was not on the public docket when the record was checked, so the substantive allegations rest on two local outlets quoting the document rather than on the document itself.

Reliability ●●●●● 5/5Why 5/5: the filing is confirmed on the official federal docket, a primary record, with the parties and the entry date read from it; the allegations inside are quoted by two independent local outlets because the complaint PDF was not yet public. · Sources: CourtListener docket · WFSU

What people type into a chatbot instead of telling a therapist is being read into public court records, with no privilege protecting it.

• A Washington Post investigation published Aug. 27 found chatbot logs cited in 12 court cases over the past two years, spanning criminal prosecutions and civil litigation, from a review of public records and local news stories.
• The lead case is a minor's own mental-health suit. A teenage boy who sued Meta, Snapchat, TikTok and YouTube in 2023 over social media addiction had his intimate October 2024 ChatGPT conversations entered into the public record by defense attorneys. His counsel Mike Morgan: "A fifteen-year-old may type things into a chatbot that he may not say to a therapist, a parent, or a friend, and he does it without realizing that it could end up in a defense expert's report."
• A federal judge in New York ruled in February that a financial executive's conversations with Anthropic's Claude about defense strategies, seized under a broader search warrant, were not attorney-client privileged, because "Claude is not a lawyer, and the executive's attorneys did not ask him to talk to the chatbot." The ruling also rejected work-product protectionWork-product protection: a rule shielding material a lawyer prepares in anticipation of litigation from being handed to the other side. The judge held it did not reach the chatbot conversations either..
• The Post reports that OpenAI "disclosed data from more than 80 accounts" to government and law-enforcement requesters in the second half of 2025, "an increase of more than four times as many over the same period the year before," and that an OpenAI spokesperson declined to say how often the company had reported users to police.

Why it matters: a single previously reported incident becomes a counted pattern, and the legal rule that was missing from it arrives with it. Candour offered to a chatbot in place of a therapist is discoverable, and on the first federal holding to reach the question it carries neither attorney-client privilege nor work-product protection.

Is there a catch? Every court event described happened before this week: the privilege ruling was in February, the settlement in the minor's suit in late July, and the only thing that happened in the past seven days is the publication of the investigation. The Post also says the true number of cases is unknowable "because police, and parties in civil cases, don't have to present in court all the evidence they obtain," the count of 12 is its own tally from public records and local news with no case list published, and the Post discloses that it has a content partnership with OpenAI.

Reliability ●●●●● 4/5Why 4/5: an original investigation read in full through a licensed reprint carrying the byline verbatim, with two of its central legal claims separately confirmed in independent sources; not 5, because no official record or docket was opened and the 12-case count is the Post's own unauditable tally. · Sources: Washington Post · Missouri Lawyers Media

On a clinician-validated mental-health test set, the leading AI models cannot be told apart.

• A Beth Israel Deaconess digital-psychiatry team screened HealthBench's 5,000 physician-rubric conversations for mental-health relevance and released 610 of them, 12.2% of the corpus, after two rounds of blinded clinician review with concealed known-exclude controls.
• The team also released the pipeline, the model responses, the grades and the analysis code.
• Across 20 frontier and open models, judged by a cross-vendor panel of three AI judges, the frontier cluster came out statistically tied, two models showed measurable refusal behaviour, and the rankings barely moved between judges (Kendall's tauKendall's tau: a measure of how closely two rankings agree, where the maximum value means identical order. At 0.92 and above, swapping the judge barely changes which model comes out ahead. of at least 0.92).
• HealthBench already appears in AI companies' model cards as the core of their health testing, so a psychiatry-resolved score becomes a target labs are measured against.

Why it matters: the benchmark labs already cite now has a psychiatry version, and the first result off it is that the frontier is not separable. On this test, picking the "best" model buys nothing you can measure.

Is there a catch? It is a preprint, labelled version 1.0.0, and has not been peer reviewed. The scores grade written answers against physician-written rubrics, so nothing here tracks what happened to a patient, and the graders doing the scoring are themselves AI models.

Reliability ●●●●● 4/5Why 4/5: methods, affiliations and every headline figure were read at arXiv and checked against the paper; capped at 4 because an arXiv preprint has not been peer reviewed. · Sources: arXiv · full text

A frontier AI lab is paying outside researchers to build the evaluations that will measure whether its own models harm the people using them.

• On August 25 Anthropic announced $5 million in grants for independent research into how AI affects users' wellbeing, supporting "open-source evaluations that help the AI industry measure how our models affect those who use them."
• It names as concerning the cases where "a user begins to seek companionship from a model, or uses AI to navigate a mental health crisis."
• Grantees get funding, model access and technical support while working fully independently. Applications are due September 21, with selection notices by October 5.
• Accompanying guidance says rigorous wellbeing evaluations should state what they measure, involve clinical experts in design and validation, test both overrefusal and harm, reflect multi-turn real usage, and validate graders against real subject-matter experts.

Why it matters: a frontier lab is capitalising the external measurement ecosystem that will set the standards it is judged by, and the guidance document is itself a normative claim about how crisis and companionship behaviour ought to be measured.

Is there a catch? Nothing has been evaluated yet, so this is the opening of a call rather than a result. The announcement is Anthropic's own with no independent confirmation, and the grantees' independence rests on terms the funder wrote.

Reliability ●●●●● 4/5Why 4/5: a first-party announcement of a real program with a verifiable artifact, an open application form and a published guidance document; no independent confirmation and no outcome data yet. · Sources: Anthropic · Social Samosa

Across 27 systematic reviews, the evidence supports AI for screening and prediction much better than for sustained treatment.

• An umbrella review in the Journal of Psychiatric and Mental Health Nursing, published online 24 August, searched five databases for peer-reviewed systematic reviews published 2021 to 2025, screened in triplicate, and included 27 reviews covering over 14 million participants, appraised with AMSTAR-2.
• Four themes converged. AI improves early detection and risk stratification, particularly for depression, anxiety, stress, PTSD and suicidal ideation, and chatbots and mobile platforms expand access and engagement. Structural and ethical fragilities persist, including data bias, limited external validity, methodological heterogeneity and insufficient transparency, and AI showed most value embedded in human-centred hybrid care that preserves clinical judgement and accountability.
• The authors state that "Evidence supporting AI applications was stronger for screening and predictive tasks than for sustained therapeutic interventions."
• A separate review published the same week found the same split in the refugee mental-health literature.

Why it matters: the highest-level synthesis available this week says the evidence base backs detection far better than it backs treatment, which is the reverse of where the consumer market has concentrated.

Is there a catch? The synthesis is thematic rather than a pooled quantitative meta-analysis, and the participant total is an aggregation across reviews that certainly overlap, so 14 million is not an independent sample. The abstract is also internally inconsistent about how many reviews were appraised for quality.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed umbrella review whose counts, design and date were confirmed at Europe PMC and PubMed after the publisher's own page refused to load; below 5 because the synthesis is thematic rather than a pooled meta-analysis. · Sources: Journal of Psychiatric and Mental Health Nursing · PubMed

The number NHS England used to justify putting AI triage into the NHS App fell short of official standards, because the data behind it was never published.

• GB News reported on 28 August that the Office for Statistics RegulationOffice for Statistics Regulation: the UK body that sets and polices standards for how public bodies publish and use statistics. It can criticise a misuse publicly, but it cannot compel a correction. found NHS England's public use of a 29% figure did not meet official standards because the underlying data was not published.
• NHS England's July 2026 announcement had said "An initial trial at a GP practice in Sussex resulted in a 29% reduction in the number of people queuing on the phone."
• In an email seen by GB News, OSR head Ed Humpherson wrote "We have found that NHS England's statement about the GP practice trial did not meet our expectations," and the regulator concluded "The use of the 29 per cent figure in this NHS England statement does not align with these expectations."
• Oxford academics Carl Heneghan and Tom Jefferson, who complained, established that the preprint linked to the announcement contained no analysis of phone queues, waiting times or call volumes, and that NHS England later said the 29% came from a separate unpublished internal evaluation.
• NHS England has told the regulator it intends to publish the underlying data and is taking steps to prevent recurrence. The tool is due to reach more than 200,000 patients within a year and every NHS App user by April 2028.

Why it matters: a national front door is being restructured on an evidence claim a regulator has now faulted. The same argument is running in the US through the bill on California's desk that would bar clinical decisions resting solely on decision-support output, and through the union fight over algorithmic mental-health triage at Kaiser Permanente, whose underlying investigation predates this week.

Is there a catch? Two limits travel with this. The OSR has published nothing, so the finding exists only in an email seen by one outlet rather than as a published regulatory ruling. And the NHS App tool is general-practice triage, not a mental-health tool, though mental-health presentations pass through it. The finding is about the transparency of a statistic, not the tool's clinical accuracy or safety.

Reliability ●●●●● 3/5Why 3/5: a real event with the regulator quoted directly and the complainants' own published account to corroborate it, but there is no official document to open: searches of the regulator's correspondence index and health pages turn up nothing, and the finding reaches the record only through an email seen by one outlet. · Sources: GB News · Trust the Evidence

A new Japanese service reads students' free-text survey answers on their school tablets and tells teachers which children need urgent support.

• SoftBank Corp.'s own release of 27 August 2026 announces Mentari, available from that date, doing three things: running school surveys from creation through analysis; analysing responses, including free-form written answers, to identify students needing urgent support; and running a chatbot character, "Mofu," designed for attentive listening and content organisation that deliberately does not give advice.
• Students respond on the one-tablet-per-student devices distributed under Japan's GIGA School Initiative.
• Teachers receive the AI analysis together with expert-reviewed guidance.
• The service is free for a set period from the start of use, for the purpose of effectiveness verification. Neither the eventual price nor the length of that period is stated.

Why it matters: urgency triage of minors' free text is now running on a national one-device-per-student base, with a teacher as the person who receives the flag and has to act on it. A deployment of this design in a US school would now need prior written parental consent rather than an opt-out.

Is there a catch? The release reports no accuracy, sensitivity, specificity or false-positive figure of any kind, and no trial or evaluation data, for a system that sorts minors' free text into urgency flags. The free period is itself framed as effectiveness verification, so the evidence is meant to arrive after the deployment.

Reliability ●●●●● 4/5Why 4/5: a first-party launch announcement for a dated, shipped service, read at SoftBank's own newsroom; held below 5 because SoftBank publishes no performance data at all for a system that triages minors' free text. · Sources: SoftBank · Impress Watch

Blinded clinicians scored AI-generated therapist replies roughly level with human ones, and rated the AI replies more empathetic.

• In JMIR Mental Health, published 26 August 2026, an Aarhus University and Aarhus University Hospital team had clinicians blind-rate 5 fictitious patient messages, each paired with one human and one AI-generated therapist response, inside a simulated internet-delivered CBT treatment for functional somatic disorder. 61 raters provided data and 54 were analysed.
• On a 1 to 5 scale the AI and human responses landed close together: overall quality 4.00 against 3.96, helpfulness 3.85 against 3.93, professionalism 4.25 against 4.11, protocol adherence 4.13 against 4.13, empathy 4.31 against 4.08.
• Raters identified the source of human responses correctly more often (mean 79%) than of AI responses (63%).
• Of the 30 raters who wrote open comments, the qualitative analysis found the AI responses were seen as polished but also generic and at times excessively empathetic.

Why it matters: internet-delivered CBT is rate-limited by the therapist time spent writing replies, and this measures whether that step comes apart from the rest of the treatment. The largest gap, empathy, ran in the model's favour.

Is there a catch? This is a blinded rating pilot on fictitious material with clinicians as raters, so it supports claims about perceived response quality only and establishes nothing about patient outcomes. No patient read any of these replies.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed preregistered study, confirmed at Europe PMC and PubMed with every rating figure after the journal's own page refused to load; below 5 because it is a rating pilot on fictitious material rather than a trial. · Sources: JMIR Mental Health · PubMed

Expressing emotion made people trust an AI depression-support agent more and worry less about privacy. Giving it a humanlike appearance changed nothing.

• A 2 × 2 experiment with 319 participants, published online 24 August 2026 in Information Systems Frontiers, separated appearance from emotional expression in how an AI "doctor" was presented.
• Emotional expression increased trust and reduced privacy concerns. Appearance had no significant effect.
• Trust raised intention to adopt and privacy concern lowered it, and the two design types interacted negatively, so each weakened when the other was high.
• The effect of emotional expression was stronger among participants who perceived high public stigma around depression.

Why it matters: it isolates which humanlike cue actually moves users, the language rather than the face, and finds that warmth buys a drop in privacy concern without any change in actual confidentiality.

Is there a catch? The outcomes measured are self-reported trust, privacy concern and intention to adopt, not symptom reduction, clinical benefit or crisis safety. The article does not state where the 319 participants were recruited.

Reliability ●●●●● 4/5Why 4/5: peer-reviewed, with the design, the sample and the direction of every reported effect confirmed on the publisher's page; below 5 because the full statistics sit behind a paywall and the study measures attitudes rather than outcomes. · Sources: Information Systems Frontiers · Scienmag

How sycophantic a model looks depends on how the user phrases the question, so a single-prompt sycophancy number does not describe the model.

• Authors at Northeastern University, Case Western Reserve University and Everpure built controlled prompt variants that hold the user's underlying situation constant while varying social cues: user confidence, hedging, emotional framing, social consensus, validation-seeking language and counter-framing.
• Each item was evaluated under eight prompt variants across nine open-weight models and three datasets derived from the earlier ELEPHANT sycophancy work.
• They introduce a per-item Sycophancy Prompt Sensitivity Score that separates a model's baseline sycophancy from the shift the phrasing induces.
• Validation-seeking and emotional-pressure phrasing raised sycophancy while counter-framing lowered it. Some models show moderate baseline sycophancy but high sensitivity, so a single-phrasing rate misses their vulnerability.

Why it matters: the cues shown to inflate sycophancy are exactly the ones a distressed person brings, which makes every single-prompt mental-health sycophancy safety number an unstable measurement. Sycophancy is meanwhile being pleaded in court as a design defect.

Is there a catch? The abstract reports only directional results with no effect magnitudes, so the size of the shift is unknown. The models tested are open-weight, so this does not directly measure the closed frontier models most people use, and no distressed person was involved: the emotional pressure is written into prompts.

Reliability ●●●●● 3/5Why 3/5: a readable preprint with a documented design and bootstrap uncertainty, but small, not peer reviewed, and reporting only directional results with no effect magnitudes in the abstract. · Sources: arXiv

The population most often invoked to justify AI mental health on access grounds has four intervention studies behind it, worldwide.

• A Yonsei University College of Nursing team searched eleven databases and preprint servers from inception through June 2026, plus the first 100 Google Scholar results, appraising quality mainly with the Mixed Methods Appraisal Tool, publishing in the Journal of Medical Internet Research on 26 August 2026.
• Of 2,726 studies identified, 23 were included, splitting into Detection, Prediction and Classification (15 of 23, 65.2%), Intervention and Clinical Support (4 of 23, 17.4%) and Information Synthesis (5 of 23, 21.7%).
• Reported strengths were scalability, accessibility in resource-limited settings, early symptom detection, reduced language barriers and synthesis of unstructured data.
• Reported limitations were bias, lack of contextual understanding, structural barriers to technology access and overreliance concerns.
• The authors state this is the first review specifically examining AI for refugee mental health care.

Why it matters: access to the underserved is this field's main moral argument, and this is the literature sitting under it: overwhelmingly about finding people, barely at all about helping them.

Is there a catch? It is a review of a very thin literature with no pooled effect estimates, so it describes what has been studied rather than what works. Counting studies also says nothing about how large or how good any of them were.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed systematic review whose counts and category split were confirmed at Europe PMC and PubMed after the journal's own page refused to load; below 5 because the underlying literature is thin and no pooled effect estimates exist. · Sources: Journal of Medical Internet Research · PubMed

AI used alongside psychotherapy should be bounded in the problem it addresses, the procedures it uses, and how much of it a patient gets.

• Published online 27 August 2026 in the Journal of Consulting and Clinical Psychology by authors at Harvard's Department of Psychology and the company Everwell.
• The proposal is that clinically deployed mental-health AI should be FIT: focal (bounded in problem focus), intervention-specific (bounded in therapeutic procedures) and titrated (dose and duration calibrated to the use case).
• The authors offer these as "practical starting guardrails, not as an exhaustive list," designed to reduce three risks linked to conversational mental-health AI.
• They state their argument concerns only AI used in routine clinical care and tools prescribed by or used with clinician guidance, and that they "leave aside fully unsupervised consumer use."

Why it matters: a leading clinical-psychology journal is drawing the supervised and unsupervised boundary explicitly, and writing design constraints for only one side of it. The other side is where most people actually meet this technology.

Is there a catch? It is explicitly a viewpoint article with no data, no sample and no effect estimates, so it is a proposal rather than evidence of anything. The second author affiliation is a company rather than a university, which is worth knowing about a piece proposing design guardrails.

Reliability ●●●●● 4/5Why 4/5: a commentary in a peer-reviewed journal, confirmed through PubMed and Crossref after the publisher's page proved unfetchable; the score grades the venue, not the evidence, and the piece carries no data at all. · Sources: Journal of Consulting and Clinical Psychology · PubMed

Facing a caregiver in an acute crisis, four widely used models were indistinguishable from one another, and all talked too much and solved too early.

• An Imperial College London team, with a co-author at Universidad Autónoma de Madrid, had four widely used models advise synthetic help-seekers with psychometrically specified profiles in one held-constant acute crisis: a caregiver learning of a relative's dementia diagnosis.
• The profiles covered the Big Five personality traits plus coping style, coping self-efficacy, resilience and reactance. Auditors blind to the profile prompt recovered the specified bands from the dialogue alone with high agreement on every instrument (ICCIntraclass correlation: a measure of how far independent raters agree, where higher means closer agreement. At 0.91 the blinded auditors were reading out of the dialogue the same personality profile the researchers had written into it.(2,4) = 0.91, range 0.79 to 0.96; band-score correlation 0.78).
• The four models were not distinguishable on emotion stabilisation and failed alike, sharing verbosity, a talk-to-listen ratio above one, and problem-solving before the situation had been explored.
• The design is a multi-turn evaluation with controlled personas rather than a single-turn refusal benchmark, which puts the failure in conversational behaviour rather than in a guardrail that either fires or does not.

Why it matters: the models are failing in the same direction rather than differing by vendor, and the failure is a manner of talking, which no safety filter is looking for.

Is there a catch? The help-seekers are synthetic, the paper is not peer reviewed, and no clinical outcomes are reported. One crisis scenario, deliberately held constant, is also a narrow basis for a claim about crisis conversations in general.

Reliability ●●●●● 4/5Why 4/5: methods are readable at arXiv, including the blind auditing and the reported reliability statistics, and every figure was checked against the paper; not peer reviewed, and the help-seekers are synthetic. · Sources: arXiv

Korean adults with subclinical to moderate depressive symptoms improved more with a messenger-app companion agent than a group carrying on as usual.

• A Yonsei University and HAII Corp team ran a proof-of-concept quasi-experimental study with 58 South Korean adults aged 20 to 39, published online 26 August 2026 in DIGITAL HEALTH: 29 in the intervention arm and 29 sequentially recruited passive controls, with 57 completing.
• Over six weeks BDI-II depression scores fell from 18.18 to 9.50 in the intervention group against 17.86 to 13.76 in controls (condition × time F(1,55) = 11.69, p = .001).
• Quality of life rose from 41.57 to 50.18 against 38.55 to 43.14 (F(1,53.53) = 5.22, p = .026). Anxiety showed no significant condition × time interaction (p = .226).
• On-time adherence to the daily check-ins was 69.5% overall, falling from 84.2% in week 1 to 59.2% in week 6. Behavioural-activation routine completion was 66.2%, and the System Usability Scale score was 74.02.

Why it matters: one of the few items this week that reports actual symptom change from a deployed conversational agent, and it ran inside an ordinary messenger app rather than a bespoke product a patient has to be sold on.

Is there a catch? The control condition was passive: participants carried on with their usual routines, with no access to the intervention and no attention-matched comparator, which inflates the apparent between-group effect. The sample is 58 adults in one country, and adherence fell steadily across the six weeks.

Reliability ●●●●● 4/5Why 4/5: peer-reviewed, with the design, sample, control condition and every figure confirmed at the publisher and in the PubMed record; capped at 4 because it is a small quasi-experimental study against a passive control. · Sources: DIGITAL HEALTH · PubMed

The FDA intends to write formal regulatory guidance for generative-AI medical devices: one broad document plus narrower ones for the harder topics.

• STAT News reported on August 24, 2026 that the FDA is developing a plan for regulating medical devices that use generative AI.
• Rick Abramson, director of the Digital Health Center of Excellence, said "Our goal is formal policy guidance."
• He said "the ecosystem can expect not only broad guidance on the overall topic of generative AI, but also some more narrowly constructed specialty guidance on particular generative AI topics of special interest or special complexity."
• The remarks follow the FDA's August 18, 2026 generative-AI discussion paper.

Why it matters: the FDA moves from a discussion paper to a stated intention to write binding guidance, and the official who owns the pen is the one saying it. Whether generative-AI mental-health tools count as devices at all runs through this document.

Is there a catch? Abramson gave no release date or timeline, named no chatbot and made no mention of mental health. This is an interview statement in trade press rather than an agency document, and an intention to write guidance is not guidance.

Reliability ●●●●● 3/5Why 3/5: the date, the speaker and the two-tier commitment were confirmed in the article, but this is an interview statement in trade press rather than a published agency document, so it cannot carry an agency grade. · Sources: STAT News · Conversational AI Watch

A startup selling health plans AI that reads unstructured behavioural-health clinical notes has raised $17 million, with strategic money from CVS Health Ventures.

• Onos announced on August 26, 2026 a $17 million Series A led by Costanoa with Flare Capital Partners participating and a strategic investment from CVS Health Ventures.
• The round follows a $6.3 million seed in October 2025, bringing total disclosed funding to $23.5 million. No valuation was disclosed.
• The platform reads claims, utilization and clinical documentation against quality guidelines to give payers a population-level view. Onos claims "more than 70% of behavioral health care quality signals are embedded in unstructured documentation," says it is "trusted by 3 of the 6 largest health plans in the country," and names Aetna as a customer.
• Onos also claims a 35% improvement in adherence to clinical standards, a 75% improvement in clinical review efficiency and a greater than 6% reduction in behavioural-health program costs within 12 months. Behavioral Health Business, which had the release in hand, reported the round without repeating that cost figure.

Why it matters: a payer's own venture arm is capitalising a vendor whose named customer is that payer's subsidiary, funding wider payer adoption of AI that reads therapists' notes to steer utilization decisions. California's sole-reliance clause is aimed squarely at this layer.

Is there a catch? Every performance figure is an unaudited company claim about its own product, with no named evaluator, method, comparison group or patient-level endpoint. The Aetna relationship rests on Onos's word alone: no filing, and neither Aetna nor CVS Health has confirmed it.

Reliability ●●●●● 3/5Why 3/5: the funding event is real and independently confirmed by an outlet that did its own interviews, but every performance number is a first-party assertion about the company's own product and there is no shipped artifact to inspect. · Sources: PR Newswire · MedCity News

A three-year South Carolina project will try to predict youth suicide risk by fusing hospital records with state mental-health data and school attendance.

• MUSC's newsroom announcement of 27 August 2026 confirms $1.23 million from The Duke Endowment for a three-year initiative that "uses AI to develop predictive models capable of detecting complex patterns of suicide risk that traditional methods may miss," led by Jihad Obeid, M.D.
• The models are to integrate electronic health records from multiple medical organizations, South Carolina Office of Mental Health service-utilisation and outcomes data, and PowerSchool attendance and demographic data.
• Named inputs include history of depression, substance use, trauma exposure and social instability, alongside suicidal ideation, mental health referrals, absenteeism and sleep disorders.
• Year 1 covers partnerships, regulatory approvals, data-sharing agreements and secure infrastructure; Year 2 student enrolment, data integration and model development; Year 3 evaluation and statewide dissemination. The release states models will be developed without using direct identifiers.

Why it matters: the novel part is the linkage rather than the algorithm: a risk model aimed at minors and fed by school records. A federal consent letter issued the day before the announcement gates exactly the school data that the first year is meant to secure.

Is there a catch? This is a grant award, not a result. No model exists yet, the school-record linkage is aspirational, and no accuracy, sensitivity or specificity figure is reported. The announcement is the university's own.

Reliability ●●●●● 3/5Why 3/5: the amount, the funder, the three-year term and the AI framing are all confirmed on MUSC's own first-party page, but the only artifact is a funding announcement: no model, no data, no validation. · Sources: MUSC · Live 5 News

Asked to read months of a person's data and predict their wellbeing, AI agents rarely beat guessing the average.

• A Dartmouth-led team, with a co-author at the University of Cambridge, released what it describes as the first systematic benchmark of LLM-based agentic systems for longitudinal mental-health sensing.
• The benchmark spans 3 real-world longitudinal datasets, 2 task families (closed-form wellbeing-score prediction and rationale generation judged by a model) and 3 agentic paradigms across 5 open- and closed-source backbones.
• Zero-shot agents rarely outperform a simple mean baseline, except with stronger backbones or compact, semantically meaningful features.
• Chain-of-thought prompting improves reasoning-oriented backbones without guaranteeing temporal grounding or numerical correctness.

Why it matters: passive-sensing products are sold on the premise that agentic models can read longitudinal signal out of a person's data, and the first benchmark of that premise says they mostly cannot.

Is there a catch? The headline negative result is stated qualitatively with no effect sizes in the abstract, and the paper is not peer reviewed. A benchmark also tests the setups its authors chose, so a different agent design could do better.

Reliability ●●●●● 4/5Why 4/5: the benchmark design and the authors' own framing were read at arXiv and check out against the paper; not peer reviewed, and the headline negative result carries no effect sizes. · Sources: arXiv

Psychiatrists argue that flattering, humanlike chatbots can amplify and co-construct unusual beliefs, and disagree about whether that deserves its own label.

• A preprint dated 25 August 2026 by authors at King's College London, Western Eye Hospital and University College London sets out the proposed mechanism: model sycophancy, "a tendency to agree with and flatter users that is reinforced through preference-based fine-tuning," combining with increasingly anthropomorphic design to create a bidirectional "echo chamber of one."
• The case for recognising it: case identification, tailored interventions, standardised research criteria, post-market surveillance, and pressure on developers and regulators.
• The case against: reifying a syndrome from anecdotal evidence, existing constructs possibly already accommodating AI use as a precipitating factor, an unproven causal claim inside the term itself, stigma, and a psychosis-centric label obscuring a broader spectrum of harms.
• The authors recommend taking a "technological history" in psychiatric assessment, pre-deployment benchmarking for sycophancy and delusion reinforcement, and post-deployment surveillance.
• They cite OpenAI's 2025 estimates that in a given week roughly 0.07% of its reported 800 million weekly active users, about 560,000 people, showed possible signs of psychosis or mania, and 0.15%, about 1.2 million, had conversations with explicit indicators of potential suicidal planning or intent.

Why it matters: whether these harms get counted at all, and by whom, turns on the diagnostic-category question. The pre-deployment sycophancy benchmarking the authors ask for was shown this week to move with the phrasing.

Is there a catch? The authors state explicitly that this is a perspective article with no new data, and that current evidence on AI-associated psychosis is limited to media reports, individual case reports and early observational data. The prevalence figures they cite are OpenAI's own, self-reported and never externally validated.

Reliability ●●●●● 3/5Why 3/5: a real, readable primary artifact whose argument, date and affiliations were all confirmed in the paper itself, but a perspective piece with no new data, so it cannot take the score reserved for a preprint with methods. · Sources: arXiv

Addiction-medicine clinicians said they would be comfortable with clinical AI if its development were transparent and it would not reinforce stigma about drug use.

• Published online in Addiction on 24 August 2026, the study ran one-on-one semi-structured interviews in 2025 with a purposive sample of 12 licensed addiction-medicine clinicians (Rhode Island 7, Massachusetts 2, Pennsylvania 3) and 25 people who use drugs recruited from four Rhode Island social service organizations.
• Five themes were generated, including that "clinician comfort with clinical AI hinges on transparent development and assurances against the perpetuation of drug-related stigma."
• Another was that "clinical AI may not keep pace with the volatile and regionally specific drug supply."
• The authors conclude that clinicians and patients see promise but are concerned about clinical relevance, stigma perpetuation and the privacy of potentially incriminating data.

Why it matters: these are adoption conditions in the clinicians' own words rather than an attitude score: tell us how the model was built, and guarantee it will not encode stigma. Both are commitments a vendor can make or refuse. (The worry about incriminating records got a concrete illustration in court this week.)

Is there a catch? It is a small purposive qualitative interview study with 12 clinicians, no quantitative outcome measure and no comparison group, so the findings are perspectives rather than measured effects, and they do not generalise beyond these three states.

Reliability ●●●●● 4/5Why 4/5: a peer-reviewed qualitative study in Addiction, with every quoted theme verbatim in the abstract and the design and date confirmed at Crossref; not 5, which is reserved for trials, systematic reviews and primary official records. · Sources: Addiction

A model trained only on patient-education material scored highest on every rubric dimension, and most psychiatrists still preferred ChatGPT's answers.

• Authors at Lurie Children's Hospital of Chicago, Relativ.ai and Northwestern Feinberg compared MIND, ChatGPT and OpenEvidence on patient questions about escitalopram, using both a computed rubric (accuracy, clarity, completeness, nuance, safety, referral appropriateness) and ratings from board-licensed psychiatrists.
• By rubric, MIND was rated highest in all domains (p < 0.001).
• By psychiatrists, ChatGPT was rated accurate more often with a negligible effect size (p = 0.021, r = 0.073), MIND more complete with a small effect size (p < 0.001, r = 0.160), and the two were rated safe with the same frequency (p = 0.955, r = 0.002).
• A majority of psychiatrists preferred ChatGPT's responses, 57.6% against 42.4% (p = 0.003).

Why it matters: a rare head-to-head between a purpose-built mental-health model and a general frontier chatbot, in which the two ways of scoring it disagree about who won. Which one a buyer trusts decides what gets bought.

Is there a catch? The preference split rests on a rater panel of only 10 psychiatrists, the work is not peer reviewed, and one author is affiliated with the vendor whose model is the subject. Every question was about a single antidepressant.

Reliability ●●●●● 4/5Why 4/5: statistics and effect sizes are reported and were checked against the paper at arXiv; not peer reviewed, the rater panel is ten psychiatrists, and one author works for the vendor whose model is being tested. · Sources: arXiv

A college has ended its two-year contract with an AI mental health app, saying students engaged more consistently with campus-based services.

• The Ithacan reported on 27 August 2026, the first public disclosure, that Ithaca College terminated its Wysa contract on 5 August 2026, ahead of the 2026-27 academic year and after a two-year contract begun in Fall 2024. The premium tool packs (stress management, grief, anger management, trauma) end; the free chatbot remains.
• The Center for Health Promotion's 2025-26 annual report records 497 Wysa enrollments across students, employees and dependents, and gives the reason as "ongoing financial constraints at both the departmental and institutional levels."
• The director of student support and wellness said the college "saw more consistent engagement with other campus-based mental health services and supports," and that the evaluation did not include testimony from an April 2026 student town hall criticising the app.
• Common Sense Media's head of AI assessments said its risk assessment, which covered only Wysa's free consumer chatbot and was built with psychiatrists across 13 mental health topics, rated Wysa at an unacceptable risk level, and that direct-to-consumer chatbots "were not better performing than the multi-use chatbots," including on recognising eating disorders and responding to crises.

Why it matters: demand-side discipline, with budget pressure and an external safety rating both on the record. Money is consolidating at the top of this market and contracts are lapsing at the bottom of it. (The consolidation side.)

Is there a catch? This is a single outlet's original reporting, and the Common Sense Media findings are quoted from that report rather than read from the assessment itself. One college's budget decision is also not evidence about whether the app works.

Reliability ●●●●● 3/5Why 3/5: original reporting by the institution's own newspaper, with the responsible administrator confirming on the record by email and the annual report cited, but a single outlet with no independent second source. · Sources: The Ithacan

A national counselor certification body argues that a pending federal bill barring chatbots from claiming a licence protects counselors, and creates a new duty for them.

• On 26 August 2026 NBCC's Government Affairs division published an analysis of 2026 federal AI legislative and executive activity, writing that "The CHATBOT Act's prohibition on chatbots claiming licensure or professional credentials protects Mental Health Counselor professional standing."
• It adds that "Chatbots cannot misrepresent themselves as possessing NCC credentials, state licensure, or clinical expertise," and that "Counselors will increasingly face requirements to evaluate AI tools' compliance before clinical adoption."
• The bill it relies on, H.R. 7985, would bar a covered entity from AI chatbot output or marketing that, from a reasonable user's perspective, indicates or implies possession of a licence to practise a covered profession, or falsely implies the output is provided or verified by a licensed human.
• Enforcement would sit with the FTC, with state attorneys general able to sue for actual damages or up to $5,000 per violation.

Why it matters: a certifying body is converting a pending bill into a professional-boundary claim and a new counselor duty to vet AI tools, which is the workforce half of the licensure question the states are already legislating. Missouri's version is already in force.

Is there a catch? H.R. 7985 was introduced in March and referred to committee with no further action: it is a pending bill, not law. A separate bill, S. 4407, carries the same CHATBOT Act short title with entirely different content, so the two must not be conflated. This is also advocacy by an interested party rather than a neutral analysis.

Reliability ●●●●● 3/5Why 3/5: the guild's own dated page was read directly and the underlying bill text checked independently on the congressional record, so the characterisation holds up on substance; but this is first-party advocacy analysis rather than a binding standard or peer-reviewed work. · Sources: National Board for Certified Counselors

Yale will track how many eligible psychiatry patients actually get prescribed a cleared prescription depression app, and how much of it they use.

• NCT07792252 was first posted on 2026-08-28: a prospective observational cohort, not yet recruiting, of 50 participants aged 22 and over at Yale New Haven and Yale Greenwich interventional psychiatry services, with an estimated start of September 2026 and primary completion March 2027.
• Yale University is lead sponsor. Otsuka Pharmaceutical, the company behind Rejoyn, is a listed collaborator.
• Primary outcomes are all feasibility and engagement measures: how many eligible patients are prescribed it, mean app engagements, mean minutes of use, mean lessons completed, and adherence defined as completing at least 12 of 18 sessions by week 12.
• Rejoyn is a 510(k)-cleared prescription digital therapeutic for major depressive disorder, and its own marketing copy calls it "the first and only prescription app for the treatment of major depressive disorder symptoms."

Why it matters: an academic centre is generating real-world data on a cleared prescription digital therapeutic in a setting where it sits beside interventional psychiatry treatments, which is where the engagement question actually bites.

Is there a catch? Neither the registry record nor the product site establishes any AI or machine-learning component in Rejoyn, which is emotional-faces memory-task training plus CBT content, so it belongs here as a prescription digital therapeutic and must not be described as an AI therapy. This is also a registration of an intention: the status is not-yet-recruiting, no results exist, and none of the protocol numbers may be read as findings.

Reliability ●●●●● 4/5Why 4/5: confirmed through the registry's own data record after the web page returned nothing to either fetcher; capped at 4 because a registration is an official but sponsor-submitted record of an intention, with no results attached. · Sources: ClinicalTrials.gov

Short, minimal counselor responses are common in real counseling transcripts and rare in AI-generated ones, and models cannot judge when brevity fits.

• Zhiyang Qi of the University of Tokyo ran a cross-lingual analysis of minimal responses, meaning backchannel cues and concise empathic statements, across multiple counseling dialogue datasets, using length-and-content filtering plus contextual verification by a model.
• Minimal responses are common in human-collected datasets and substantially underrepresented in the AI-generated ones.
• Strong commercial models can produce them when explicitly instructed but struggle to judge when they are appropriate. Counseling-specific models trained on synthetic data perform particularly poorly, producing longer, more information-rich replies.
• The author further notes that response-quality evaluation done by models may itself undervalue minimal responses.

Why it matters: it puts the verbosity problem upstream, in training-data provenance and reward design. Synthetic counseling corpora appear to encode the bias, and the judge models then reward it. The same talk-to-listen problem turned up independently this week in a crisis evaluation.

Is there a catch? The paper reports its findings qualitatively with no rates or effect sizes in the abstract, and the datasets are not enumerated there, so the size of the gap is unknown. It is a single-author preprint that has not been peer reviewed.

Reliability ●●●●● 3/5Why 3/5: the methods are readable at arXiv and every clause of the finding is supported there, but the paper reports qualitatively with no rates or effect sizes and does not name its datasets in the abstract. · Sources: arXiv

A registered randomized trial will test whether a chatbot-enhanced depression-literacy course changes high schoolers' willingness to seek professional help.

• NCT07790731 was first posted on 2026-08-27: an interventional randomized parallel trial, single-masked and prevention-purpose, with an estimated 720 high-school students aged 15 to 19 at National Yang Ming Chiao Tung University, an estimated start of August 2026 and primary completion May 2027.
• The experimental arm is a 4-week AI chatbot-enhanced depression-literacy programme of four weekly 50-minute sessions on causes of depression, symptom recognition and stigma reduction, plus a printed "Resource Navigation Booklet" of local help-seeking channels.
• The control arm receives no intervention and continues the standard curriculum.
• Primary outcomes are change in stigmatising attitudes, depression literacy and attitudes toward seeking professional psychological help.

Why it matters: one of the few registered adolescent chatbot trials aiming at the help-seeking pathway rather than at symptoms, which is the route by which a tool could help without treating anyone itself.

Is there a catch? The control arm is no-intervention rather than an active comparator, and the outcome set contains no symptom measure, so a future readout can speak only to literacy, stigma and help-seeking attitudes. This is a registration: no results exist, and the 720 figure is an estimated enrolment rather than a studied sample.

Reliability ●●●●● 4/5Why 4/5: confirmed through the registry's own data record; capped at 4 because a registration is an official but sponsor-submitted record of an intention, and here the sponsor is an individual investigator rather than an institution. · Sources: ClinicalTrials.gov

A Nottingham study of an app that reads facial expression, speech and movement is flagging roughly one in three participants above its depression threshold.

• The BBC reported on 24 August 2026 that University of Nottingham and Nottinghamshire Healthcare NHS Foundation Trust researchers, with the company Blueskeye, are running a study of TrueBlue, an app that "works by asking users to complete tasks while software analyses their facial expressions, speech and automatic movements" and "flags if patients are getting above a certain depression score." Tasks range from reading a novel to journalling mood.
• The registration (NCT06364488, sponsor BlueSkeye AI) gives a start of 1 October 2024, an estimated enrolment of 125 perinatal participants recruited between 12 weeks of pregnancy and 12 weeks post-partum, and a single-group non-randomised design.
• Endpoints are safety, feasibility, acceptability, usability and agreement of the automated mood scores with PHQ-9, EPDS, GAD-7 and TEPS. The study is heading for October completion.
• The lead researcher says the technology will not be used to diagnose but to support clinicians. A charity representative told the BBC that "women need to get support from an actual human" and that "we don't want women being supported by an AI."

Why it matters: a flag rate that size, if it holds, is a referral-capacity problem before it is a detection problem. Somebody has to see everyone the app flags.

Is there a catch? The "about a third" figure is a lead researcher's spoken statement with no denominator, no confidence interval, no comparison against the validated scales and no published interim analysis. The design is single-group and non-randomised, so it cannot show that flagging anyone changed an outcome.

Reliability ●●●●● 3/5Why 3/5: a single reputable outlet reporting a real, properly registered trial whose enrolment, design and dates were independently checked; the new content is an investigator's spoken interim figure rather than a published result. · Sources: BBC · Resultsense

Nearly two in five physicians say proficiency with AI tools already factors into who gets hired and who gets promoted.

• Doximity published its 2026 Physician Compensation Report on 25 August 2026, drawing on roughly 250,000 compensation survey responses collected over seven years, including nearly 23,000 US physicians surveyed in 2025.
• The headline finding is quoted directly: "39% said AI proficiency already factors into hiring and promotion decisions."
• 66% use AI daily or weekly for clinical or administrative work, and 78% say AI has increased or maintained their job security.
• 67% believe physicians who stay current with AI tools will have a meaningful earnings advantage over the next 12 months, while 72% say AI is not changing the physician's role in their specialty.

Why it matters: this is the only measurement this week of what AI is doing to clinician hiring and pay, and it says the labour market has begun pricing the skill while most doctors still say their own role is unchanged.

Is there a catch? These are all-specialty figures for US physicians with no psychiatry or mental-health breakout, and no non-physician mental-health clinicians (psychologists, counselors, social workers, therapists) were surveyed at all, so applying any of it to the mental-health workforce is an inference the source does not support. The survey is also run by an interested party on its own members.

Reliability ●●●●● 3/5Why 3/5: a first-party survey report with real data and a large panel, read at Doximity's own release with every headline figure confirmed; self-reported, run by an interested party on its own members, and not independently audited. · Sources: Doximity · Healthcare Dive

Asked what they wanted from AI in disaster recovery, Los Angeles wildfire survivors picked forms and resources over emotional support.

• On 26 August 2026 The After Collective, a fiscally sponsored project of a registered 501(c)(3) co-founded by a Stanford clinical research psychologist, opened a closed beta of a free trauma-informed AI coach reachable by text message with no app, account or cost, with broader availability planned for late 2026.
• The same day it published a mixed-methods pilot evaluation of 115 survivors of the 2025 Eaton and Palisades wildfires (mean age 49.31, 73% female), 16 of whom completed interviews and gave feedback on a prototype.
• At 12 to 16 months post-fire, 67% were still displaced in temporary housing. The top recovery challenges were emotional and mental health stress (73%), insurance and rebuilding paperwork (61%) and financial strain (52%), and 30% rated their mental health as poor or very poor.
• Barriers to support were lack of time (50%), cost (49%) and difficulty finding a trusted provider (30%), and 22% wanted mental health support but had not received it.
• 69% were interested in AI tools for recovery, but 81% raised accuracy concerns, 64% data-privacy concerns and 49% lack of human connection. 68% said a human in the loop alongside an AI tool was moderately or very important.

Why it matters: a rare case where a population's own stated preferences are measured before the tool reaches it, and what they ask for is help with forms and resources rather than with feelings.

Is there a catch? The deployed coach itself has not been evaluated: the survivor study is formative research plus prototype feedback from 16 people, and the authors say larger controlled studies are needed to test effectiveness. The organisation publishing the survey is the one building the tool.

Reliability ●●●●● 3/5Why 3/5: the full report was downloaded and read, and it carries a named methods section, participant demographics and a stated limitations section; but it is a nonprofit white paper written by the tool's own builders, not peer reviewed and not a controlled evaluation. · Sources: announcement · The After Collective research

A Japanese ambient-documentation AI built for psychiatry says its drafts arrive 18 seconds after a consultation ends and that nearly all of them are adopted.

• On 2026-08-28 MENTRA Inc. announced a limited-time free trial of a medical AI specialised for psychiatry and psychosomatic medicine, open to psychiatric hospitals and mental health clinics applying by 2026-10-31, running one month alongside existing electronic charts with no automatic conversion to a paid contract.
• The product covers five record-type-specific models (consultation and SOAP notes, nursing, interview, multidisciplinary conference minutes, welfare home visits) plus AI intake questioning and scheduling.
• The company published telemetry from all voice records at a single deployed psychiatric hospital between 2026-02-25 and 2026-08-22: 1,356 conversations totalling 190.1 hours became record drafts, the median time from end of consultation to draft was 18 seconds, and 96.7% of drafts were adopted as the formal record after clinician review or correction.
• Roughly 70% were adopted unchanged and over 90% were finalised within 3 minutes. Staff using it grew from 2 to 7 and monthly use from 9 to 346 records over six months.

Why it matters: ambient documentation is the most widely adopted clinical AI in general medicine, and this is a psychiatry-specific instance, where the material being recorded is a psychotherapy session rather than a physical exam. California's new bill carves out documentation systems that do not involve professional judgment, which is precisely this boundary.

Is there a catch? Every figure is vendor telemetry, self-measured and self-tabulated at one hospital with no controlled comparison. The roughly 135 hours of writing time saved is a modelled estimate against an assumed 8 minutes of handwriting per record rather than anything that was measured.

Reliability ●●●●● 3/5Why 3/5: a genuine first-party launch with an independently verifiable live product, and the release text was retrieved in full from two outlets that match word for word; but the effectiveness claim carrying the item is uncontrolled single-site vendor telemetry. · Sources: Sankei (release) · Jiji

Forcing a language model to follow an eleven-stage CBT session structure produced more engagement from simulated clients than six alternative designs.

• Authors at Virginia Tech, UC Davis and the International Computer Science Institute represent the first CBT session as eleven stages with explicit therapeutic objectives, use an external controller to decide when a stage is complete, and use a deep reinforcement-learning model to select the therapeutic intentions guiding response generation.
• Against six retrieval-, prompting-, stage- and policy-based alternatives, it elicited higher simulated client engagement and openness.
• DeepSAGE achieved the strongest balance of stage-goal completion and dialogue efficiency among the stage-structured systems.
• Expert review found broadly plausible emotional trajectories and recognisable CBT processes.

Why it matters: it is a datapoint on whether manualised therapy structure can be imposed on a generative model at all, which is the technical premise under most of the clinically backed chatbot pipeline. The alternative is to skip the generative model and write rules.

Is there a catch? The authors state that because the evaluation relies primarily on simulated clients and model-based metrics, the findings demonstrate comparative dialogue-control improvements rather than clinical effectiveness. No person in distress was involved, and the paper is not peer reviewed.

Reliability ●●●●● 3/5Why 3/5: a readable preprint whose design and results were checked in the full text, but every headline result rests on simulated clients and model-based metrics with no effect sizes reported, which is thin support for a comparative superiority claim. · Sources: arXiv

A platform that reads video, movement, gaze and voice is selling ADHD assessments at consumer prices, with a licensed clinician authorising the report at the end.

• A release datelined Newark, Del., 24 August 2026 announces the NVOX Network, inviting psychologists, psychiatrists and other qualified clinical professionals to join at no cost.
• The platform combines clinical history, developmental information and objective performance data with behavioural signals from "video observation, movement, gaze and voice," and states that "final report release is subject to licensed clinical review and authorization."
• Pricing is $349 per assessment through the network and $499 through NVOX Direct, which adds professional review, a comprehensive report and a feedback session.
• The release cites more than 1,000 completed assessments during an initial operating phase.

Why it matters: gaze, voice and movement inference is being sold at a consumer price point as an ADHD assessment input, with the clinician positioned as sign-off rather than as the assessor. Missouri now penalises representing AI as capable of a mental health diagnosis.

Is there a catch? NVOX states explicitly that it does not present its internal evidence as a substitute for independent research and is preparing for external validation. No peer-reviewed accuracy, sensitivity or specificity figures exist for the gaze, voice or movement signals that carry the product.

Reliability ●●●●● 3/5Why 3/5: every headline figure was confirmed verbatim in the company's own release and the product verified as live and on sale; but the only substantive support is that release, since the secondary is a syndication of the same text. · Sources: PR Newswire · Health Technology Net

A university institute and a youth suicide-prevention charity will try to translate clinical crisis expertise into the technical choices that shape how chatbots behave.

• On 26 August 2026 Columbia University's Technology & Democracy Initiative, part of SIPA's Institute of Global Politics, and The Jed Foundation announced a partnership to bridge mental-health expertise and AI safety practice.
• They frame the gap as one where "the two fields use different vocabularies and often lack a shared framework for translating what a young person in crisis needs into the technical choices that shape how AI systems behave."
• The announcement cites figures that 64% of American teenagers use AI chatbots and that 12% of those teens use them for emotional support or advice. Both are figures the partners cite rather than measure.
• The first phase is an expert roundtable with timing to be announced, followed by a jointly authored working paper mapping where expert judgment needs to be built into technical systems, then a reference document, "Anatomy of a Mental Health Technical Response."

Why it matters: a named, dated cross-sector attempt at the translation problem, aimed at unsupervised teen chatbot use and at the point where clinical expertise has to become a model specification.

Is there a catch? Nothing has been produced yet: the roundtable, the working paper and the reference document are all forthcoming, so this is agenda-setting rather than a result. The two usage figures are borrowed from elsewhere rather than measured here.

Reliability ●●●●● 3/5Why 3/5: a first-party announcement read directly, with the date, the partners, the cited figures and the three named deliverables all checking out; held at 3 because no artifact has shipped, only forthcoming ones. · Sources: The Jed Foundation

A hospital will test a deliberately rule-based chatbot delivering six weeks of mindfulness and compassion modules to breast cancer survivors.

• NCT07791368 was first posted on 2026-08-27: an interventional, randomized, double-masked parallel trial with an estimated 68 participants aged 18 and over, allocating survivors to a chatbot arm or a waitlist control.
• The registry describes the intervention as "a rule-based AI-Chatbot implemented through Line platform" guiding survivors step by step through six weekly modules of body-mind-spirit empowerment strategies, structured mindfulness skills and compassion-based skills.
• Primary outcomes are PHQ-9 and the WHO-5 Well-Being Index at baseline, week 6 and week 12.
• The trial tests a clinician-designed conversational agent delivering a manualised psychotherapy inside an oncology care pathway.

Why it matters: this is the supervised end of the spectrum, and it was built rule-based rather than generative on purpose. The generative alternative has so far been tested only on simulated clients.

Is there a catch? The estimated start date is August 2027 with primary completion in 2032, so this is a registration of a trial that has not begun and cannot report for years. The comparator is a waitlist, which cannot separate the chatbot from attention or expectancy.

Reliability ●●●●● 4/5Why 4/5: confirmed through the registry's own data record, including the start date and comparator that the summary alone would not reveal; capped at 4 because a registration is a sponsor-submitted record of an intention, with no results. · Sources: ClinicalTrials.gov

Psychiatry trainees should master interviewing, observation and diagnosis before being allowed even supervised use of clinical AI.

• Published in Cureus on 2026-08-25, the review surveys AI across psychiatric education: personalised learning, virtual patient encounters, clinical decision support, EHR synthesis, automated feedback and documentation.
• The review warns that "algorithmic bias, misinformation, privacy, academic integrity, cost, and overreliance may undermine independent clinical judgment."
• It proposes "a staged, competency-based curriculum in which trainees establish core interviewing, observation, and diagnostic skills before progressing to supervised clinical AI use," and argues AI should augment rather than replace human interpretation and professional accountability.
• The review notes limited empirical outcome data on AI curricula in residency and fellowship programs.

Why it matters: deskilling is a sequencing problem before it is a competence problem, and this is a concrete proposal about the sequence. The duty to vet AI tools that certifying bodies are now claiming presumes exactly this training.

Is there a catch? It is a narrative review with no new data, no systematic search protocol and no measured outcome, and the curriculum is the authors' recommendation rather than an evaluated or adopted training standard.

Reliability ●●●●● 3/5Why 3/5: a review article in a peer-reviewed venue with the publication date independently confirmed, but the venue's peer review is light, the piece contributes no new data, and its central claim is a recommendation. · Sources: Cureus · article page

A newly launched app scans your face through the phone camera, states your emotional state, and recommends music, breathwork, yoga or meditation.

• Nihilent Limited announced nSEPIA's global launch on 28 August 2026. Per the release, "The experience begins with a brief, non-invasive interaction using the device's camera."
• The release continues: "Using Emotion AI, nSEPIA analyses subtle spatial and temporal movement patterns across the face and eye region, together with head and upper-body movement, to assess the user's emotional state. The assessment is processed on device to protect user privacy."
• Nihilent describes the underlying Emoscape platform as "medical-grade," built on a patent-protected framework rooted in the Navarasa principles of the Natyashastra, and says it "has been evaluated through scientific and academic research for its ability to assess human emotions objectively." The product page calls it a "medically validated Emotion AI engine."
• The app is live with a Google Play download, and the user can reassess after following the recommendations.

Why it matters: a live consumer product is asserting clinical-grade facial affect inference, a class of inference with a weak evidence base generally. Missouri's new prohibition is aimed at exactly this kind of claim, though it binds only in Missouri.

Is there a catch? No study, journal, sample size, accuracy figure or regulatory clearance is cited anywhere for the "medical-grade" or "medically validated" claims, and none was findable. Both phrases are the vendor's own.

Reliability ●●●●● 3/5Why 3/5: the launch and the shipped app were verified beyond the release, including a live product page and a store listing; held at 3 because the announcement is paid distribution with a single corporate voice, no independent coverage, and no evidence at all behind the load-bearing claims. · Sources: MTI News