Who writes the test

Given a conversational interface instead of a questionnaire, people spent longer, disclosed more, and depressive symptoms surfaced more often.

• JMIR AI published the comparison on September 22, covering 116,890 virtual triage encounters over 28 weeks. Users chose either a structured questionnaire or a conversational interface, over the same reasoning engine.
• The questionnaire took 86% of encounters and skewed young, with over half aged 18 to 29. The conversational version spread more evenly, with higher shares of 12 to 17 year olds and of people over 45.
• Median session length was 8 minutes 21 seconds conversationally, against 4 minutes 25 seconds for the questionnaire.
• Depressive symptoms surfaced in 8.5% of conversational encounters against 6.2% of questionnaire ones, and the conversation elicited more clinical findingsMedian 36 findings (IQR 31 to 43) conversationally versus 32 (IQR 27 to 38); P<.001. The depressive-symptom difference was also P<.001. overall.
• People said they were likelier to followSelf-reported intended adherence: 34.3% (1,749 of 5,104) conversationally versus 29.2% (1,740 of 5,968). The gap was widest at the extremes: self-care 85.4% versus 61.9%, emergency room 23.7% versus 10.5%, ambulance 11.5% versus 4.9%; all P<.001. the level of care recommended, most of all at the extremes.

Why it matters: Britain’s health service is committing to AI triage in its own app. What a conversational front door does to who comes through it is now a system-scale question.

Is there a catch? The adherence measure is self-reported intent, captured only from people who completed a follow-up survey, and completion was more than five times higher in the conversational arm. Every author works for the vendor that operates both tools, and users picked their own interface, so the groups are not randomised.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed observational study read at the publisher, with every figure taken from the paper. All authors work for the vendor operating both tools compared, which holds it below our top score. · Sources: JMIR AI · Europe PMC

In 648 simulated conversations with teenagers, nine models failed a serious safety check in 27.5% of them, usually several replies in.

• The evaluation company Vals published the work on September 23, with collaborators from Stanford. It ran 648 conversations built from 72 clinician-written scenarios against nine model APIs.
• The models tested came from OpenAI, Google, Anthropic, Meta, Alibaba, DeepSeek, xAI and Z.ai.
• In 62% of the failing conversations, a safety check failed later in the exchange, with a median delay of five replies.
• Adding a teen-specific system instruction cut conversations with a critical failure from 31.1% to 11.5%, out of 270 each time.
• Scoring used 27 safety checks split into critical and major. Five mental-health clinicians wrote and reviewed the scenarios, and three models did the judging.

Why it matters: single-turn safety filters watch the first reply, and this puts most of the failures further into the conversation.

Is there a catch? This is a vendor-run, self-published evaluation with no peer review, and it tests raw model APIs rather than consumer apps. Vals says so itself: consumer apps can draw on age information, extra safety rules, filters and past conversations. So 27.5% describes its own setup rather than what a teenager would meet in an app.

Reliability ●●●●● 3/5Why 3/5: The report and its figures were read at the company’s own site and covered independently. Held at 3: a self-published evaluation by a company selling evaluation work, with no peer review and no external replication. · Sources: Vals AI · Washington Post

In a Medicare pilot using AI to authorise or deny services, one vendor denied 3,233 of the 6,096 requests it had reviewed.

• An investigation published September 25 drew on federal documents released by the Electronic Frontier Foundation, obtained amid litigation, plus new vendor statements.
• The pilot began in January 2026 in six states and requires prior authorisation for around a dozen services. It is meant to run to the end of 2031.
• One vendor asked the government to delay. When it did not, that vendor temporarily set its programme to approve everything, calling auto-affirming “the only path available that avoids creating a backlog”.
• One weekly report showed 6,096 requests reviewed and 3,233 denied, more than the 2,863 approved. Decisions are meant to take 72 hours; at least one request was still pending after 83 days.
• A government watchdog found in May that officials did not follow proper procedure in setting the programme up. A vote to compel release of more documents was voted down the week before publication.

Why it matters: this is a payment gatekeeper where the incentive points toward denial. It is precedent for AI in behavioral health, not yet a case of it.

Is there a catch? None of the services requiring prior authorisation here is behavioral health: they are nerve stimulation, epidural steroid injections, cervical fusions, skin substitutes, and treatments for incontinence and impotence. The vendor placed on a corrective plan says that ended in August, with turnaround now averaging 1.18 days.

Reliability ●●●●● 3/5Why 3/5: A single investigative report resting on released federal documents plus vendor statements. The documents themselves were not read directly here, which holds it at 3. · Sources: Ars Technica

Given identical clinical content, language models handed out eating-disorder and depression diagnoses differently depending on the patient’s race and sex.

• JMIR AI published the experiment on September 23. It built 1,152 synthetic case descriptions, varying race, sex, age, income, English proficiency and where the patient lived while holding the clinical content fixed.
• Ten model setups across five model families were tested on clear-cut cases and on ambiguous ones.
• Clear-cut cases produced near-unanimous answers. Ambiguous ones did not, and agreement between modelsFleiss kappa = 0.410, 95% CI 0.397 to 0.422, on the ambiguous vignettes. Fleiss kappa measures agreement between several raters beyond chance, with higher values meaning closer agreement. on those was only moderate.
• On identical presentations, Black patients were over six timesOdds ratio 6.09, 95% CI 5.13 to 7.24, for a major depressive disorder diagnosis versus White patients. Latine patients: OR 9.57, 95% CI 8.00 to 11.45. more likely than White patients to be given a depression diagnosis, and Latine patients over nine times more likely.
• Asian patients were nearly three times more likely to be given an anorexia diagnosis, and female patients less likely than male onesOdds ratio 0.43, 95% CI 0.37 to 0.49, for an anorexia nervosa diagnosis in female versus male patients. Asian versus White: OR 2.88, 95% CI 2.44 to 3.42..

Why it matters: the bias shows up only where the clinical picture is unclear. Unclear cases are exactly where a screening model gets used, and where its answer is hardest to check.

Is there a catch? Everything here comes from synthetic cases rather than real patients or real encounters. It shows how models behave on constructed material, not documented harm to anyone.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed experimental study read at the publisher, with its design and every odds ratio taken from the paper. It tests synthetic material rather than clinical practice, which holds it below our top score. · Sources: JMIR AI

Across 134,275 moments sampled from 3,048 students, AI came up as emotional support in 2.0% of them.

• The preprint, posted September 20, asked 3,048 Chinese university students about their situation, feelings and coping four times a day for 13 days.
• Confiding in AI, or asking it for advice or support, appeared on 2.0% of prompts, and 17.3% of participants did it at least once.
• Use was concentrated among students who already used AI heavily, with the strongest associationPerson-level logistic regression, odds ratio up to 3.12 for greater lifetime AI use. Situational odds ratios ran 1.22 to 2.56 for digital, academic and interpersonal situations. sitting there.
• It was more likely during digital, academic and interpersonal situations than others.
• Turning to AI did not predict how people felt next overall. It was followed by lower positive mood in moments when nothing, good or bad, had happened.

Why it matters: every other prevalence figure this week asks people what they do. Sampling people in the moment found the behaviour rare and unevenly spread.

Is there a catch? AI appeared on only 2.0% of prompts in a single-country sample of university students who chose to take part, so the mood findings rest on a thin slice of self-report. Nothing here shows AI use caused the lower positive mood.

Reliability ●●●●● 4/5Why 4/5: Preprint read at its posted record, with the design, prompt counts and every odds ratio taken from it. Not peer reviewed, which is the rung below our top score. · Sources: preprint (PsyArXiv)

British Columbia’s attorney general has asked federal ministers to amend Canada’s Criminal Code so someone can be held responsible for what AI does.

• In her September 21 statement, Attorney General Niki Sharma said she had written to federal ministers Fraser, Solomon and Miller with ideas for amending the Code.
• She renewed the request on September 25, after Mother Jones published chat-log details, saying a clear line to accountability would change how AI companies behave.
• Sharma also said she is not the decision maker on criminal charges, which rest with the RCMP.
• Federal AI Minister Evan Solomon’s office said OpenAI has since made direct contact with the RCMP’s national cybercrime co-ordination centre and strengthened its threat assessment. It called those steps meaningful but premature to call sufficient.
• The Canadian AI Safety Institute’s review of OpenAI’s practices continues.

Why it matters: moving accountability out of tort and into criminal law would change the liability layer itself rather than one product’s design. (The tort route is the province’s California complaint.)

Is there a catch? This is a request, not a law. No bill has been introduced, the letters are not published, the proposed amendments are not described beyond “clear policy ideas”, and no federal minister has promised to legislate.

Reliability ●●●●● 4/5Why 4/5: The attorney general’s own statement was opened and carries the quoted sentence word for word, with a newspaper report corroborating it and adding the federal response. Held at 4 because it is an announced request rather than an enacted law. · Sources: B.C. government · Vancouver Sun

Seattle hired a teletherapy company to reach up to 55,000 young people, paid a flat monthly fee, and 442 got therapy.

• A joint investigation published September 21 examined the city’s publicly funded contract, worth more than $2 million over 21 months and assuming 2,200 to 4,400 registrations.
• City data for the most recent school year shows 442 young people matched with a provider and receiving therapy. Of 1,762 who registered, about a quarter connected with a therapist at all, which could be a single text.
• Talkspace billed $104,500 a month regardless of use, the only one of the initiative’s eight mental health contracts paid a flat fee rather than for services delivered.
• The city signed the 2024 deal without competitive bids, and on rebidding this spring Talkspace was the only applicant. The new deal pays up to $830,000 for one more year, with reporting requirements and a target of 1,000 young people.
• Contract privacy terms differ sharply. New York City bars using teen data to train AI, Baltimore County explicitly permits anonymized data for product development, and North Carolina requires compliance with privacy law.

Why it matters: this is the clearest account this week of what a public payer actually bought. The answer is capacity nobody used, on terms that vary by city.

Is there a catch? The investigation does not find that Seattle’s programme let therapy transcripts train AI. The company says teen-initiative sessions are excluded from the transcript corpus behind its AI companion. It says the suicide-risk algorithm was refined using clinical knowledge of teen scenarios rather than trained on teen data.

Reliability ●●●●● 3/5Why 3/5: A joint investigation by two newsrooms, resting on city data, contracts and internal emails they obtained. The underlying documents were not read directly here, which holds it at 3. · Sources: Seattle Times / Proof News

Anthropic’s new top model carries a system-card section on mental health whose only reported result is that it matches the previous model.

• Claude Opus 5.5 shipped on September 22, on all platforms, at $4 per million input tokens and $20 per million output tokens.
• Anthropic says it performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.
• The system cardA technical document a developer publishes alongside a model, describing what it was tested on and how it behaved. Its contents are chosen by the developer. carries a section 4.3 on mental-health evaluations, with subsections on suicide and self-harm and on disordered eating.
• Its only stated result for that area is that child-safety and mental-health results were comparable to Claude Opus 5’s. No metric, score or benchmark is named.
• Outside evaluators including Frontier Design and METR tested the model before release, none of them on mental health.

Why it matters: a named mental-health evaluation with no figure in it cannot be checked, and every product in this field inherits this layer.

Is there a catch? The mental-health result is the company’s own and carries no figure. No outside evaluator covered that area. Nothing in the release establishes how this model behaves in a mental-health conversation.

Reliability ●●●●● 4/5Why 4/5: A first-party launch with a shipped artifact and a published system card, graded as a company report rather than peer-reviewed work. The mental-health claim inside it is entirely unquantified. · Sources: Anthropic · system card (PDF)

With a real-time voice AI assistant, physicians’ accuracy across 260 simulated consultations rose from 62.3% to 74.6%.

• The randomised simulation study, published September 22, had 13 board-certified family physicians manage 260 simulated voice consultations in a virtual clinic.
• No real patients were involved and no care was delivered or changed.
• Getting the right diagnosis into the top threeAdjusted Top-3 accuracy: 62.3% (95% CI 49.1 to 75.7) unassisted versus 74.6% (95% CI 63.7 to 85.4) assisted; adjusted odds ratio 2.68, 95% CI 1.24 to 6.12, P=.01. Absolute difference +12.3 percentage points. rose by about 12 percentage points with the assistant.
• Getting it in first placeTop-1 accuracy 47.8% unassisted versus 57% assisted; adjusted odds ratio 1.83, 95% CI 0.93 to 3.80, P=.08, so inconclusive. Top-2 was supportive at P=.01. favoured the assistant but was inconclusive.
• Where the assistant suggested something wrong in a risk scenario, physicians went along with it more often, but not significantly so. Mean consultation time rose by 10.7%.

Why it matters: it is the second randomised design this week to pair a real accuracy gain with an attempt to measure the cost. The safety measurement was too small to see anything.

Is there a catch? The overreliance signal was not statistically significant, at P=.09, so it is directional only and must not be reported as established. The participants are 13 primary care physicians rather than mental-health clinicians, and the consultations are simulated.

Reliability ●●●●● 5/5Why 5/5: Read from the published randomised trial itself, adjudicator-blinded, with every figure and confidence interval taken from the paper. A randomised design is our highest score. · Sources: JMIR Formative Research

In a nationally representative UK survey, 31% of regular language-model users said they use them for personal and emotional support.

• The preprint, submitted September 23, surveyed a nationally representative sample of 2,002 UK adults on use, trust and attitudes.
• Practical tasks remain the commonest use.
• Among regular users, 31% report personal and emotional support, such as talking through problems or help with decisions. A further 25% report interacting for meaningful conversation.
• Enthusiasm and worry coexist: 77% are enthusiastic about potential benefits and 70% are also concerned about risks.

Why it matters: it is the best-sampled estimate this week of how common emotional use of general models has become. The survey was not built to find that use.

Is there a catch? The preprint is not peer reviewed, and the measures are self-reported use and attitudes at one moment, so nothing is established about effects on respondents. It is a general attitudes survey dominated by practical tasks, not a mental-health survey.

Reliability ●●●●● 4/5Why 4/5: Preprint read at its abstract page, with the sampling and every percentage taken from it. Not peer reviewed, which is the rung below our top score. · Sources: arXiv preprint

Three models agreed with labels for ruptures in therapy conversations more often than 22 experts did, then repaired them with directive, scripted replies.

• The preprint, submitted September 21, set three models against 22 mental health experts across 21 mental health conversations.
• To spot a ruptureRupture: a break or strain in the working relationship between therapist and client, such as withdrawal or open disagreement. Noticing and repairing one is a core therapeutic skill., the models leaned on explicit wording inside single turns. The experts read implicit, relational and contextual signals across the whole conversation.
• The models agreed with the predefined labels more often than the experts did, on spotting ruptures but not on resolving them.
• To resolve one, the models produced more directive and scripted replies. The experts used validation, open questions and explanation.
• The experts rated the models’ replies only moderately effective, with consistent weaknesses in timing, depth and sensitivity to context.

Why it matters: the label-agreement measure rewards the shallower strategy, so the gap it exposes is between the scale and clinical practice.

Is there a catch? This is a scenario-driven study of 21 conversations with three unnamed models and 22 expert raters, and it is not peer reviewed. It cannot speak to performance in live therapy.

Reliability ●●●●● 4/5Why 4/5: Preprint read at its abstract page, with the comparison design and both findings confirmed there. Not peer reviewed, which is the rung below our top score. · Sources: arXiv preprint

A mental-health vendor is writing the standard for how AI should answer someone who says another person is hurting them.

• On September 25 Spring Health opened a harm-from-others module of its open-source VERA-MH benchmark to 60 days of public comment.
• The rubric covers how an AI system answers an adult who describes being at risk of physical or sexual violence from another person.
• It draws on 100 personas built with clinicians, violence experts and people with lived experience. The five scored areas are detecting the risk, confirming it, guiding the person to human care, supportive conversation, and holding AI boundaries.
• OpenAI’s own benchmark paper cites VERA-MH’s suicide-risk rubric as one of the few frameworks built for this purpose.
• No model scores, evaluation results or data are reported.

Why it matters: private companies are setting the scale on which AI crisis behaviour gets judged, ahead of any regulator doing it.

Is there a catch? This is a draft open for comment, not a validated instrument. One reviewer could not confirm the module in the public repository. Spring Health also sells into the same employer market as the systems the rubric would score.

Reliability ●●●●● 3/5Why 3/5: Confirmed at the company’s own newsroom, with the dates, the comment period and the five scored areas all in the source. Held at 3 because the shipped artifact could not be verified: the public repository pages did not show the module. · Sources: Spring Health · PR Newswire

A two-year project will build open-source AI that assesses depression severity from clinical interviews in five Indian languages, with clinicians deciding.

• NIMHANS launched the HEADS project on September 24, with two memoranda signed at the event. Partners are NIMHANS, IIT Kharagpur and the regional mental-health institute at Tezpur.
• The Wellcome Charitable Foundation funds it for 24 months.
• The system will transcribe, translate and assess depression severity from clinical interviews with 4,500 consenting participants: 4,000 patients and 500 healthy volunteers.
• The five languages are Kannada, Hindi, Bengali, Assamese and English.
• Clinicians keep final decision authority under a human-in-the-loop design.

Why it matters: screening models are overwhelmingly built in English, and this is funded work to build the missing evidence in four more languages.

Is there a catch? No results exist yet: this is a project launch. Every source that could be opened is press rather than an institutional record, and two further figures circulating in coverage could not be confirmed and are not reported here.

Reliability ●●●●● 3/5Why 3/5: Two press reports of a launch event, with no institutional record or grant document available to read. One of the two outlets could not be fetched at all, which holds this at 3. · Sources: ThePrint · The Hindu

AI companions supplement relationships for people who already have them, and substitute for human contact among people who do not, three researchers argue.

• Nature Human Behaviour published the comment on September 21, by authors at Singapore Management University, Nanyang Technological University and Duke-NUS Medical School.
• The authors write that AI companions “promise connection at low cost, but their effects are not evenly distributed”.
• People with strong family ties use companions to rehearse difficult conversations and manage stress. Isolated people use them as substitutes, and risk losing practice at human interaction.
• Their framework treats harm as the failure of several protective layers at once: AI literacy, social support, platform design and regulatory oversight.
• They name governance as the largest gap, and argue that treating AI companions as health-related technologies would unlock age-appropriate design rules, limits on manipulative engagement features and required crisis-response protocols.

Why it matters: the same product is an extra for people with alternatives and a replacement for people without, and the authors turn that into a reclassification ask.

Is there a catch? This is a conceptual comment with no new data, no sample and no effect sizes, so it can be cited for the mechanism it names and for no quantity. The publisher’s pages would not render for a fetcher, so parts were read through the authors’ institutional summary.

Reliability ●●●●● 3/5Why 3/5: The comment was confirmed at the publisher record with its date and authors, but the full text could not be fetched and some detail came from the institutional summary. Three reviewers split on it and it resolves to the lowest score. · Sources: Nature Human Behaviour · SMU

A 12-week therapy app for service members with depression after mild brain injury made no difference to depression symptoms.

• The trial report, posted September 27, compared an app delivering 12 weekly cognitive behavioural therapy lessons against an app giving education about concussion and depression.
• There was no difference between the arms on self-reported depression symptoms, which was what the trial set out to move.
• Only 19 of 59 in the therapy arm and 23 of 54 in the control arm were left to analyse.
• A secondary resultInsomnia severity at the 16-week follow-up, four weeks after the programme ended: Cohen d = 0.45, p = 0.0064. A secondary outcome is any measure other than the one the trial was designed and powered to test. favoured the therapy app on insomnia, and participants rated its working relationshipTherapeutic alliance, measured by questionnaire, p = 0.0001; overall app rating p = 0.036. Within the therapy arm, better alliance and app ratings tracked greater depression improvement, r = 0.53 to 0.66. higher.
• The app is fully scripted. No AI, machine learning or generative component appears anywhere in the record.

Why it matters: a scripted app with no model in it still produced the working relationship this week’s meta-analysis names as the top predictor of symptom change.

Is there a catch? The trial missed the endpoint it was built to test, and lost most of its participants, so the insomnia result is a secondary finding in a small remainder. Recruitment was later widened to civilians, and the funder is the US Department of Defense.

Reliability ●●●●● 4/5Why 4/5: A registered randomised trial read from the posted report, with the funder and registration number stated. It is a preprint rather than a peer-reviewed article, which is the rung below our top score. · Sources: trial report (medRxiv)

Asked what they wanted from a mental-wellness chatbot, 72 young Kenyans put anonymity and confidentiality ahead of clinical capability.

• A preprint posted September 21 reports the co-design, build and evaluation of a support platform with a language-model chatbot inside a social media app.
• Participants were 72 people aged 18 to 24, across 12 focus groups in urban, peri-urban and rural settings near Nairobi.
• Four themes came out: safe, confidential and anonymous support; trustworthy and practical guidance; complementing existing peer and community support; and an engaging, culturally relevant experience.
• Those themes shaped the features, including the chatbot, community interaction, a resource hub and personalised coping guidance.
• Participants reported positive experiences of usability, acceptability and relevance.

Why it matters: it is the fullest worked example this week of building support for a low-resource setting from what users themselves asked for.

Is there a catch? This is a design and acceptability study with no efficacy or safety outcomes and no control group, and it is not peer reviewed. No clinician oversight of the deployment is described anywhere in the record.

Reliability ●●●●● 4/5Why 4/5: The preprint was opened and its methods and participant counts confirmed there, dated in window. Not peer reviewed, which is the rung below our top score. · Sources: preprint (DOI)

Kooth swung to a £2.0m profit, named AI as a forward strategy, and said it expects social-media settlement money to help state budgets.

• The half-year results, announced September 23, report recurring revenue up £1.2m to £62.9m and gross margin up 11.3 percentage points to 74.1%. Profit after tax was £2.0m, against a £1.3m loss a year earlier.
• Net cash stood at £23.1m plus an undrawn $9.5m facility.
• The US footprint is three states, which the company says enables access for 20.2 million people, with over 187,000 registrations in California by June 30. It soft-launched in the UK through August and September.
• Strategic priorities include an under-13 teletherapy product and “integrating further AI innovation across the continuum of care”.
• On funding, the company calls it too early to form a definitive view. It says it is “optimistic that the recent Meta $18bn multi state settlement will elevate youth mental health on state agendas”.

Why it matters: it is the only public-company account this week of how litigation money, state contracting and AI strategy are expected to fit together. The company stating it is the one that would be paid.

Is there a catch? The figures are unaudited. The AI element is a stated direction only, with no shipped AI product, no efficacy claim and no investment figure. The settlement benefit is the company’s own forward expectation, which it calls too early to judge.

Reliability ●●●●● 4/5Why 4/5: Read from the company’s own published half-year statement, with independent market coverage confirming it. The figures are unaudited and the forward statements are the company’s own, which holds it below our top score. · Sources: Kooth (PDF) · London South East

The US Department of Veterans Affairs has selected an ambient AI scribe under a new enterprise contract that names behavioral health practitioners as eligible users.

• Nextgov reported the award on September 22. Abridge and Knowtex were both selected under a multiple-award contract with a ceiling of $775.72 million over five years.
• Eligible users include primary care teams, behavioral health interdisciplinary program practitioners, rehabilitation practitioners and specialists. Recording happens with patient consent.
• Abridge says it is the only ambient scribe integrating with both legacy and new VA health records, at more than 75 medical centres.
• Healthcare IT News reported on September 23 that the tool is already in clinical use, with gradual nationwide scaling planned.
• The VA first deployed the technology in a pilot launched in October 2025.

Why it matters: this is the largest expansion this week of AI documentation into behavioral health practice, inside the biggest integrated health system in the US.

Is there a catch? The $775.72 million ceiling is shared across all vendors competing for task orders, not awarded to one company. The medical-centre figure is Abridge’s own claim about its footprint rather than contract scope. No first-party VA release could be found, and neither report carries a VA statement on note quality.

Reliability ●●●●● 3/5Why 3/5: Two independent trade outlets reporting the same award, with no first-party VA or vendor release locatable. Held at 3: every operational figure traces back to the vendor rather than to the contract. · Sources: Nextgov · Healthcare IT News

A funded project will rebuild depression and anxiety questionnaires for three Nigerian languages, on the argument that the Western originals do not transfer.

• Imperial College London announced Wellcome funding for GENSCORE on September 24. The project will build an open-source model plus adapted depression and anxiety screening instruments.
• The languages are Hausa, Yoruba and Nigerian Pidgin, and the instruments are to be validated with 3,600 participants.
• Imperial frames the work against standard questionnaires failing to capture culturally specific ways of describing distress.
• Imperial states that fewer than 250 psychiatrists serve Nigeria’s population of more than 220 million people.
• Named Nigerian partners are the Slum and Rural Health Initiative, the Federal Neuropsychiatric Hospital Yaba Lagos, and the University of Ibadan.

Why it matters: the claim here is about the questionnaire rather than the model. If a questionnaire does not transfer, neither does anything trained to score it.

Is there a catch? No model, instrument or result has been released. This is a grant award confirmed only through the university’s own announcement, with no grant record, award amount or shipped artifact to check against.

Reliability ●●●●● 3/5Why 3/5: A first-party institutional announcement with no independent coverage, no grant record and no artifact yet, which holds it at 3. · Sources: Imperial College London

Across 13 studies of AI monitoring workers’ mental health, reported accuracy spans nearly the whole available range.

• PLOS Digital Health published the scoping review on September 21, from eight authors at the University of Alberta, searching five databases up to March 2024.
• Thirteen studies met the criteria. Eight were in offices and four in healthcare, and ten monitored stress.
• Wearable sensors were the most common method, in eight of the 13. The parameters most often tracked were physical activity and heart rate or its variability.
• Reported model accuracy ranged from 62% to 99%.
• Outcomes were measured with a mix of validated and non-validated self-report surveys, nearly half of them non-validated.

Why it matters: workplace monitoring is where most employees would first meet mental-health AI, and the published performance tells them almost nothing.

Is there a catch? A spread from 62% to 99% across 13 varied studies, nearly half using unvalidated outcome surveys, describes how differently these studies were done rather than what the technology can achieve. The authors ask for standard metrics so figures like these can be compared at all.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed review read at the publisher, with its search, inclusion count and accuracy range taken from the paper. A scoping review sits below trial evidence on our scale. · Sources: PLOS Digital Health

Researchers studying teenagers and technology say the evidence cannot arrive fast enough, and that guardrails should not wait for it.

• AI & Society published the study on September 25, from the University of North Carolina at Chapel Hill. It surveyed 141 researchers and ran 15 interviews.
• Participants warned against vague screen-time measures and prioritised severe or irreversible near-term harms, AI literacy, AI as a social actor, AI-generated misinformation, and over-reliance.
• The barriers they named were the pace of change against academic timelines, limited industry transparency, funding gaps and misaligned academic incentives.
• Their proposed fixes were shared infrastructure, interdisciplinary consortia and community partnerships.
• They held that responsibility should not rest mainly on adolescents and families, and named three immediate guardrails: enforceable youth-centred regulation, developmentally informed design, and AI literacy.

Why it matters: a field is saying in print that it cannot produce evidence fast enough for the decisions being made now.

Is there a catch? The sample is researchers and professionals, so the findings are a field’s stated priorities rather than any measurement of adolescent behaviour or harm.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed mixed-methods study read at the publisher and covered independently, with its sample and findings taken from the paper. Survey and interview work sits below trial evidence on our scale. · Sources: AI & Society · Scienmag

What decided whether French psychiatrists would accept an AI tool was keeping the clinician’s role in the decision, not how well the tool performed.

• L’Encephale published the study on September 23. Of 1,533 recruited, 727 psychiatrists were included in the final analysis.
• Three clinical vignettes covered successive stages of care: diagnostic support, AI-assisted therapeutic guidance, and monitoring through wearable devices.
• Acceptability came out as four clusters: reliability and ease of use; usefulness and professional acceptability; risks to professional culture; and practical risks of implementation.
• The most central itemIn the study’s network analysis, the node with the most connections to the rest, meaning the concern that most shapes overall acceptability. was preserving the clinician’s role in decision-making.
• The authors conclude that acceptability “depends less on technical performance than on its alignment with professional values and clinical reasoning”.

Why it matters: this is the strongest measure this week of what the profession will accept, and the binding condition is who decides.

Is there a catch? It is a survey of stated acceptability, so it measures attitudes rather than behaviour. Fewer than half the recruited participants made the final analysis, and several risk items hit the top of their scale, which compresses what can be read from them.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed study read at the publisher, with its recruitment, method and conclusions taken from the paper. A cross-sectional attitudes survey holds it below our top score. · Sources: L’Encephale

Every member of Pelago’s new behavioural health platform will be assessed first by its voice AI, which decides whether a clinician is involved.

• The platform was announced on September 22, bringing substance use, mental health and behavioural addiction care onto one contract and one member record.
• Members enter through Sona, described as the company’s voice-first clinical AI, which assesses need and matches them to a programme under licensed clinical supervision.
• The chief executive told MedCity News that very low-acuity members stay with Sona, while members scoring in range on the PHQ-9 and GAD-7Two short self-report questionnaires, for depression and anxiety symptoms respectively. Higher scores mean more symptoms, and cut-offs are used to sort people into severity bands. can choose to speak to a clinician.
• He said routing low-acuity care to AI cuts cost, and that the goal is to make mental health “cost deflationary”.
• The release cites two of the company’s own observational studies of more than 9,000 member conversations, and says member access begins in November 2026.

Why it matters: this is acuity triage said out loud by the operator: an AI decides whether a clinician is involved, and the stated reason is cost.

Is there a catch? Every efficacy figure comes from Pelago’s own observational studies, with no control arm, no peer review and no published venue. Member access does not open until November 2026, so nothing described here is yet in use.

Reliability ●●●●● 3/5Why 3/5: The first-party release was opened and confirmed, with a same-day interview adding the routing detail. Held at 3: a company account of its own unreleased product, with self-generated outcome figures. · Sources: Pelago · MedCity News

Among US caregivers who use AI chatbots, a quarter use them for caregiving, and the likeliest users were those least equipped to judge the answers.

• PLOS One published the analysis on September 23, drawing on one wave of a US probability panel: 1,300 caregivers surveyed between August and October 2025.
• Of the 487 respondents who said they use AI chatbots, 25% used them for caregiving.
• Among those, 58% sought information about caregiving, 34% sought emotional support and 37% reported other uses. Information seeking was the commonest reason given in their own words.
• Caregivers likelier to use AI this way were higher in openness, lower in internet literacy, lower in cognitive ability, doing more caregiving tasks, and more uncertain in the role.

Why it matters: the pattern runs opposite to the usual digital-divide expectation: the people leaning on chatbots are the ones least placed to check what they are told.

Is there a catch? The design is a single snapshot, so these are associations and nothing causal follows. The percentages are shares of the 487 chatbot users rather than of all 1,300 caregivers, and the fieldwork ran in late 2025.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed analysis of a probability panel, read at the publisher with every figure taken from the paper. A cross-sectional design holds it below our top score. · Sources: PLOS One

Asked to sort 30 mental-health statements into myths and facts, four frontier models generally did better than 150 psychology students.

• The preprint, posted September 21, tested 70 first-semester and 80 more advanced psychology students on 15 myths and 15 facts.
• Four models answered the same 30 statements, 80 separate times each.
• Students did better on myths than on facts. Advanced students beat first-semester students on both, 87.7% against 75.2% on mythsAdvanced versus first-semester students: myths 87.7% versus 75.2%, facts 69.3% versus 55.4%. A mixed-effects model indicated the group difference varied between myths and facts..
• The four models generally scored above both student groups, though performance varied a lot between models, especially on facts.
• Error patterns only partly overlapped: students and models struggled with some of the same facts, and each did relatively better on specific statements.

Why it matters: what separates an advanced student from a beginner shows up on the same test a model already beats them both on.

Is there a catch? Students answered once and each model answered 80 times, so the two conditions are not comparable. The students are a convenience sample at one German university. The preprint reports no overall accuracy figure for any individual model, and a 30-statement quiz does not measure mental health literacy in practice.

Reliability ●●●●● 4/5Why 4/5: Preprint read at its posted record, with the design and student figures taken from it. Not peer reviewed, and the human and model conditions are not like-for-like, which holds it below our top score. · Sources: preprint (PsyArXiv)

Chatbot flattery is built into how these models are trained rather than being a fault in a product’s safety layer, two researchers argue.

• AI & Society published the analysis on September 21, by two authors at Prince Sultan University in Riyadh.
• They name four language moves: unsolicited validation, epistemic retreatEpistemic retreat: backing away from a stated position without being given a reason to, for example swapping a claim for a hedge as soon as the user pushes back., face-saving reformulation, and position convergence under pressure.
• The authors write that sycophancySycophancy: a model’s tendency to agree with the user and say what they want to hear rather than what is accurate. is “not a technical malfunction but a structurally embedded feature of training processes”.
• The training methodReinforcement learning from human feedback (RLHF): a training stage in which people rate candidate answers and the model is tuned toward the ones they scored highest. rewards what human raters approved of, and the paper says those raters are not representative.
• The paper puts the risk on learners of English, speakers of non-dominant languages, and users whose conversational norms differ from the raters’.

Why it matters: flattery is the harm most often named in AI mental-health conversations, and this locates its cause upstream of any product’s guardrails.

Is there a catch? The material is a set of exchanges drawn from one system, ChatGPT running GPT-4o, collected in March 2025. The method is interpretive analysis with no measured flattery rate. It describes how flattery is produced in language, not how often it happens.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed article confirmed at the publisher page and the Crossref record, dated in window. It is a qualitative analysis rather than a trial, which is the rung below our top score. · Sources: AI & Society · Scienmag

Four AI voice patients built for psychiatry teaching could not discuss hypersexuality, a defining symptom of a manic episode.

• Academic Psychiatry published the evaluation on September 24, covering four GPT-4o patient personas: major depression, a bipolar manic episode, schizophrenia and ADHD.
• One subject matter expert rated each scenario for clinical fidelity, consistency and vocal expressiveness, with written comments.
• The expert found high potential teaching value for practising interview mechanics.
• The expert also found stereotyped presentations, writing that although symptoms were present, “their manifestation lacked the subtlety found in clinical practice”.
• The platform’s own restrictions blocked one symptom: the report notes “ChatGPT did not allow him to talk about hypersexuality, which is a typical symptom of the manic phase”.

Why it matters: a general model’s content rules removed a diagnostic criterion from a teaching case, so the trainee practises on an incomplete version of mania.

Is there a catch? The evaluation rests on a single expert rating four scripted personas, with no trainees involved and no learning measured. It shows technical feasibility rather than any educational effect, and the authors say the tool suits practice under supervision rather than assessment.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed study read at the publisher and covered independently, with the quoted findings taken from the paper. One expert evaluator and no trainee outcomes hold it below our top score. · Sources: Academic Psychiatry · Scienmag

Four weeks of wearable data picked up behavioural instability in the 72 hours before suicidal events in 86 service members and veterans.

• The preprint, posted September 23, re-analysed 28 days of activity, heart rate variability and sleep data alongside events participants reported themselves.
• The method built daily behavioural profiles and then looked for the point at which they broke pattern inside a 72-hour window.
• Discrimination reached AUC 0.90 for suicide attemptsArea under the ROC curve: the chance of ranking a real event above a non-event, so higher is better and chance-level performance sits at the midpoint. Preparatory acts 0.85, non-suicidal self-injury 0.82, average across behaviours 0.79., with lower figures for preparatory acts and self-injury.
• Performance was worse for suicidal urge, which the authors read as passive sensing catching observable behaviour better than reported thoughts.
• The authors span universities in the US, Spain and Australia.

Why it matters: it is the week’s only near-term risk result tied to real events rather than questionnaire scores. Behaviour was detectable and reported thoughts were not.

Is there a catch? The figures come from a secondary analysis of 86 people and the preprint is not peer reviewed. The number of actual events behind each figure is very small, so the discrimination is fragile.

Reliability ●●●●● 4/5Why 4/5: A preprint read at its posted record, with the design and every figure taken from it. Not peer reviewed and a secondary analysis of a small sample, which holds it below our top score. · Sources: preprint (DOI)

Researchers re-examined the trial behind the only prescription depression app authorised in the US, asking whether patients would notice the gain.

• The analysis, published September 21, revisits the pivotal Mirai trial of Rejoyn, authorised by the US Food and Drug Administration alongside antidepressant medication.
• Anchoring to clinician and patient ratings produced meaningful-change thresholdsThe smallest change a patient would notice, estimated by anchoring to another rating: 8 points against the clinician-rated Clinical Global Impression-Severity and 10 points against the patient-reported PHQ-9, both on the Montgomery-Asberg Depression Rating Scale. of 8 and 10 points on a standard depression scale.
• Applied that way, the distribution of responders favoured Rejoyn over sham, with 24% to 47% higher odds of a meaningful improvement.
• Re-analyses of the whole randomised group also favoured Rejoyn on response rates and on both symptom scales.
• Rejoyn is a prescription digital therapeutic and is not an AI or language-model product.

Why it matters: whether an effect is large enough for a patient to notice is the question payers and prescribers actually ask.

Is there a catch? These are after-the-fact, unblinded re-analyses of an already-reported trial, not new randomised data. They reinterpret the existing effect rather than confirming it independently.

Reliability ●●●●● 4/5Why 4/5: Read from the published trial analysis itself. Held below our top score because it is a post hoc re-analysis of an earlier trial rather than a new randomised result. · Sources: JMIR Formative Research

Men using chatbot therapy described a consistently available, attentive relationship while doubting that any of it was authentic.

• Counselling Psychology Review published the study on September 22: an online survey of 69 men who had used chatbot-mediated therapy.
• Participants described a reparative relationship “characterized by consistent availability, responsiveness, and emotional attunement”.
• Many called the support “good enough”, and said it helped them speak freely while easing the fear of being judged.
• The same men raised concerns about authenticity, the impersonal feel of the exchanges, and the risk of building unrealistic expectations of relationships.
• The authors conclude the support “remains distinct from human interaction”, leaving open questions of depth and long-term impact.

Why it matters: this is the alliance question in first-person form, from the group least likely to walk into formal therapy.

Is there a catch? The 69 men chose to take part and reported looking back, with no outcome measured. The study does not say whether the chatbots were purpose-built therapy tools or general-purpose assistants.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed qualitative study confirmed at the publisher with its date and sample. A qualitative survey sits a rung below trial evidence on our scale. · Sources: Counselling Psychology Review

An addiction-clinic chatbot sorts messages into three risk tiers and hands every serious one to a person, by design.

• npj Digital Medicine published the description on September 22, from authors at Massachusetts General Hospital, Harvard Medical School, UCLA and Dimagi.
• Suzy supports adults on medication treatment for opioid use disorder in primary care addiction clinics at a large urban hospital.
• A fine-tunedA general language model further trained on examples from a specific task, here classifying the risk in a patient message, so that it performs that one job more reliably. classifier sorts messages in real time: tier 1 for explicit suicidality, tier 2 for ambiguous distress, tier 3 for minimal concern.
• A staff member reviewed conversations twice daily on weekdays and ran a standard suicide-risk scale whenever a tier 1 message appeared, with psychologists assessing further when needed.
• The authors recommend pairing built-in guardrails with human oversight and repeated expert testing as models change. A UK regulator testing AI in health services reached the same conclusion about disclaimers, which do not stop a system straying outside what it was meant to do.

Why it matters: it is the only published account this week of a crisis-escalation design actually running in a clinical service, and the model never escalates alone.

Is there a catch? This is a description of a design, not an outcome study. The only figure in it is 200 simulated messages reviewed by four expert raters, and it reports no detection performance, efficacy or patient outcomes. The authors say the design still needs validating beyond their pilot.

Reliability ●●●●● 3/5Why 3/5: A peer-reviewed comment piece read in full at the publisher, describing the authors’ own deployment. Held at 3 because it reports a design with no outcome or performance data to check it against. · Sources: npj Digital Medicine

A decade of building digital child mental health tools produced two working applications, and the limits were engagement, integration and money.

• The Journal of Child and Family Studies published the case history on September 22, from a Mayo Clinic team in psychiatry and paediatrics.
• The programme ran six rounds of gathering information from intended users, six prototypes and six rounds of pilot testing, with seven earlier publications behind it.
• The result was two working applications.
• It also produced a conclusion about direction: focus on primary-care tools that promote assessment and referral and support in-person care, rather than standalone apps.
• The barriers identified were participant engagement, integrating with hospital systems, and securing resources.

Why it matters: ten years of documented attempts put the ceiling on redesigning a care pathway at engagement, integration and money rather than at capability.

Is there a catch? This is one institution’s account of its own development history, with no comparison group. No AI or language-model component is described anywhere in it.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed original paper read at the publisher and covered independently, with its process and results taken from the paper. A self-reported single-institution case history holds it below our top score. · Sources: Journal of Child and Family Studies · Scienmag

A model trained to spot clinical reasoning in therapy supervision agreed reasonably with human coders, but systematically missed reasoning that was there.

• Psychotherapy published the feasibility study on September 21, analysing 631 supervision transcripts in which supervisors and therapists discussed youth cases with engagement problems.
• Human coders had labelled the transcripts across three categories: reasoning activities, content focus, and what the reasoning was about.
• A small transformer model beat a bag-of-words baseline on every measure, reaching moderate to substantial agreementCohen’s kappa above 0.58 against human ratings. Kappa measures agreement beyond chance, with higher values meaning closer agreement. It measures agreement with coders, not accuracy. with the human ratings.
• Statistical tests showed systematic underlabeling, meaning the model missed reasoning the human coders had marked.
• Agreement was better in one trial arm than the other, and the authors present the work as an early step rather than a usable measure of supervision quality.

Why it matters: supervision is the least-measured part of the mental-health workforce pipeline, and this is the only attempt this week to instrument it.

Is there a catch? Agreement with human coders is not accuracy, and the underlabeling means the model misses reasoning that was present. The transcripts come from one trial, so nothing here shows the approach transfers.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed study read at the publisher, with its sample and agreement figures taken from the paper. A feasibility design with no deployment holds it below our top score. · Sources: Psychotherapy

Google has given its live conversational model animated or photorealistic human faces that lip-sync whatever it says.

• Live Avatar arrived on September 24 for Gemini 3.8 Live, with preset characters and custom avatars built from a reference image.
• Google describes “precise lip-syncing, natural expressions, and fluid turn-taking”, across 97 languages.
• Availability is enterprise-only, through the Gemini Enterprise console, the API and the Vertex AI Agent Platform.
• Google frames the uses as customer service and interactive walkthroughs, and says all output is watermarkedGoogle says Live Avatar output carries SynthID, a marker embedded in generated audio and video so the content can later be identified as AI-made..

Why it matters: human likeness is being added to the general layer that companion and support products are built on top of.

Is there a catch? Google makes no claim about companionship, emotional support or wellbeing here, and offers no caution about human likeness. The feature is enterprise-only for customer service, so what is relevant is the mechanism and not a stated mental-health use.

Reliability ●●●●● 4/5Why 4/5: A first-party launch with a shipped feature, opened and dated at Google’s own blog and covered independently. The capability claims are Google’s own. · Sources: Google · The Register

Three general models scored clinician-run parent interviews for autism almost exactly as clinicians did, on a deliberately clean comparison.

• BMC Psychiatry published the study on September 26, from Nanjing Brain Hospital and Xuzhou Medical University in China with Seoul National University in South Korea.
• Gemini-2.5, GPT-5-mini and Qwen-Plus each scored a Chinese-language structured parent interview used to screen toddlers.
• In a cohort of 256 children, Gemini-2.5 reached the highest discriminationROC AUC 0.9961 (95% CI 0.9916 to 0.9991) for Gemini-2.5, 0.9861 for GPT-5-mini, 0.9770 for Qwen-Plus. AUC is the chance of ranking a true case above a non-case, so higher is better and chance-level performance sits at the midpoint. Gemini-2.5 also reached sensitivity 0.9692 and specificity 0.9206. of the three, with close agreement against clinician scoring.
• In a second, smaller cohort of 60 children, Gemini-2.5 classified 53 correctly, with accuracy 0.8833.
• The authors say prospective studies in larger, clinically mixed groups are needed before claiming real-world effectiveness.

Why it matters: a general model can reproduce clinician scoring of a structured form, and the size of the numbers shows how clean the comparison had to be.

Is there a catch? The design compares autistic children against typically developing children with no other neurodevelopmental group in between, which is what a near-perfect score is actually measuring. The authors note that discrimination, score agreement, repeatability and clinical usefulness are four separate claims.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed diagnostic study read at the publisher with its cohorts and figures taken from the paper. A retrospective case-control design holds it below our top score. · Sources: BMC Psychiatry

A review of wearable and phone sensing for perinatal mental health found ten studies and concluded the evidence does not yet support screening.

• JMIR Mental Health published the systematic review on September 25, from the University of Turku. Six databases yielded 1,952 unique records, of which 10 studies met the criteria.
• Those studies covered postpartum depression, prenatal stress, discrete emotions in pregnancy and maternal loneliness.
• Individual studies reported high discriminationA multiclass AUC of 0.85, a binary AUC of 0.871 and an F1-score of 0.9872 in individual included studies. F1 combines precision and recall into one figure from 0 to 1., with heart rate variability, location-derived mobility, activity and sleep as the informative signals.
• The review notes that perinatal mental health disorders affect about 20% of pregnant and postpartum people.
• Its own conclusion is that the current evidence does not support clinical screening or decision-making, and that independent validation in perinatal populations is missing.

Why it matters: individually high figures and an unusable evidence base turn out to be compatible, and the review shows how.

Is there a catch? Ten studies, assessed for bias and synthesised in words rather than pooled, cannot support a performance claim. The authors say so directly: no external validation in these populations exists yet.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed systematic review, preregistered, read at the publisher with its search and inclusion counts confirmed. A narrative synthesis of ten studies sits below trial evidence on our scale. · Sources: JMIR Mental Health

A review of digital depression treatments in low- and middle-income countries found 12 eligible studies, and not one of them used AI.

• BMC Psychiatry published the scoping review on September 24, reporting that a final set of 12 articles was eligible for synthesis.
• The interventions were a guided digital psychological programme, mobile phone approaches, home-based neuromodulation, internet-delivered cognitive behavioural therapy, and a virtual reality cognitive model.
• None was AI-based or chatbot-based.
• The authors call the current evidence limited and heterogeneous, and ask for well-powered trials with standard outcomes.
• They also ask for attention to digital access and literacy, and for cost-effectiveness work on affordability and scale.

Why it matters: AI mental health is argued for hardest on treatment-gap grounds, and the gap is widest exactly where this evidence base is thinnest.

Is there a catch? This is a scoping review of 12 varied studies with no pooled effect estimate, so it cannot say how well anything works. Its finding is about what has been studied, not about what is deployed.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed review read at the publisher with its inclusion count quoted directly, and covered independently. A scoping review sits below trial evidence on our scale. · Sources: BMC Psychiatry · Scienmag

Scoring a model for flattery may penalise the same language that makes advice worth taking, a new preprint argues.

• The preprint, posted September 22, says validation and positivity mark both flattery and receptivenessReceptiveness: a construct from social psychology for how openly someone engages with a view they may not share, measured through language that acknowledges the other side..
• On a moral-advice dataset, answers rated more flattering were also rated more receptive.
• Making human-written answers more receptive, without changing their conclusions, made them score as more flattering.
• In a preregisteredThe hypotheses and the analysis plan were filed publicly before the data were collected, which stops the authors picking a test after seeing the result. experiment, people preferred the more receptive answers and were readier to seek advice from their authors. That held even among people who thought the asker was in the wrong.
• The authors also describe a way to raise receptiveness without raising deference.

Why it matters: if the safety metric punishes the language that makes advice land, the scales now being written into mental-health benchmarks measure something other than harm.

Is there a catch? This is a preprint, not yet peer reviewed, and the abstract reports no correlation figures, sample sizes or benchmark names. The paper is about flattery scoring in general and not about mental-health use.

Reliability ●●●●● 4/5Why 4/5: The abstract page was opened and the argument confirmed there, including the preregistered experiment and the coupling result. It is not peer reviewed, which is the rung below our top score. · Sources: arXiv preprint

Meta’s AI glasses will run its Muse agent, and a pocket device is coming that gives the same agent a cartoon face.

• Meta’s own post, timestamped September 24, says its AI glasses will be powered by the personal agent from Muse.
• Meta separately announced the Muse Charm, a pocket-sized device with a live product page.
• Dezeen describes it on September 25 as an Apple Watch on a lanyard, with a two-inch screen, fingerprint sensor, 5G modem and small camera.
• A customisable cartoon character fills the whole screen. The choices include Jolly the Muse mascot, an anthropomorphic mango, a labradoodle and a piece of toast wearing sunglasses.
• It is set to launch before Christmas 2026. Zuckerberg says that without glasses it will be “by far the fastest way to talk to your Muse”.

Why it matters: a company with Meta’s reach is shipping an always-carried agent with a deliberately cute face, which is the surface companion-attachment research is about.

Is there a catch? Meta makes no claim about companionship, wellbeing, emotional support or mental health in any source that could be read, so the tie here is inferred rather than stated. Age-gating reported elsewhere could not be verified and is not claimed.

Reliability ●●●●● 4/5Why 4/5: The announcement was read at Meta’s own blog with its timestamp, and the device details come from independent coverage. The company’s site serves only page metadata to fetchers, so the product page itself was not read. · Sources: Meta

Seventeen therapists said their worry is not whether an AI can be validating, but whether it knows when validating is the wrong move.

• JMIR Mental Health published the interview study on September 25, with 17 Swedish psychologists who each have at least a year of clinical dialectical behaviour therapyDialectical behaviour therapy (DBT): a structured therapy for people with intense emotions and self-harm risk, best evidenced in borderline personality disorder. It deliberately balances accepting distress against pushing for change. experience.
• Three themes came out. Clinicians placed the agent variously as tool, team member, supervisor or harmful competitor, and where they placed it decided what it was allowed to do.
• A second theme was the agent as a tireless, always-available version of the ideal helper, mirroring qualities the clinicians found hard to sustain themselves.
• The third was the worry that it reinforces dependency, working as a reassurance habit rather than building autonomy.
• Benefits centred on access and on practising DBT skills; concerns were alliance, dependency and who is responsible when risk is high.

Why it matters: in this therapy a validating reply can be the clinically wrong one. A flattery metric scores the reply, not the timing.

Is there a catch? Participants did not test any specific AI agent, so these are anticipated views rather than evaluations of a deployed system. The sample is 17 clinicians in one country and one therapy.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed qualitative study read at the publisher, with its sample, method and themes confirmed. Interview work sits a rung below trial evidence on our scale. · Sources: JMIR Mental Health

Resilience Care has bought a Berlin psychiatry platform whose central asset is a contract with Germany’s largest statutory insurer.

• The acquisition was announced on September 24 by Resilience Care, described in the release as a French company founded in 2021, headquartered in Biarritz.
• Recovery Cat runs a digital platform for psychiatry, and says its therapy-monitoring tool is deployed in nearly 20 hospitals and practices.
• Its selective care contractA direct agreement with a statutory health insurer, here Techniker Krankenkasse, Germany’s largest. It makes the app free to that insurer’s members and lets providers bill for therapy monitoring. with Techniker Krankenkasse is what the buyer is acquiring.
• The deal value was not disclosed.

Why it matters: reimbursement, not capability, is what limits digital mental-health deployment in Europe, and this is a company buying a route through it.

Is there a catch? Neither company’s product is described as using AI or machine learning in the announcement, so this is a channel purchase rather than an AI development. No deal value is public.

Reliability ●●●●● 4/5Why 4/5: The first-party release was opened and its claims read verbatim there, with independent French coverage confirming the deal. The operating figures are the companies’ own. · Sources: Resilience Care · FrenchWeb

Across 19 studies, users valued AI for spiritual conversation precisely because it is not human, and told it things they told nobody.

• A scoping review posted September 21 searched ten databases for English-language studies from November 2022 to May 2026, and 19 met the criteria.
• Four patterns emerged: AI as spiritual authority, as a companion for disclosure, as a facilitator of ritual, and as a governance concern.
• Users often valued these systems “because they are not human”, and disclosed to them what they kept from clergy, family or therapists.
• The same users voiced scepticism about those systems’ authority and reliability.
• Most of the included sources raised concerns about bias, privacy, dependency and weak oversight.

Why it matters: the absence of a person is being described as the feature, in a disclosure setting next to therapy and outside any safeguarding frame.

Is there a catch? The authors say the evidence base is small, cross-sectional and concentrated in institutional religious settings. The review synthesises 19 studies and is not yet peer reviewed.

Reliability ●●●●● 4/5Why 4/5: The preprint was opened and its search strategy, inclusion count and quoted findings confirmed there. Not peer reviewed, so it sits below our top score. · Sources: preprint (DOI)

Twenty hospital physicians welcomed AI taking documentation and data work, and named de-skilling and lost room for judgment as their central worries.

• JMIR AI published the study on September 23: interviews with 20 physicians across different Dutch hospital types and specialties, plus two focus groups at a national symposium.
• Physicians described craftsmanship as delivering the best possible care through human judgment, empathy and understanding of context.
• They saw AI supporting workflow, documentation, data integration and parts of analytical reasoning.
• They made adoption conditional on human oversight, explainability, traceability and AI literacy.
• Empathy, interpreting context and ethical judgment were seen as inherently human. The concerns were de-skilling, less room for independent judgment, and threats to professional autonomy.

Why it matters: the tasks these physicians are happiest to hand over are exactly the ones the US veterans’ health system just contracted out.

Is there a catch? These are 20 physicians’ reported perceptions rather than measured skill loss. The sample is hospital physicians across specialties rather than mental-health clinicians, so it is evidence for the mechanism from a neighbouring profession.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed qualitative study read at the publisher, with its two-phase design and themes taken from the paper. Interview work sits below trial evidence on our scale. · Sources: JMIR AI

Among 685 Chinese adolescents, anxiety about parental attachment tracked more honest disclosure to generative AI, through rumination and depression.

• Health Communication published the study on September 23. It surveyed 685 adolescents offline, 324 of them girls, with a mean age of 13.17.
• Attachment anxiety toward both father and mother was linked to more ruminationRumination: repeatedly turning the same distressing thought over without resolving it. It is one of the better-established risk factors for depression., which was in turn linked to more depression.
• Both rumination and depression were linked to more honest self-disclosure to generative AI.
• Those two together accounted for the path from attachment anxiety to disclosure.
• The link from rumination to depression was significantly stronger among girls than boys.

Why it matters: it connects the state of a teenager’s human attachments to how much they tell a model. The age band is the one this week’s policy items are about.

Is there a catch? This is one survey at one moment, so the direction of these relationships is not established. Disclosure was self-reported, not observed.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed study read at the publisher, with its sample and mediation findings taken from the paper. A single cross-sectional survey holds it below our top score. · Sources: Health Communication

Reading what people in addiction recovery said about chatbots produced seven kinds of harm, two of which no safety rubric this week covers.

• The preprint, posted September 23, analysed user-reported chatbot interactions across 20 online substance use disorder recovery communities.
• The seven kinds are failing to recognise or escalate a crisis, helping someone return to use, claiming clinical authority it does not have, and giving inaccurate guidance.
• The remaining three are undermining non-professional recovery support, undermining a person’s recovery identity, and inappropriate emotional attachment or dependency.
• Two of those, undermining non-professional support and undermining recovery identity, appear in none of the safety rubrics published this week.

Why it matters: it is the week’s only harm account built from what users said rather than from a benchmark. Two of its categories appear in no benchmark.

Is there a catch? The posting is not peer reviewed and its abstract reports no post or participant counts, no sampling frame and no coding procedure. The data are self-selected public forum posts, so the taxonomy describes the shape of reported harms and cannot say how often any occurs.

Reliability ●●●●● 3/5Why 3/5: A preprint posting read at its deposited record, with the seven domains confirmed there. Held at 3: no peer review, an unstructured abstract and no counts, sampling frame or reliability procedure. · Sources: preprint (DOI)

Across four countries, 66% of chatbot users said they had discussed a sensitive topic with an AI assistant, mental health among the top three.

• Proton published the survey on September 24: 4,014 current chatbot users across France, Germany, the UK and the US, asked about 12 questions online.
• Personal finances, mental health and career problems were the three most-discussed sensitive areas. Mental health topped the list in Germany.
• Only 11% to 20% reported high trust in AI companies to protect that information, while 42% had little or no trust. Distrust was commonest in France at 47.7%.
• A guarantee that conversations would not be used for training would make 44% more likely to share sensitive information.
• UK respondents cited freedom from judgement, over convenience, as the reason for choosing a chatbot for mental health.

Why it matters: this is the gap between what people tell these systems and what they believe happens to it afterwards.

Is there a catch? The survey was commissioned by a privacy company that sells a competing private assistant. It observes no real conversations, does not name the fieldwork provider, and discloses no fielding date, weighting or margin of error.

Reliability ●●●●● 3/5Why 3/5: The published figures were read at the company’s own post and echoed independently. Held at 3: a self-report survey commissioned by a firm with a commercial interest in the answer, with no methodology disclosed. · Sources: Proton · TechRadar

A framework sets out twenty-two AI competencies for behavioral health graduate training, at a point when the tools are already in student placements.

• Mental Health and Digital Technologies published the framework on September 23, from West Virginia University’s School of Medicine.
• It is a narrative synthesis of the existing literature rather than new data.
• The authors grouped that literature into twenty-two AI competencies recommended for practitioner education.
• The literature pointed to several good moments across a student’s training to teach this, and to several ways of checking what was learned.
• The authors report the literature is fragmented and lacks common guidance, and that teaching on ethical, legal and safe practice remains scarce.

Why it matters: it is the clearest statement this week of what the next cohort of clinicians would have to be taught about the tools already around them.

Is there a catch? The twenty-two competencies are the authors’ own grouping of a fragmented literature, not a validated or consensus-endorsed set. Nothing here has been tested in a curriculum.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed article read at the publisher, with its method and competency count taken from the paper. A narrative synthesis with no primary data sits below our top score. · Sources: Mental Health and Digital Technologies

Trying to predict which mildly depressed patients worsen, ordinary logistic regression beat every fancier method, and the neural network lost to guessing.

• Scientific Reports published the study on September 24, using routine health records for 803 people with mild depression and two years of continuous follow-up.
• The records come from MedStar Health, a mid-Atlantic network of 10 hospitals and over 300 connected points of care.
• Four approaches were compared: logistic regression, random forest, gradient boosted trees and a deep neural network.
• Logistic regression discriminated best and generalised best. The deep neural network was less accurate than a null classifierA model that ignores the input and always predicts the commoner outcome. Anything performing worse than this has learned nothing useful from the data..
• The authors describe the improvement over simply knowing how common worsening is as modest.

Why it matters: the assumption that a bigger model finds more signal in a medical record did not hold here.

Is there a catch? The authors call the gain over base rates modest, and no discrimination figures are published on the article page, so no accuracy number can be quoted here. One health network and 803 patients is a narrow base.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed study read at the publisher, with its cohort, methods and comparative result taken from the paper. A retrospective single-network cohort holds it below our top score. · Sources: Scientific Reports

An automated therapy agent beat an electronic leaflet on depression symptoms in international medical students over two weeks.

• The pilot trial, published September 22 at Medical University Sofia, screened 228 students and randomised 138, 69 to each arm.
• The agent delivered cognitive restructuring, behavioural activation, mood monitoring, psychoeducation, emotional reflection and mindfulness. Retention at two weeks was 74.6%.
• Depression scores fell in the agent arm and rose slightly in the control arm, a large differencePHQ-9 moved from 15.39 (SD 3.70) to 13.84 (SD 3.49) with the agent, and from 17.47 (SD 3.84) to 17.78 (SD 5.06) with the leaflet; Cohen d = 0.84..
• Anxiety scores showed a moderate differenceGAD-7 moved from 14.41 (SD 3.33) to 12.62 (SD 3.17) with the agent, and from 14.53 (SD 3.43) to 14.80 (SD 4.13) with the leaflet; Cohen d = 0.57. in the same direction.

Why it matters: it is one of few randomised tests this week of a genuinely AI-enabled therapy agent rather than a scripted one.

Is there a catch? Two weeks against a passive leaflet is a weak comparison. The control arm started about two depression points worse and then got worse still, which inflates the gap. The agent is unnamed and unbranded, so these are preliminary feasibility estimates.

Reliability ●●●●● 4/5Why 4/5: A randomised trial read from the published paper with its screening, randomisation and score figures. Held below our top score: a two-week pilot feasibility design against a passive comparator. · Sources: Psychiatry International · full text

The measures used to claim a chatbot relationship works were built for human therapy, and a new review examines whether they survive the move.

• The International Journal of Human-Computer Interaction published the review online on September 23.
• Its two authors are at Shahid Beheshti University and Islamic Azad University.
• The subject is the transport of therapeutic allianceTherapeutic alliance: the working bond between a client and a therapist, covering agreement on goals, agreement on tasks, and the emotional tie. It is measured by questionnaires designed for human therapy. measures into mental health chatbot research, and whether measuring it that way is valid.
• Nothing more can be said: the abstract could not be retrieved from any source.

Why it matters: alliance scores are the field’s main instrument for claiming a chatbot relationship works, and this is the week’s only item asking whether the instrument holds.

Is there a catch? With no abstract available, the direction and strength of the conclusion, the number of studies screened and included, and any synthesis are all unverified. No finding may be attributed to this item.

Reliability ●●●●● 4/5Why 4/5: The title, venue, both authors and the online publication date were confirmed at the publisher record. No abstract exists there or anywhere else, so the primary is confirmable but thin. · Sources: journal (DOI)

laya healthcare extended round-the-clock human support to members from 16 and added digital therapy from age six, alongside survey findings on AI use.

• The announcement of September 22 paired Irish findings from a parent-company survey with a change to the insurer’s own service.
• Round-the-clock mental wellbeing support now covers all plans for members aged 16 and over, at no extra cost. Digital therapy is offered from age six for the first time.
• The survey figures come from fieldwork among 19,000 adults across 18 countries, including 1,000 in Ireland, conducted in January and February 2026.
• As reported for Ireland, 61% have used AI platforms for mental health support against 45% who saw a professional in the past year.
• Among Irish AI users, 42% say they almost always follow AI guidance without question, 39% felt uneasy afterwards, and a third say AI advice led to harmful behaviour.

Why it matters: an insurer is restructuring its own service tiers around a shift to AI as the first door. Human support now extends down to age six.

Is there a catch? The 61% figure is identical to the global figure the parent company published in June, and the UK edition reported 46%. It is not a new Irish measurement. The fieldwork predates this week by eight months, the harmful-behaviour item is an unvalidated self-assessment, and no first-party release or methodology document could be opened.

Reliability ●●●●● 3/5Why 3/5: Two press reports of a first-party service change plus previously published survey figures. No company release or methodology document could be opened, and one outlet was blocked, which holds it at 3. · Sources: Irish Tech News · RTE

Three AI companies describe their systems in human terms in three recurring ways, a content analysis of their own documents finds.

• AI & Society published the analysis on September 21, by a single author at the University of Münster, covering 35 documents from OpenAI, Meta and Mistral AI.
• The three recurring framings are AI as a tool with human capabilities, AI as an empowering helper, and AI as a human-like superintelligence.
• Those imply three kinds of power the author names instrumental, collaborative and autonomous.
• The argument is that these framings take politics out of AI, cast the companies as authoritative creators, and encourage dependence.

Why it matters: human likeness is usually studied as something users do, and this documents it being manufactured on the supply side.

Is there a catch? The 35 documents are company-authored material, so the paper evidences how these firms write about themselves. It cannot show that users read the systems as persons.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed qualitative analysis read at the publisher, with its corpus and findings taken from the paper. A single-author content analysis sits below trial evidence on our scale. · Sources: AI & Society

Model performance, deployment and engagement figures establish nothing about whether practice changed, and a proposal sets out what would.

• The Journal of Medical Internet Research published the viewpoint on September 23, from authors at Wuhan Asia Heart Hospital and Wuhan University of Science and Technology.
• The proposal starts from a defined AI output and records the chance to act, who was accountable, whether they acted, and the evidence about the AI’s role.
• It sets rules for repeated outputs, completed action, justified inaction, missed action, unresolved cases and lost denominators.
• It pairs those with monitoring of safety, workload, privacy, equity and model driftModel drift: the gradual decay of a model’s accuracy as the world it was trained on changes, so that performance measured at launch no longer describes performance in use..
• The authors state their own limits: it is not a causal outcome, a reporting guideline, or a stand-in for population health benefit. An action rate alone does not show AI caused the action.

Why it matters: it is a concrete alternative to the metric every screening item this week reports. The gap between a benchmark score and a defensible claim about care widened this week.

Is there a catch? It addresses AI in public health generally and is not specific to mental health. It is a proposal for an evaluation standard with no empirical validation behind it.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed viewpoint read at the publisher, with the proposal and the authors’ own stated limits taken from the paper. An unvalidated proposal sits below trial evidence on our scale. · Sources: Journal of Medical Internet Research

Adults with serious mental illness said they trusted AI tools more when people with their own experience had helped design them.

• Community Mental Health Journal published the study on September 25, a secondary analysis of three focus groups across three supportive housing facilities in New York City.
• Participants had schizophrenia spectrum disorder, bipolar disorder or major depressive disorder.
• They raised strong concerns about data privacy and misuse, while recognising AI could help with reminders, mental health and daily functioning.
• Concerns remained about losing critical thinking and about over-reliance.
• Trust rose when tools were personalised, human-centred and co-developed with people who understand living with serious mental illness.

Why it matters: this is the population the access argument is usually made on behalf of. The condition they name for trust is taking part in the design.

Is there a catch? Three focus groups in a secondary qualitative analysis give no prevalence estimate and no comparison group. Nothing here shows that AI companions reduce loneliness.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed qualitative study read at the publisher and covered independently, with its design and themes taken from the paper. A secondary analysis of three focus groups sits below trial evidence on our scale. · Sources: Community Mental Health Journal · Scienmag

A speech-based depression screening system scored in the seventies once it had to work across corpora rather than inside one.

• MDPI Computers published MindVoice on September 23, from authors at Amity University in Noida, India and Lusófona University in Porto.
• Three separate speech datasets were combined into one mixed, cross-lingual three-class task, with a fourth dataset evaluated separately.
• On the combined corpus76.5% accuracy on the combined three-class dataset; 81.4%, 72% and 80% on the three constituent corpora individually, and 78.5% on a fourth dataset for four-class classification. the system reached 76.5% accuracy, against higher figures on the individual datasets.
• The system is deployed as a web tool with a visual explanation of what drove each prediction.
• The authors frame it as a screening aid rather than a diagnostic.

Why it matters: this is the size of the generalisation gap in one task, in one week. A single-corpus model published two days later reports far higher figures.

Is there a catch? The venue is a general computer-science journal rather than a clinical one, which is a real limit on a screening claim. Nothing here was tested with patients in a care setting.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed article read at the publisher, with every accuracy figure taken from the paper. A computer-science venue with no clinical evaluation holds it below our top score. · Sources: Computers (MDPI)

Across three countries, openness to AI in mental health drops sharply once diagnosis, treatment or prescribing is involved.

• The survey, published September 25, polled 1,509 adults across the US, UK and Australia, fielded on 9 September 2026.
• About 29% cite cost as a barrier to professional support, and about 27% say they prefer to handle things on their own.
• In the UK 19% point to long wait times, nearly triple the US rate of 7%.
• Nearly one in five adults currently use an AI companion or chatbot for mental health.
• Openness is highest for general advice and self-guided support. For diagnosis, treatment or prescribing, 54% to 74% still prefer a licensed professional depending on the market.

Why it matters: public consent is broad at the advice layer and narrow at the decision layer. This week’s triage deployments sit at the decision layer.

Is there a catch? This is self-serve panel research published on a company blog, with no weighting or margin of error disclosed and an explicitly commercial framing. It observes attitudes at one moment, not behaviour.

Reliability ●●●●● 3/5Why 3/5: Figures read at the publisher’s own post with the fielding date and sample sizes stated. Held at 3: self-serve panel market research on a vendor blog, with no weighting or margin of error disclosed. · Sources: SurveyMonkey

Six studies exist on algorithm-assisted monitoring in psychiatric wards, and not one measured the nurse-patient relationship it is meant to affect.

• The Journal of Psychiatric and Mental Health Nursing published the scoping review on September 21, searching five databases for studies from 2010 to 2025.
• Six studies were included, all from the United Kingdom. Five looked at non-contact vision-based systems and one at wearable physiological sensing before it was installed.
• Nurses read the outputs inside their observation routines, adding explanation, reassurance, in-person checking and boundary-setting.
• No included study used a validated measure of therapeutic alliance or of the relationship. What was reported instead was acceptability, privacy, dignity and the sense of being watched.
• The authors conclude monitoring should stay an add-on, with rules covering explanation, verification and privacy boundaries.

Why it matters: continuous monitoring is already installed on wards, and whether it changes the relationship that ward care depends on has not been measured once.

Is there a catch? The whole evidence base is six UK studies, and none measured the relational outcome the review set out to examine. So monitoring is unmeasured in both directions here: neither preserving nor damaging that relationship is shown.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed review read at the publisher, with a preregistered protocol, duplicate screening and its inclusion count confirmed. A scoping review of six studies sits below trial evidence on our scale. · Sources: Journal of Psychiatric and Mental Health Nursing

Twenty older adults using a voice-based memory companion reported better mood and more sense of connection right after a session.

• The preprint, posted September 21, describes MeBo, designed with 11 older adults and then evaluated with 20 more.
• Participants rated it highly usable, at 87.75 on the System Usability ScaleSystem Usability Scale (SUS): a ten-question usability survey. Higher means easier to use, and a score of 87.75 sits in the range usually read as excellent..
• After the session they reported higher positive mood and more momentary social connection, and lower negative mood, than before it.
• The companion is built around recalling and retelling the user’s own memories.

Why it matters: it is the week’s only built-and-tested companion for older adults, and it reports the mood gain the category usually just asserts.

Is there a catch? There was no control group, and the mood and connection results are single-session self-report from 20 people before and after. The study cannot show any effect that outlasts the session.

Reliability ●●●●● 4/5Why 4/5: The preprint was opened and its design, sample sizes and usability figure confirmed there, dated in window. Not peer reviewed, which is the rung below our top score. · Sources: arXiv preprint

Lifting US or EU AI frameworks wholesale into low-resource health systems loads them with compliance work they cannot absorb, a group of authors argue.

• npj Digital Medicine published the perspective on September 21, from authors at Erasmus MC Rotterdam, Toronto, Stellenbosch, Delft and Duke.
• The proposal is regionally anchored networks tailored to local context, with a primary-care-first approach and local governance so decisions are context-driven.
• It adds data sovereigntyData sovereignty: the principle that data about a population is governed under that population’s own laws and institutions, rather than held and ruled elsewhere. to build trust and protect privacy, plus community-driven infrastructure and capacity building.
• The authors describe the conditions these systems actually face: nascent regulation, fragmented data, under-represented languages, intermittent connectivity and power, and workflows centred on primary care.

Why it matters: this week’s two new screening awards are the kind of project this argues has to be governed locally. The argument is about institutions, not models.

Is there a catch? This is a normative perspective with no new data. It addresses clinical AI generally and contains no mental-health content, with examples drawn from diagnosis, surveillance, telemonitoring and screening.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed perspective read at the publisher, with its proposals taken from the paper. A normative argument with no new data sits below trial evidence on our scale. · Sources: npj Digital Medicine

Vitality is taking its AI engagement platform to US employer plans, citing 4% lower average claims costs and a 180% return.

• A report of September 22 covers the expansion ahead of 2027 renewals, for a platform built on Google’s Gemini technology.
• Vitality says engaged clients recorded claims costs 4% lower on average, a 180% return, and that members at higher participation levels had 15% lower risk-adjusted claims costs.
• The first two figures were reviewed by one outside body and the third endorsed by another, but the underlying data comes from Vitality.
• The platform pairs Gemini with a dataset spanning more than 2,800 health and behavioral dimensions. Since June 2025 more than 210,000 people have received an AI-driven personalised action, and 54% completed at least one.
• The one mental-health-specific figure: members receiving personalised recommendations were 3.3 times likelier to complete a mental wellbeing assessment.

Why it matters: it is the only item this week attaching a cost-of-care number to an AI behavioural deployment.

Is there a catch? This is broad health engagement rather than mental health care, and its only mental-health datapoint is completion of an assessment. The claims and return figures are the company’s own data reviewed rather than independently generated, with no methodology published.

Reliability ●●●●● 3/5Why 3/5: A single trade report of a company expansion, which itself states that the underlying data comes from the company. No first-party release or methodology is available, which holds it at 3. · Sources: Human Resources Director

A framework tracks emotion through crisis-counselling conversations at three levels, and counts the volunteer’s trajectory as part of the measurement.

• The preprint, submitted September 24, works at three levels: labels per turn, the probability of moving between states, and whole-conversation shapes.
• It was applied to text-based crisis conversations with self-identified Black texters discussing grief.
• It reports persistent negative feeling, gradual movement toward hope, distinct emotional roles for texter and volunteer, and varied recovery paths.
• The corpus is described only as text-based crisis conversations and is not attributed to any named service.

Why it matters: crisis-line analytics rarely treat the volunteer’s own trajectory as data, and this one does.

Is there a catch? The abstract reports no sample size, no corpus provenance and no evaluation against a ground truth, so these are descriptive patterns rather than validated measurement. The preprint is not peer reviewed.

Reliability ●●●●● 3/5Why 3/5: The abstract page was opened and the framework and reported patterns confirmed there. Held at 3: no sample size, no corpus provenance and no evaluation metric, with no peer review. · Sources: arXiv preprint

Research on human-machine bonds uses the word attachment without the evidence the word carries elsewhere, a new preprint argues.

• The preprint, posted September 24, says studies “frequently produc[e] expansive relational claims that far outstrip their underlying evidence”.
• It proposes a four-question procedure that sorts a human-machine tie into attachment, an affective social bond, or no relationship.
• The questions are grounded in criteria from developmental, ethological and neuroendocrineConcerning the hormones that accompany social bonding, such as oxytocin and cortisol. Attachment research in animals and infants uses these measures alongside behaviour. research, each with a stated minimum of evidence.
• Intermediate verdicts are allowed where the evidence is incomplete.
• The authors run the procedure over a set of published human-robot studies as a demonstration.

Why it matters: attachment is the word doing most of the work in claims about companion harm, and this sets out what would have to be shown first.

Is there a catch? The preprint adds no new empirical data of its own and is not peer reviewed. Its subject is human-robot pairs, so applying the criteria to text-based companion chatbots is an extension the paper does not make.

Reliability ●●●●● 4/5Why 4/5: The abstract page was opened and the argument and procedure confirmed there, including the quoted claim. Not peer reviewed, so it sits below our top score. · Sources: arXiv preprint

Conversational AI for older adults should aim to widen what they can do rather than stand in for social contact, two researchers argue.

• Frontiers in Psychology published the analysis on September 22, from authors at Wuhan Sports University and Central China Normal University.
• The argument is that research has read conversational AI mainly through companionship, and should read it through capability instead.
• Its framework pairs what the system offers, such as responsiveness and availability, with personal, social, environmental and system-level conditions that decide whether the offer converts.
• Under adverse conditions the same system “may contribute to capability erosion through over-reliance, cognitive offloading, reduced human contact, or diminished autonomous judgment”.

Why it matters: it names an alternative design goal for the group most often offered companionship, and the conditions under which the same tool does the opposite.

Is there a catch? This is a conceptual analysis with no new data. The framework is untested in either direction, so neither capability building nor capability erosion is demonstrated here.

Reliability ●●●●● 4/5Why 4/5: Peer-reviewed article confirmed at the publisher with its date, authors and affiliations. A conceptual analysis rather than a trial, which is the rung below our top score. · Sources: Frontiers in Psychology

Rula Health has created an external council of clinicians and academics to review its AI behavioural-health products every quarter.

• The council was announced on September 22, to advise on clinical safety, AI ethics, informed consent, outcome measurement and emerging standards.
• Four members are named: Tanzeem Choudhury of Cornell Tech, Dori Steinberg of Steinberg Research Consulting, Camille Nebeker of UC San Diego, and Nick Jacobson of Dartmouth’s Geisel School of Medicine.
• Chief Medical Officer Doug Newton is quoted saying the future of mental health care “should be AI and therapy, not AI or therapy.”
• Rula describes a network of more than 22,000 licensed providers, with partners covering more than 200 million lives across the US.

Why it matters: a provider is installing its own AI safety review at the moment other vendors are writing the standards such a review would apply.

Is there a catch? The release describes no charter, no promise to publish findings, and no power to block a product decision. No independent coverage of the council exists, so its effect on Rula’s AI products cannot be checked from outside.

Reliability ●●●●● 3/5Why 3/5: The release was opened and every detail confirmed in it, including all four members. Held at 3 because the announced thing is a governance body rather than an inspectable shipped product. · Sources: Rula Health

Companion providers owe users a managed exit when the companion is withdrawn, under consumer law that already exists, a legal paper argues.

• The preprint, posted September 23 by Christophe Boeraeve, is listed as under review at the Journal of Consumer Policy.
• It builds on EU consumer and data-protection rules, among them the Digital Content DirectiveDigital Content Directive: EU rules on digital content and services sold to consumers, covering what a trader owes a buyer when a digital service changes or stops., the Data Act and the GDPR.
• The argument is that policy should protect what a user can still do after an ending, without promising that the companion continues.
• It proposes separating generation, records, transactions and recourse, and assigning duties by who ended the relationship.
• It rejects compulsory farewells, requiring diagnostic proof of attachment, and keeping models alive indefinitely.

Why it matters: companion shutdowns are now a recurring event, and this is the first attempt to find the remedy in law that already exists.

Is there a catch? This is a single-author legal argument with no empirical data, and it is not peer reviewed. It reasons from EU law only, so it says nothing about users elsewhere.

Reliability ●●●●● 3/5Why 3/5: The posting date was confirmed at three independent records after an earlier misreading, and the argument was read in full. A single-authored preprint with no data caps at 3. · Sources: preprint (DOI)

A three-month automated text-message treatment for anxiety produced a medium reduction in depressive symptoms in young adults.

• The trial, published September 23, tested CBT-txt-A in 102 young adults recruited from social media.
• Depression symptoms fell by a medium amountCohen d = .61 on the PHQ-8, a short self-report depression scale. Depression was a secondary outcome; the trial targeted generalized anxiety disorder. at three months.
• At three months, 27% of the treatment group had no or minimal depressive symptoms and 33% had mild symptoms.
• The mechanisms linked to that fall were more behavioural activation, less repetitive worrying, and fewer distorted thoughts.

Why it matters: this is the bar an AI-delivered equivalent has to clear: a manual-following automated programme with a medium effect and a named mechanism.

Is there a catch? Depression was a secondary outcome of a trial aimed at anxiety, in 102 young adults recruited from social media. No AI or language-model component is established anywhere in the paper, and it does not say whether the messages are pre-written or generated.

Reliability ●●●●● 4/5Why 4/5: A randomised trial read from the published paper, with independent coverage. Held below our top score because the depression result is a secondary outcome rather than the trial’s target. · Sources: journal (DOI) · Scienmag

CentralReach has released a tool that assembles authorisation and progress data into draft reports for behaviour analysts.

• The company’s September 22 release says ReportDraftAI has moved into general availability, producing a “structured, review-ready draft”.
• The users are Board Certified Behavior Analysts serving people with autism and with intellectual and developmental disabilities.
• CentralReach says reauthorization reportingThe paperwork a therapy provider files to get a payer to keep funding a course of care. Much of it is unbilled work, and a refusal ends or reduces the funded therapy. typically takes 9 to 12 hours, much of it non-billable, but gives no time-saving figure for this product.
• It says an earlier tool has assisted with 5.8 million session note summaries and saves technicians an average of 8 hours a week on note writing.
• The company describes itself as trusted by more than 200,000 professionals globally.

Why it matters: the document that decides whether funding for a child’s care continues is now being drafted by a model.

Is there a catch? Every figure is the company’s own and none of them measures this product. No accuracy, error-rate or clinician-review data is reported anywhere.

Reliability ●●●●● 3/5Why 3/5: A first-party release confirmed at the wire and echoed by trade coverage, but the announced tool has no independently inspectable artifact and every figure is self-reported, which holds it at 3. · Sources: CentralReach · Health Technology Net

AI systems modelling human behaviour are built on what people say about themselves, and psychiatry’s evidence on how unreliable that is goes unused.

• A BJPsych Bulletin piece, published September 24, argues “the assumption that this data is reliable remains largely unexamined”.
• The author points to distorted thinking, the gap between how someone presents and what they are experiencing, and conditions that make accurate self-report impossible.
• The piece also holds that “the populations most affected by AI-driven decisions can be the least represented in training data”.
• Its conclusion is that this expertise is largely absent from AI development, that the omission is fixable, and that “psychiatry must enter the conversation”.

Why it matters: it names a specific technical contribution the profession could make, rather than a general claim to be consulted.

Is there a catch? It presents no new data and stakes a claim about professional role rather than reporting a finding. It is a single-author editorial.

Reliability ●●●●● 3/5Why 3/5: A peer-reviewed journal editorial read at the publisher with its arguments quoted directly. Held at 3: a single-author opinion piece with no data behind it. · Sources: BJPsych Bulletin

Mainstream psychology’s own literature now carries a study of what people want from AI companionship and what they think they gain.

• Technology, Mind, and Behavior, an American Psychological Association journal, published the article online on September 24.
• It is titled “Digital intimacy: Motivations, perceptions, and perceived growth in human-artificial intelligence companionships”, by three authors.
• The subject is motivations, perceptions and perceived growth in human-AI companionship.
• No method, sample size, design or finding could be retrieved from any source.

Why it matters: a flagship psychology venue publishing on companionship marks the topic’s arrival in that field’s own literature.

Is there a catch? The publisher platform refused every fetch, so only the existence, venue, date and title are confirmed. Nothing about what the study did or found is verified, and no figure or finding may be attributed to it.

Reliability ●●●●● 4/5Why 4/5: Existence, venue, title, authorship and publication date were confirmed at the publisher deposit for a peer-reviewed article. No abstract or full text was reachable, so the content is unverified. · Sources: journal (DOI)

Presence, participation, discernment and comfort with uncertainty are the physician skills that survive as AI takes over retrieval and protocol, a correspondence argues.

• Academic Psychiatry published the correspondence on September 24, by a single author at Mayo Clinic.
• The argument is that four relational capacities built in psychiatric training endure while AI absorbs “fact retrieval, protocol recitation, and drug-interaction flagging”.
• It states that “Medical training has long treated knowledge as the central currency of doctoring”.
• It argues “a reflective self-awareness about patient interactions should be considered as important as diagnostic accuracy and pharmacology”.
• Its ask is curriculum reform that measures and rewards relational capacities alongside technical knowledge.

Why it matters: stated as a curriculum requirement, the therapeutic-identity claim becomes actionable rather than rhetorical.

Is there a catch? This is an argument, not a study: no data, no participants and no evaluation. It is one author writing in a correspondence section.

Reliability ●●●●● 3/5Why 3/5: A peer-reviewed correspondence read at the publisher and covered independently, with its claims quoted directly. Held at 3: a single-author argument with no data. · Sources: Academic Psychiatry · Scienmag

A peer-reviewed psychiatry journal has published a call for India to regulate AI companion chatbots for adolescent safety.

• The International Journal of Psychiatry in Medicine published the piece online on September 24.
• Its two authors are affiliated with CHRIST (Deemed to Be University) and Colorado State University.
• India has no rule specific to companion chatbots.
• What the piece proposes is not known: no abstract was retrievable from the publisher or anywhere else.

Why it matters: a clinical journal arguing in print for a rule is the professional route into policy, and the rule it asks for does not exist yet.

Is there a catch? With no abstract available, the measures proposed and any evidence or prevalence figures cited are unverified. No finding can be attributed to this item.

Reliability ●●●●● 3/5Why 3/5: Title, venue, both authors and the publication date were confirmed at the publisher’s deposit. Held at 3: it is a call to regulate rather than a research article, and with no abstract in hand the primary is thin. · Sources: journal (DOI)

A peer-reviewed study asked online peer counsellors what they would want from AI-simulated clients used to train them.

• The paper appears in the Proceedings of the 2026 ACM Conference on Human-AI Complementarity and Alignment.
• Its authors are at Carnegie Mellon University’s Human-Computer Interaction Institute and School of Computer Science, and the University of Pittsburgh Department of Psychiatry.
• The subject is peer counsellors’ own preferences for AI simulation-based communication training.
• Nothing more can be said: no abstract was deposited anywhere, and the publisher page does not render for a fetcher.

Why it matters: simulated clients are a live proposal for training peer counsellors, and asking the counsellors what they want is rare in this literature.

Is there a catch? The claim is confirmed at title, venue, authorship and date level only, so no findings, sample size or design detail could be read. The date itself is contested: the publisher says September 27 while the deposit records say September 16.

Reliability ●●●●● 3/5Why 3/5: Title, venue, authorship and peer-reviewed status confirmed at the deposit, but no abstract exists anywhere and the publisher page cannot be fetched, so the content is unverified and the date is contested. · Sources: conference paper (DOI)

AI in mental health has moved from assessment tools to companions that remember you, a review of the literature argues.

• The survey preprint, posted September 21, sorts the literature into three phases. They run from pattern recognisers for assessment, through stateless empathetic conversation, to long-term personalised companions.
• It puts the current frontier at the third phase, built as agents that keep a profile, a memory, reasoning and planning.
• Across the papers collected, psychotherapy leads as the application at 58%, followed by diagnosis at 25%, education at 9% and assistance at 8%.
• Papers came from the ACL Anthology, PubMed, arXiv, Web of Science and other scholarly sources.

Why it matters: it is the week’s only attempt to name where the whole field is heading, and it points at companions rather than assessment.

Is there a catch? This is a narrative review with no original data and no peer review. It states no inclusion protocol and no screening counts, so the three phases are the authors’ way of organising the field rather than a measured finding.

Reliability ●●●●● 3/5Why 3/5: The preprint was opened and its submission date confirmed in window, and the three-phase claim is exactly what the paper asserts. Held at 3: a survey with no experiments and no peer review. · Sources: arXiv preprint