What the early treatment trial demonstrates
Heinz et al. (2025) randomized 210 adults with selected clinically significant symptoms or elevated eating-disorder risk to Therabot or a waitlist. The four-week intervention produced greater symptom reductions, with assessment through eight weeks. It is a meaningful proof of concept for that expert-developed system and study population. It is not a comparison with a psychologist, and participants’ alliance ratings cannot establish comparable treatment effectiveness. Because the accessible material for this review was the publisher abstract, detailed safety procedures and attrition were not independently extracted. Interpretation needs to keep the short duration, selected sample and inactive comparator visible rather than turning this trial into a general statement that AI therapy works.
Sources: Heinz et al., 2025
Engagement is not symptom superiority
More demanding controls can change the conclusion. In the 2026 trial by McFadyen et al., a generative-AI-enabled CBT tool increased use compared with digital CBT workbooks, but did not show overall superiority on depression or anxiety symptoms. Exploratory differences among users who chose particular features are not randomized tests of those features. Kuta et al. (2026) likewise found no clear advantage of a structured solution-focused chatbot over ChatGPT in a small feasibility trial, despite a PHQ-9 advantage over assessment-only control. High attrition and the exclusion of people receiving psychiatric treatment restrict that study’s interpretation. Together these findings support a distinction between making an intervention more engaging and demonstrating added clinical benefit.
Sources: McFadyen et al., 2026; Kuta et al., 2026
Automation is not one uniform technology
So et al. (2026) examined natural-language feedback added to self-guided cognitive-restructuring exercises. Its adherence findings are relevant to automated support, but an automated-feedback system should not be relabeled a current generative chatbot. Findings in selected efficacy-analysis subsets also cannot replace the randomized intention-to-treat clinical comparison. In a separate two-week exploratory study, Campellone et al. (2025) compared generative and rules-based versions of a conversational intervention. Relationship and user-experience measures were broadly similar; no serious or device-related adverse events were reported. This is useful evidence about that controlled implementation. Its modest size and duration do not establish safety against uncommon events, and the authors’ employment by the developer is relevant context.
Sources: So et al., 2026; Campellone et al., 2025
What the newer synthesis can and cannot settle
Gong et al. (2026) provide a more recent synthesis of CBT-oriented chatbot trials. Their final published version reports improvement in depressive symptoms, but rates the certainty of the main symptom findings as low, with very low certainty for follow-up anxiety. Its prediction interval for postintervention depression includes the possibility of no benefit in a future setting. The synthesis combines different chatbot architectures and controls. It should therefore be described as evidence about the included conversational interventions, not as a verdict on all large language models. The contrast between its final version and its earlier preprint also illustrates why publication version and search cutoff need to be recorded.
Sources: Gong et al., 2026
Clinical safety needs different tests
Trials of symptom change do not answer every safety question. Moore et al. (2025) demonstrated stigma and inappropriate responses in experimental therapy-like scenarios with specific LLMs, including responses that could reinforce delusional beliefs. These experiments identify failure modes; they do not estimate how often users are harmed. A November 2026 preprint by Weilnhammer et al. extends testing to multi-turn triage and reports substantially more missed emergencies than when the relevant information is supplied together. It is preliminary simulation evidence, not clinical incidence. Its design nevertheless highlights a useful evaluation question: can the system revise an early benign interpretation as risk becomes clearer?
Sources: Moore et al., 2025; Weilnhammer et al., 2026
Observational findings should not become causal headlines
General AI use is a different exposure from a mental-health intervention. Perlis et al. (2026b) observed an association between frequent use and depressive symptoms in a cross-sectional US survey; confounding and reverse causation remain plausible. Their later longitudinal target-trial emulation did not find a significant subsequent symptom difference after adjustment (Perlis et al., 2026a). That analysis still had substantial follow-up loss and was not a randomized experiment. Neither study proves that chatbots generally cause depression or that vulnerable people face no risk. The relevant conclusion is that associations, controlled treatment effects and reports of particular harms answer different questions.
Sources: Perlis et al., 2026b; Perlis et al., 2026a
Application to Ella and the research agenda
The American Psychological Association’s advisory distinguishes consumer wellness products from regulated treatments and emphasizes evidence, privacy and safeguards. For Ella, a reasonable research agenda is to test the actual system rather than borrow results from Therabot, ChatMind or an earlier model. Evaluation should include credible active controls, clinician-reviewed multi-turn cases, unsupported reassurance, bias, data handling and escalation performance, followed by prospective monitoring. This is a proposed program of work, not a claim that those checks or a clinical trial have already occurred. These sources do not establish therapy replacement, diagnostic accuracy, or guaranteed crisis safety for Ella.
Sources: American Psychological Association, 2025; Heinz et al., 2025; Moore et al., 2025; McFadyen et al., 2026; Kuta et al., 2026; Weilnhammer et al., 2026
References
Links open the original papers or official sources. Access labels describe the material assessed for this review.
Heinz, M. V., Mackin, D. M., Trudeau, B. M., Bhattacharya, S., Wang, Y., Banta, H. A., Jewett, A. D., Salzhauer, A. J., Griffin, T. Z., & Jacobson, N. C. (2025). Randomized trial of a generative AI chatbot for mental health treatment. NEJM AI, 2(4), AIoa2400802. 10.1056/AIoa2400802
Access assessed: Publisher indexed abstract; direct publisher full page returned 403; no full-text safety extraction
McFadyen, J., Habicht, J., Dina, L.-M., Harper, R., Hauser, T. U., & Rollwage, M. (2026). Increasing engagement with cognitive-behavioral therapy (CBT) using generative AI: A randomized controlled trial (RCT). Communications Medicine, 6(1), 129. 10.1038/s43856-025-01321-8
Access assessed: PubMed and publisher indexed abstract/results excerpts; direct publisher page redirected to inaccessible cookie service
Kuta, B., Novak, L., Zidkova, R., Furstova, J., Malinakova, K., De Winter, A., & Husek, V. (2026). Effectiveness of a fully automated mobile therapeutic versus a general chatbot in reducing depression and anxiety and improving well-being: Feasibility randomized controlled trial. JMIR Mental Health, 13, e82642. 10.2196/82642
Access assessed: Publisher full text; abstract, participant flow, analysis and editorial notice inspected
So, M., Sekizawa, Y., Hashimoto, S., Kashimura, M., Yamakage, H., & Watanabe, N. (2026). Effect of AI-based natural language feedback on engagement and clinical outcomes in fully self-guided internet-based cognitive behavioral therapy for depression: 3-arm randomized controlled trial. Journal of Medical Internet Research, 28, e76902. 10.2196/76902
Access assessed: Publisher full text; abstract, participant flow, funding and outcome interpretation inspected
Campellone, T. R., Flom, M., Montgomery, R. M., Bullard, L., Pirner, M. C., Pavez, A., Morales, M., Harper, D., Oddy, C., O’Connor, T., Daniels, J., Eaneff, S., Forman-Hoffman, V. L., Sackett, C., & Darcy, A. (2025). Safety and user experience of a generative artificial intelligence digital mental health intervention: Exploratory randomized controlled trial. Journal of Medical Internet Research, 27, e67365. 10.2196/67365
Access assessed: Publisher full text; abstract, methods and conflicts inspected
Gong, B., Yao, N., Xie, H., Huang, C., Kishimoto, T., Berenbaum, H., & Mu, W. (2026). Efficacy, user engagement, and acceptability of cognitive behavioral therapy–oriented psychological chatbots for adults with depressive and/or anxiety symptoms: Systematic review and meta-analysis of randomized controlled trials. Journal of Medical Internet Research, 28, e82677. 10.2196/82677
Access assessed: Final publisher full text; Results/GRADE and author record inspected
Moore, J., Grabb, D., Agnew, W., Klyman, K., Chancellor, S., Ong, D. C., & Haber, N. (2025). Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers. Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency, 599–627. 10.1145/3715275.3732039
Access assessed: Published conference paper PDF on author institution website; experimental scope inspected
Weilnhammer, V., Dolan, R., Luettgau, L., Corno, V., Wilkinson, E., Sounderajah, V., Richardson, H., Kelly, C., & Nour, M. M. (2026). Path-dependent triage: Early judgments lead conversational AI chatbots to miss mental-health emergencies. arXiv preprint, arXiv:2611.09647v1. 10.48550/arXiv.2611.09647
Access assessed: Primary arXiv abstract and submission history; not peer reviewed
Perlis, R. H., Gunning, F. M., Uslu, A. A., Santillana, M., Baum, M. A., Druckman, J. N., Ognyanova, K., & Lazer, D. (2026b). Generative AI use and depressive symptoms among US adults. JAMA Network Open, 9(1), e2554820. 10.1001/jamanetworkopen.2025.54820
Access assessed: Publisher full text; abstract, methods, Results and correction note inspected
Perlis, R. H., Gunning, F. M., Uslu, A., Santillana, M., Baum, M. A., Druckman, J. N., Ognyanova, K., & Lazer, D. (2026a). Emulated trial of artificial intelligence use and subsequent depressive outcomes in a survey of US adults. BMJ Mental Health, 29(1), e302609. 10.1136/bmjment-2026-302609
Access assessed: Publisher indexed PDF abstract and author institutional abstract; direct publisher full text returned 403
American Psychological Association (2025). Use of generative AI chatbots and wellness applications for mental health: An APA health advisory. American Psychological Association.
Access assessed: Official advisory and issuing-body publication information
This evidence review is educational. It does not provide an individual diagnosis or treatment plan. Read the review methods, access limits, and clinical-review status.