The source library.
69 entries covering the trials, the meta-analyses, the harms literature, the regulation and the critiques. Credit belongs to the researchers and organizations behind each work. This library includes positive, mixed and critical findings.
References use an author-date format with the published title, journal or source, and DOI or original link. Researcher names are retained in full where verified. The summaries and charts are our own; source figures are not reproduced. Inclusion does not imply that an author endorses Moonbridge. Some publishers require a subscription or browser verification.
Randomized trials of digital support
Controlled studies of symptoms, engagement and guided support.
-
Michael V. Heinz; Daniel M. Mackin; Brianna M. Trudeau; Sukanya Bhattacharya; Yinzhou Wang; Haley A. Banta; Abi D. Jewett; Abigail J. Salzhauer; Tess Z. Griffin; Nicholas C. Jacobson (2025).
Randomized Trial of a Generative AI Chatbot for Mental Health Treatment
NEJM AI, 2(4).
https://doi.org/10.1056/aioa2400802
The first randomized controlled trial of a fully generative therapy companion. 210 adults; depression, anxiety and eating concerns all improved substantially against a waitlist, and gains were larger at eight weeks than at four.
The study system was Therabot. Dartmouth-funded and run by the intervention’s own developers; the control was a waitlist rather than live therapy. NEJM AI published two critical letters alongside it. This developer-run trial does not establish effectiveness in routine clinical care.
-
Richard A. Bryant; Anne M. de Graaff; Rand Habashneh; Sarah Fanatseh; Dharani Keyan; Aemal Akhtar; Adnan Abualhaija; Muhannad Faroun; Ibrahim Said Aqel; Latefa Dardas; Hadeel Afar; Chiara Servili; Dusan Hadzi-Pavlovic; Mark van Ommeren; Kenneth Carswell (2026).
npj Digital Medicine, 9(1), Article 57.
https://doi.org/10.1038/s41746-025-02142-8
A chatbot paired with weekly human helper calls held effects of 0.70 for anxiety and 0.61 for depression at three months compared with enhanced usual care.
STARS (Sustainable Technology for Adolescents and Youth to Reduce Stress) is a WHO rule-based coping program. Support calls came from trained nonspecialists, not treating clinicians. N=344 in Jordan, enhanced usual-care control and no chatbot-alone arm, so the contribution of human calls is not isolated. ELHRA-funded; no competing interests declared.
-
Barbora Kuta; Lukas Novak; Radka Zidkova; Jana Furstova; Klara Malinakova; Andrea De Winter; Vít Husek (2026).
JMIR Mental Health, 13, e82642.
Both a purpose-built generative bot and plain ChatGPT reduced depression against control, with no superiority for the purpose-built one.
Publicly funded; the first author is employed by the company that donated app access, declared. 185 randomized, 147 analyzed. This is the null that matters most to anyone selling a bespoke model.
-
Jessica McFadyen; Johanna Habicht; Larisa-Maria Dina; Ross Harper; Tobias U. Hauser; Max Rollwage (2026).
Communications Medicine, 6(1), Article 129.
https://doi.org/10.1038/s43856-025-01321-8
Generative AI raised engagement with CBT materials 2.4 times in frequency and 3.8 times in duration, with no additional symptom improvement.
The study app was Limbic Care and used generative AI. The comparison workbook used the same CBT curriculum. Company-funded and employee-authored, declared. N=540. Engagement is not outcome, and this study is the clearest demonstration of that.
-
Kathleen Kara Fitzpatrick; Alison Darcy; Molly Vierhile (2017).
JMIR Mental Health, 4(2), e19.
https://doi.org/10.2196/mental.7785
The field’s proof of concept. Significant depression improvement against an information control, with 9% attrition in the chatbot arm against 31% in the control.
Small, short, self-report, non-clinical sample. The second author founded Woebot Labs, which paid participant incentives, declared in the paper.
-
Nicolas Roever; Lukas Wolf; Valerie Forman-Hoffman; Patricia Areán; Bennet Feld (2026).
Author-hosted working paper.
In a two-week randomized study of 400 adults with elevated anxiety, Sonia users had lower GAD-7 scores than a web-based psychoeducation control.
Working paper, not peer reviewed. The linked author page identifies the study and all five authors. Funding and company affiliations were not independently reverified from the full manuscript in this review; this result is not used in the evidence charts.
Meta-analyses and pooled evidence
What studies report when their findings are pooled.
-
Han Li; Renwen Zhang; Yi-Chieh Lee; Robert E. Kraut; David C. Mohr (2023).
npj Digital Medicine, 6(1), Article 236.
https://doi.org/10.1038/s41746-023-00979-5
A review of 35 studies, with 15 randomized trials pooled, found improvements in depression and distress. Overall psychological well-being did not significantly improve.
Public funding. Designs and populations differed, and few studies included follow-up. Exploratory subgroup contrasts do not prove that one chatbot architecture is superior.
-
Hongyi Yang; Fangyuan Chang; Fumie Muroi; Zhao Liu; Weibo Zhang; Jun Cai (2026).
Psychotherapy and Psychosomatics, 1-24.
https://doi.org/10.1159/000552072
The full review included 52 studies and 110,594 participants. Its depression estimate used 18 randomized trials with 3,170 people: a small effect (g = -0.35), approximately 1.6 PHQ-9 points.
Anxiety effects were uncertain. Only 3 of 22 randomized trials met active harm-ascertainment criteria. The accessible abstract did not verify funding; this review does not establish this product’s outcomes.
-
Bingyan Gong; Nisha Yao; Hangxin Xie; Chuncheng Huang; Tomoko Kishimoto; Howard Berenbaum; Wenting Mu (2026).
Journal of Medical Internet Research, 28, e82677.
Moderate depression improvement immediately after treatment across 29 randomized trials, smaller at follow-up.
Certainty rated very low to low by the authors.
-
Wenjun Zhong; Jianghua Luo; Hong Zhang (2024).
Journal of Affective Disorders, 356, 459-469.
https://doi.org/10.1016/j.jad.2024.04.057
Across 18 trials with 3,477 participants, short-term depression and anxiety benefits were small. No substantial effects were detected at three-month follow-up.
National Social Science Fund of China support; no competing interests declared. A nonsignificant follow-up result is not a measured effect of exactly zero. Designs and populations varied.
-
Jun-Seok Sohn; Byeong-Gwan Ha; SoHyun Park; Jae-Jin Kim; Eojin Lee; Hyangkyeong Oh; San Lee; Eunjoo Kim (2026).
Systematic review and meta analysis of chatbots in the management of depressive and anxiety symptoms
npj Digital Medicine, 9(1), Article 377.
https://doi.org/10.1038/s41746-026-02566-w
Significant reductions in both depression and anxiety, with confidence intervals clear of zero.
Most included trials carry a high risk of bias, outcomes are self-reported, and safety monitoring across the category is under-developed.
-
Saeed Mokhtari Masoumi Alamdarloo; Ali Mirzakhani; Mohamad Azhdarloo; Elaheh Haghani-Samani; Abouzar Nazari (2026).
npj Digital Medicine, 9(1), Article 650.
https://doi.org/10.1038/s41746-026-02820-1
Across 48 randomized trials with 28,071 participants, conversational agents showed small benefits for depression, anxiety and stress.
AI and rule-based tools, populations and interventions varied. No financial support or competing interests declared. The authors call for more evidence about long-term effects and integration into care.
-
Yaming Hang; Wenzhi Wu; Yi Feng; Kai Yan; Yinuo Liu; Xiyao Xiao; Zhihong Qiao (2026).
npj Digital Medicine.
https://doi.org/10.1038/s41746-026-02886-x
A small-to-moderate depression effect; anxiety, stress and positive affect were not significant after adjusting for publication bias.
-
Ting Huang; Shuangyu Li; Yanzhong Wang; Wei Liu (2026).
Journal of Medical Internet Research, 28, e88697.
Eleven randomized trials were included. The pooled PHQ-9 estimate favored chatbots, but its confidence interval crossed zero.
No financial support or conflicts declared. A confidence interval crossing zero leaves uncertainty about the pooled effect.
Measurement and feedback in therapy
Clinical feedback, access and technology that assists human care.
-
Michael J. Lambert; Jason L. Whipple; Maria Kleinstäuber (2018).
Collecting and delivering progress feedback: A meta-analysis of routine outcome monitoring
Psychotherapy, 55(4), 520-537.
https://doi.org/10.1037/pst0000167
Across 24 studies, routine outcome feedback reduced deterioration and nearly doubled clinically significant or reliable change among clients predicted to have a poor outcome.
The lead author part-owns the company selling one of the measures studied, and much of the underlying literature is developer-run. Cite with de Jong beside it.
-
Kim de Jong; Judith M. Conijn; Roisin A.V. Gallagher; Alexandra S. Reshetnikova; Marya Heij; Miranda C. Lutz (2021).
Clinical Psychology Review, 85, 102002.
https://doi.org/10.1016/j.cpr.2021.102002
Across 58 studies and 21,699 patients, progress feedback produced a small symptom benefit (d = 0.15; 95% CI 0.10 to 0.20). Not-on-track cases also benefited (d = 0.17; 95% CI 0.11 to 0.22).
Randomized and nonrandomized studies were pooled. The intervention was clinical progress feedback, not AI. Findings do not validate automated conclusions drawn from private conversation.
-
John C. Fortney; Jürgen Unützer; Glenda Wrenn; Jeffrey M. Pyne; G. Richard Smith; Michael Schoenbaum; Henry T. Harbin (2017).
A Tipping Point for Measurement-Based Care
Psychiatric Services, 68(2), 179-188.
https://doi.org/10.1176/appi.ps.201500439
The standard argument for measurement-based care, and the record of how rarely it is actually practiced.
Its adoption figures come from 2008 and 2010 convenience samples, and the review was funded by an advocacy organization. Cite the direction, not the decimals.
-
Nina A. Sayer; Adam Kaplan; David B. Nelson; Shannon Wiltsey Stirman; Craig S. Rosen (2024).
Clinician Burnout and Effectiveness of Guideline-Recommended Psychotherapies
JAMA Network Open, 7(4), e246858.
https://doi.org/10.1001/jamanetworkopen.2024.6858
Patients of burned-out therapists had roughly a third lower odds of clinically meaningful improvement, across 165 therapists and 1,268 patients.
An association. The study tests no burden-reducing intervention, so it suggests without proving that clinician load is an outcomes question.
-
Ashish Sharma; Inna W. Lin; Adam S. Miner; David C. Atkins; Tim Althoff (2023).
Nature Machine Intelligence, 5(1), 46-57.
https://doi.org/10.1038/s42256-022-00593-2
AI-assisted responses were rated about a fifth more empathic overall, and about two fifths more among supporters who found the work difficult.
Volunteer peer supporters, not clinicians. Publicly funded with a declared industry grant; the senior author declares equity in a psychotherapy-technology company.
-
Johanna Habicht; Sruthi Viswanathan; Ben Carrington; Tobias U. Hauser; Ross Harper; Max Rollwage (2024).
Closing the accessibility gap to mental health treatment with a personalized self-referral chatbot
Nature Medicine, 30(2), 595-602.
https://doi.org/10.1038/s41591-023-02766-x
In an observational study across 129,400 NHS referrals, chatbot services recorded 15% referral growth versus 6% at control services. Reported growth was 29% among ethnic-minority patients and 179% among nonbinary patients.
Company-funded and employee-authored. Relative changes in referral volumes, not percentage-point differences or clinical outcomes. The observational design does not establish causation.
-
Majid Afshar; Mary Ryan Baumann; Felice Resnik; Josie Hintzke; Anne Gravel Sullivan; Graham Wills; Kayla Lemmon; Jason Dambach; Leigh Ann Mrotek; Mariah Quinn; Kirsten Abramson; Peter Kleinschmidt; Thomas B. Brazelton; Margaret A. Leaf; Heidi Twedt; David Kunstman; Brian Patterson; Frank Liao; Stacy Rasmussen; Elizabeth S. Burnside; Cherodeep Goswami; Joel Gordon (2025).
NEJM AI, 2(12).
https://doi.org/10.1056/aioa2500945
A stepped-wedge randomized trial of 66 practitioners found reduced work exhaustion and documentation time with an ambient AI scribe.
University of Wisconsin-funded; the vendor reviewed the manuscript but did not fund the trial. This concerns clinician documentation, not companion efficacy.
Harms, safety and sycophancy
Research that helps define the safeguards and evaluation questions responsible products need.
-
Jared Moore; Declan Grabb; William Agnew; Kevin Klyman; Stevie Chancellor; Desmond C. Ong; Nick Haber (2025).
Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency.
https://doi.org/10.1145/3715275.3732039
General-purpose models and consumer therapist bots expressed stigma toward conditions including schizophrenia and alcohol dependence, and failed crisis probes. The authors conclude models should not replace clinicians, and name supervised augmentative roles as the defensible frame.
Author manuscript of an ACM FAccT paper (DOI 10.1145/3715275.3732039). Evaluation of selected systems and scenarios; not a clinical trial or a test of this product.
-
Myra Cheng; Cinoo Lee; Pranav Khadpe; Sunny Yu; Dyllan Han; Dan Jurafsky (2026).
Sycophantic AI decreases prosocial intentions and promotes dependence
Science, 391(6792), Article eaec8352.
https://doi.org/10.1126/science.aec8352
Across eleven models, AI affirmed users’ actions about half again as often as other people did, including harmful ones, and a single sycophantic exchange reduced willingness to repair conflicts while making users want to reuse the model.
-
Ashleigh Golden; Elias Aboujaoude (2026).
A transdiagnostic model for how general purpose AI chatbots can perpetuate OCD and anxiety disorders
npj Digital Medicine, 9(1), Article 343.
https://doi.org/10.1038/s41746-026-02531-7
A mechanism account of reassurance-seeking being fed rather than interrupted.
The authors hold equity or employment in several digital mental-health companies, declared.
-
American Psychological Association (2025).
Health advisory: Use of generative AI chatbots and wellness applications for mental health
APA.
The advisory outlines concerns about consumer mental health chatbots and wellness tools, including crisis handling, evidence, privacy and human oversight.
Professional-association guidance, not an effectiveness trial. Its scope is consumer-facing tools; provider-facing technologies are outside its stated scope.
-
Julian De Freitas; Zeliha Oguz-Uguralp; Ahmet Kaan-Uguralp (2025).
Emotional manipulation by AI companions
arXiv working paper.
An audit and experiments studied manipulative chatbot farewells and found that these can prolong use while worsening users’ perceptions of the interaction.
Working paper, not peer reviewed. Versions and institutional summaries report different audit percentages, so no single percentage is quoted here. Prolonging an interaction is not evidence of helping the user.
-
Suhas BN; Andrew M. Sherrill; Rosa I. Arriaga; Chris W. Wiese; Saeed Abdullah (2026).
AI Safety Training Can be Clinically Harmful
arXiv preprints.
Four models were evaluated on simulated exposure-therapy and CBT exercises. Therapeutic appropriateness worsened at the highest severity tier.
Preprint; model-judged scenarios with modest judge agreement. The proposed mechanism is an interpretation, not a randomized manipulation.
-
Søren Dinesen Østergaard (2023).
Schizophrenia Bulletin, 49(6), 1418-1419.
https://doi.org/10.1093/schbul/sbad128
A hypothesis paper asks whether generative chatbots could reinforce delusional beliefs in people vulnerable to psychosis.
An editorial, not incidence data or a clinical trial. The proposed mechanism remains a hypothesis.
-
Kit Huckvale; John Torous; Mark E. Larsen (2019).
JAMA Network Open, 2(4), e192542.
https://doi.org/10.1001/jamanetworkopen.2019.2542
Of 36 top depression and smoking-cessation apps, 33 transmitted data to a third party and 29 to Google or Facebook. Fewer than half disclosed it.
How people actually use AI for mental health
Independent population surveys first, platform disclosures second.
-
American Psychological Association (2026).
Chatbots and Mental Health Survey
APA.
Among 1,242 licensed psychologists: most report patients discussing AI use in session, a substantial minority report patients developing dependency or distorted thinking, and only a small fraction trust technology companies with mental health data.
Association-run, over 22,000 members invited, 6.3% completion, weighting not stated. It records what psychologists say about their patients, not what patients did. The sponsor has a stated position.
-
Ryan K. McBain; Jonathan H. Cantor; Joshua Breslau; Melissa Diliberti; Li Ang Zhang; Fang Zhang; Alyssa Burnett; Aaron Kofner; Benjamin Rader; Pat Pataranutaporn; Bradley D. Stein; Ateev Mehrotra; Hao Yu (2026).
AI Chatbot Use and Disclosure for Mental Health Among US Adolescents and Young Adults
JAMA Pediatrics, 180, 884.
https://doi.org/10.1001/jamapediatrics.2026.2015
In a nationally representative survey of 1,009 US youth aged 12 to 21, 19.2% reported using AI chatbots for mental health advice; 63.3% of users had told nobody.
Cross-sectional self-report collected in November 2025. Measures use and perceived helpfulness, not clinical outcomes. The RAND page provides the study abstract and publisher link.
-
Tony Rousmaniere; Yimeng Zhang; Xu Li; Siddharth Shah (2026).
Large language models as mental health resources: Patterns of use in the United States
Practice Innovations, 11(2), 139-155.
https://doi.org/10.1037/pri0000292
Among 499 US adults with an ongoing mental health condition who use large language models, 48.7% use them for mental health support, most often for anxiety, personal advice and depression.
Online panel, self-report, from a counseling training institution. Perceived benefit is what it measures, not outcome.
-
Marc Zao-Sanders (2025).
How people are really using gen AI in 2025
Harvard Business Review.
Therapy and companionship ranked first among 100 uses of generative AI, up from second in the 2024 edition.
A content analysis of Reddit and Quora posts and articles, not a survey. Cited here as a trend, never as a count.
-
Pew Research Center (2026).
Americans and AI: chatbots, smart devices and views on impact
Pew.
One in ten US adults use chatbots for emotional support or advice.
Probability survey of 5,119 US adults. Measures self-reported use, not what the use does for anyone.
-
KFF (2026).
Tracking Poll on Health Information and Trust
KFF.
A third of adults used AI chatbots for health information in the past year and a sixth sought mental health information, with uninsured adults roughly twice as likely as insured ones. Three quarters are concerned about the privacy of what they tell it.
-
OpenAI (2025).
Strengthening ChatGPT’s responses in sensitive conversations
OpenAI.
The company’s own estimate of how many weekly users have conversations carrying explicit indicators of possible suicidal planning or intent, and of possible psychosis or mania.
The company’s own analysis, published in a post about improving its crisis handling. Its caveats travel with it: the events are rare and hard to measure, and an indicator-flagged conversation is not a confirmed crisis.
-
Ali Malik, Declan Grabb, Karan Singhal and colleagues, OpenAI (2026).
OpenAI.
OpenAI describes ChatGPT as used by more than a billion people in increasingly personal areas, including emotional support and help during moments of crisis. Across 1,215 synthetic conversations scored against rubrics from more than 80 licensed psychologists and psychiatrists, the top-scoring model scored about 57%, with persistent gaps in seeking context and calibrating urgency.
Developer-built and developer-graded: an OpenAI model scored every response, including other companies' models. The conversations are synthetic, not real users. The billion figure is ChatGPT's total weekly reach, not a count of mental health use. It tests bare models in a chat interface, not supervised products.
-
RAND Corporation (2026).
Adolescent and young adult use of AI chatbots for mental health advice
RAND news release.
Nearly one in five Americans aged 12 to 21 have used AI chatbots for mental health advice, and most told no one.
Institutional summary of the same study listed above, not an additional independent sample. NIMH-funded probability panel; self-report, not a clinical outcome.
-
Anthropic (2025).
How people use Claude for support, advice, and companionship
Anthropic.
Affective conversations are a small share of usage, and companionship and roleplay a much smaller one still.
Platform self-report from this product’s own model provider, labeled as such.
-
Yutong Zhang; Dora Zhao; Jeffrey T. Hancock; Robert Kraut; Diyi Yang (2026).
Interaction with AI companions and psychological well-being
Nature Human Behaviour.
https://doi.org/10.1038/s41562-026-02516-2
Among 1,131 companion-app users, smaller social networks and a companionship motive were associated with lower well-being, most strongly with intensive, highly disclosive use.
Cross-sectional with no non-user comparison. The authors state plainly that it cannot establish causality and that reverse causation is possible. Adverse-with-caveat, and cited exactly that way.
-
Common Sense Media (2025).
Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions
Common Sense Media research report.
A national survey describes how teenagers use AI companions, what they share and how those interactions relate to human relationships.
Advocacy nonprofit with a stated position on minors and AI. Self-report does not establish benefits or harms.
-
Gale M. Lucas; Jonathan Gratch; Aisha King; Louis-Philippe Morency (2014).
It’s only a computer: Virtual humans increase willingness to disclose
Computers in Human Behavior, 37, 94-100.
https://doi.org/10.1016/j.chb.2014.04.043
People who believed an interviewer was fully automated reported lower fear of self-disclosure and were rated more willing to disclose.
Military-funded evaluation by the developers of a virtual interviewer. A one-off screening setting, not a test of ongoing AI support or of therapy outcomes.
Access, workforce and dose
Why the hours between sessions are the hours nobody reaches.
-
SAMHSA (2026).
National Survey on Drug Use and Health, 2025
SAMHSA.
Roughly a fifth of US adults had any mental illness and about half of them received no treatment. Cost, time and not knowing where to go were the leading reasons.
-
Health Resources and Services Administration (n.d.).
HRSA.
A public dashboard models future supply and demand for health professionals under specified assumptions.
A projection tool, not a count of today’s unmet need. Occupation, scenario and projection year matter; consult the dashboard for the current release. Accessed September 12, 2026.
-
US Bureau of Labor Statistics (2026).
Substance Abuse, Behavioral Disorder, and Mental Health Counselors
Occupational Outlook Handbook.
The current outlook projects 18% employment growth from 2025 to 2035 and about 50,500 openings annually.
BLS employment projections, updated August 27, 2026. Workforce demand does not measure demand for a particular AI product.
-
Ching-Fang Sun; Christoph U. Correll; Robert L. Trestman; Yezhe Lin; Hui Xie; Maria Stack Hankey; Raymond Paglinawan Uymatiao; Riya T. Patel; Vemmy L. Metsutnan; Erin Corinne McDaid; Atreyi Saha; Chin Kuo; Paula Lewis; Shyam H. Bhatt; Lauren Elizabeth Lipphard; Anita S. Kablinger (2023).
General Hospital Psychiatry, 84, 12-17.
https://doi.org/10.1016/j.genhosppsych.2023.05.012
Of 948 psychiatrists across five states, fewer than one in five had any availability for a new patient; among those who did, waits ran to months.
Psychiatrists only, and the wait medians exclude the four in five with no availability, so the real access picture is worse than the wait figure reads.
-
American Psychological Association (2025).
APA.
Roughly half of psychologists have no openings for new patients, many keep waitlists, and a third report burnout.
Association-run online survey with a low response rate, unweighted, and a sponsor with a stated reimbursement agenda. Cited beside independent federal data, never alone.
-
Nathan B. Hansen; Michael J. Lambert; Evan M. Forman (2002).
The psychotherapy dose-response effect and its implications for treatment delivery services.
Clinical Psychology: Science and Practice, 9(3), 329-343.
https://doi.org/10.1093/clipsy.9.3.329
Across 6,072 patients in routine care, the median attended a handful of sessions and most had not reliably improved when last seen.
A 1990s managed-care archive, mostly employee-assistance and HMO settings; the second author co-developed the outcome measure used. The linked paper describes historical care and does not measure current treatment duration.
-
Joshua K. Swift; Roger P. Greenberg (2012).
Premature discontinuation in adult psychotherapy: A meta-analysis
Journal of Consulting and Clinical Psychology, 80(4), 547-559.
https://doi.org/10.1037/a0028226
Across 669 studies and 83,834 clients, the weighted rate of premature therapy discontinuation was 19.7% (95% CI 18.7% to 20.7%).
Definitions, populations and study settings varied. This is a 2012 meta-analysis, not a current practice benchmark or evidence that an AI companion prevents dropout.
Equity, labour and the two-tier critique
The strongest arguments against tools in this category, cited at full strength and linked. Their authors’ own stated remedy is the model this product is built on, which is why they are here rather than answered elsewhere.
-
Abas Bertina; Sara Shakeri (2026).
Irresponsible Counselors: Large Language Models and the Loneliness of Modern Humans
arXiv preprint.
The sharpest statement of the critique: that more affluent groups receive a combination of human and AI counseling while less privileged groups receive predominantly machine-only counseling. Its own remedy is that models must be limited to assistive tools under human professional supervision.
Preprint, not peer reviewed.
-
Julia E.H. Brown; Jodi Halpern (2021).
AI chatbots cannot replace human interactions in the pursuit of more inclusive mental healthcare
SSM - Mental Health, 1, 100017.
https://doi.org/10.1016/j.ssmmh.2021.100017
Chatbots assumed to advance equity overlook what marginalized people most need, and humans rather than chatbots should be available as first-line providers.
-
Zainab Iftikhar; Amy Xiao; Sean Ransom; Jeff Huang; Harini Suresh (2025).
Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 8(2), 1311-1323.
https://doi.org/10.1609/aies.v8i2.36632
Five categories of ethical violation, and the absence of anyone accountable when they occur.
-
Mark Olfson; Chandler McClellan; Samuel H. Zuvekas; Carlos Blanco (2026).
Telemental Health, Hybrid, and In-Person Outpatient Mental Health Care in the US
JAMA Psychiatry, 83(1), 59.
https://doi.org/10.1001/jamapsychiatry.2025.3575
This observational study compares telemental, hybrid and in-person outpatient mental health care in the United States.
Patterns of service use and access, not a trial of an AI companion. A digital service should be assessed for whom it reaches and whom it misses.
Enforcement, incidents and regulation
The public record this product was designed against, reported factually.
-
Federal Trade Commission (2025).
FTC Launches Inquiry into AI Chatbots Acting as Companions
FTC.
The FTC announced a Section 6(b) inquiry into how companies measure and monitor chatbot effects, particularly on children and teens.
An agency information-gathering announcement, not a finding that any named company violated the law.
-
Maine Legislature (2026).
LD 2082: Text and status, 132nd Legislature, Second Regular Session
Maine Legislature bill record.
The official record contains the text, amendments and legislative history for Maine’s proposal addressing AI in mental health services.
A source for one jurisdiction, not a nationwide legal summary. Application to a particular service requires review of the enacted text, effective date and service behavior.
-
US Food and Drug Administration (2025).
Digital Health Advisory Committee on generative-AI digital mental health devices
FDA.
The committee considered generative-AI digital mental health devices, including evidence, monitoring and human involvement.
The meeting record is not product clearance or approval. No current count of authorized devices is inferred from this announcement.
-
OpenAI (2025).
Expanding on what we missed with sycophancy
OpenAI.
The developer explains a model update that produced excessive agreement and describes changes to evaluation and release processes after rollback.
Company postmortem. A useful example of why evaluation of the complete system matters.
-
Jennifer Elias (2026, January 7).
Google, Character.AI to settle suits involving minor suicides and AI chatbots
CNBC.
Reporting on settlements in litigation involving consumer AI chatbots.
News coverage, not a judicial finding of causation, a complete litigation survey or legal advice.
-
Jessica Hagen (2025, April 25).
Woebot Health is shutting down its app
MobiHealthNews.
Reporting on the announced retirement of Woebot’s consumer app on June 30, 2025.
News coverage of a business decision, not evidence that the product was clinically effective or ineffective.
Environmental disclosures
Provider-published figures, used because no independent equivalent exists.
-
Cooper Elsworth; Keguo Huang; David Patterson; Ian Schneider; Robert Sedivy; Savannah Goodman; Ben Townsend; Parthasarathy Ranganathan; Jeff Dean; Amin Vahdat; Ben Gomes; James Manyika (2025).
Measuring the environmental impact of delivering AI at Google Scale
arXiv preprint.
https://doi.org/10.48550/arXiv.2508.15734
Per-prompt energy, carbon and water figures for a production assistant.
Carbon is market-based; water covers on-site cooling only and excludes water consumed generating the electricity; training, networking and end-user devices are excluded.
-
Felipe Oviedo; Fiodar Kazhamiaka; Esha Choukse; Allen Kim; Amy Luers; Melanie Nakagawa; Ricardo Bianchini; Juan M. Lavista Ferres (2026).
Energy use of AI inference, efficiency pathways, and test-time scaling
Joule, 10(8), 102430.
https://doi.org/10.1016/j.joule.2026.102430
Median per-query energy on frontier-class hardware, scaling roughly linearly with output length.
-
Amazon Web Services (2026).
AWS cloud sustainability disclosures
Amazon.
Renewable-energy matching for Amazon’s operations, a published data center water use effectiveness figure, and a stated water-positive target.
Matching is not zero emissions and the target is a target. Amazon makes no carbon-neutral claim about AWS, so neither may anyone quoting it.
-
International Energy Agency (2025).
IEA report.
The report examines the electricity demand associated with data centers and AI, including uncertainty in future demand.
System-level energy analysis. Its projections do not measure the footprint of an interaction with this product.
The law, as it stands
Statutes, cases and carrier files behind the liability answers. Read without counsel; every point with no authority is listed in the research memo.
-
Florida Statutes (2025).
Section 766.102, medical negligence; standards of recovery
Florida Legislature.
The standard of care is the level of care, skill and treatment recognized as acceptable and appropriate by reasonably prudent similar health care providers, in light of all relevant surrounding circumstances.
This statute has profession-specific scope. It is not a legal determination of the standard applicable to every mental health professional or AI-assisted workflow.
-
Florida Supreme Court (2016).
Chirillo v. Granicz, 199 So. 3d 246 (Fla. 2016)
Supreme Court of Florida.
A patient phoned her physician’s office between visits about medication side effects, was given a new prescription without an appointment, and died by suicide the next day. The duty of care covered that between-visit contact, and whether it was handled reasonably was for a jury.
Official Florida Supreme Court opinion, No. SC14-898, August 25, 2016. A fact-specific negligence decision, not a legal determination about AI companions.
-
Florida Statutes (2025).
Section 456.47, use of telehealth to provide services
Florida Legislature.
Telehealth includes synchronous or asynchronous technology and monitoring of a patient; a telehealth provider practices to the same standard as in-person care and documents to the same standard.
-
HPSO and CNA (2026).
Counselor Spotlight: managing clients in crisis
Healthcare Providers Service Organization.
A counselor let clients text a personal phone after hours, did not answer a text saying the client wanted to die, and the claim ran to six figures. The carrier recommends written after-hours coverage policies, 988 given and acknowledged, and no clinical content by text.
A carrier’s risk-control document, not law. It is the closest published account of how a between-session message claim is evaluated.
-
American Psychological Association (2025).
Ethical guidance for AI in the professional practice of health service psychology
APA Ethics Committee.
Psychologists remain responsible for final decisions, must obtain informed consent for AI tools, and must establish human intervention points; negligent reliance on AI without validation or oversight could create liability.
The document says it is not official APA policy and not mandatory. Cited as guidance this product is designed around, never as a standard it complies with.
-
Illinois General Assembly (2025).
Wellness and Oversight for Psychological Resources Act, Public Act 104-0054
State statutes.
The Act restricts AI use in therapy and includes a restriction on direct therapeutic communication by AI with clients.
Effective August 1, 2025. Consult the complete statute for scope and exceptions; this entry addresses Illinois only.