The DOI matches its Crossref record; 85 of the 85 references it cites were found in a registry or an index; 9 of the 71 sources its claims cite were checked for support.
The Opportunities and Risks of Large Language Models in Mental Health
Open accessAbstract
Global rates of mental health concerns are rising, and there is increasing realization that existing models of mental health care will not adequately expand to meet the demand. With the emergence of large language models (LLMs) has come great optimism regarding their promise to create novel, large-scale solutions to support mental health. Despite their nascence, LLMs have already been applied to mental health–related tasks. In this paper, we summarize the extant literature on efforts to use LLMs to provide mental health education, assessment, and intervention and highlight key opportunities for positive impact in each area. We then highlight risks associated with LLMs’ application to mental health and encourage the adoption of strategies to mitigate these risks. The urgent need for mental health support must be balanced with responsible development, testing, and deployment of mental health LLMs. It is especially critical to ensure that mental health LLMs are fine-tuned for mental health, enhance mental health equity, and adhere to ethical standards and that people, including those with lived experience with mental health concerns, are involved in all stages from development through deployment. Prioritizing these efforts will minimize potential harms to mental health and maximize the likelihood that LLMs will positively impact mental health globally.
Overview
CoChat Summary
TL;DR
This narrative review summarizes early applications of large language models (LLMs) to mental health education, assessment, and intervention, arguing that they offer scalable opportunities but also substantial risks. It recommends domain-specific fine-tuning, equity, ethical safeguards, and human involvement across development, testing, and deployment.
Key findings
- LLMs show promise in mental health education: ChatGPT responses to postpartum depression questions were rated near ceiling for clinical accuracy (3.93/4) and more aligned with ACOG responses than Bard or Google Search; ChatGPT also improved provider efficiency by 37.5% when generating/editing behavioral health learning content.
- In assessment, Med-PaLM 2 predicted the correct DSM-5 diagnosis 77.5% of the time and 92.5% when asked for the correct diagnostic category, but it classified possible PTSD with high specificity (0.98) and low sensitivity (0.30) and correctly predicted comorbid diagnosis or diagnostic modifier only 20% of the time.
- In intervention, chatbots trained in empirically supported treatments show initial evidence of reducing depressive, anxiety, and stress symptoms, but they can be nonpersonalized, forget prior interactions, give nontherapeutic or iatrogenic advice, and inconsistently respond to suicide risk.
- The paper identifies cross-cutting risks including perpetuating inequalities/stigma, unethical service provision, insufficient reliability, inaccuracy, lack of transparency/explainability, and neglecting human involvement; it argues risks are generally higher for assessment and especially intervention than for education.
Methods
- Design: narrative review/commentary summarizing extant literature on LLM applications to mental health education, assessment, and intervention, plus proposed risk mitigation strategies.
- Data/sample: no primary data collection; the paper synthesizes cited studies, examples, and prior reviews, distinguishing general-purpose LLMs (eg, ChatGPT, Gemini) from domain-specific LLMs (eg, Med-LM, Mental-Alpaca, Psy-LLM).
- Analysis: thematic synthesis across education, assessment, and intervention, with a table assigning qualitative risk levels (lower, medium, higher) for potential harms across those domains.
Limitations
- The paper states no explicit limitations section.
- Evident from the text: it is a narrative review without a described systematic search strategy, inclusion criteria, or formal quality appraisal.
- Evident from the text: many cited LLM studies are preprints or early evaluations, and the risk estimates in Table 1 are qualitative and not empirically validated; the authors note the table is not meant to represent every possible LLM use case.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Figures




Verification
Claims the paper makes
7 Valid · 3 Not valid · 50 Not checked · 28 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“existing models of mental health care will not adequately expand to meet the demand”
- Also stated in: § Introduction · plain
“access to mental health care has not expanded to adequately meet the demand [4]”
§ The Opportunities and Risks of Large Language Models in Mental Health > Abstract
What the sources say
- Supportedcites [4] Wainberg, Challenges and opportunities in global mental health: a research-to-practice perspective, Curr Psychiatry Rep source
The section on integrating treatment into primary care says community mental health services have generally not been prepared to meet care needs, while many people cannot access care or remain untreated. Its subsection on severe mental illness adds that inpatient care contributes only modestly to treatment access and specialized services alone cannot cope with the burden in low- and middle-income countries. These passages support the claim that existing models are inadequate to meet demand. (LLM, conf=0.89)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § Introduction · plain
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“research finds that chatbots can be trained to express empathy [37-39]”
§ The Opportunities and Risks of Large Language Models in Mental Health > Applications of LLMs to Mental Health > Intervention
What the sources say
- Not checked yetcites [37] Brocki L Dyer GC Gładka A Chung NC . Deep learning mental health dialogue system. Presented at: 2023 IEEE International Conference on Big Data and Smart Computing (BigComp); Feb 13-16, 2023:395-398. Jeju, Korea. source
- Not checked yetcites [38] Martinengo, Evaluation of chatbot-delivered interventions for self-management of depression: content analysis, J Affect Disord source
- Partly supportedcites [39] You, Beyond self-diagnosis: how a chatbot-based symptom checker should respond, ACM Trans Comput-Hum Interact sourceExperimental — accuracy not yet measured
Our copy of You et al. (2023) clearly supports the general idea that chatbots can convey/express empathy: §3.2 states 'most research has proposed that healthcare chatbots be human-like and convey empathy,' cites studies showing users desire empathetic interactions and that an empathic chatbot was preferred over an advice-only one for self-diagnosis, and §7.3 repeats that 'chatbots should show empathy and emotion to promote the interaction.' However, no passage in our copy (searched: 'trained to express empathy,' machine-learning/affective-computing training, NLP training data) states that chatbots can be *trained* to express empathy — the paper's own emotional-support responses were conversationally designed (including a toggle button, §7.2), not learned. The 'trained' mechanism attributed here is therefore not this source's claim. (LLM, conf=0.70)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“maintain therapeutic conversations [14]”
§ The Opportunities and Risks of Large Language Models in Mental Health > Applications of LLMs to Mental Health > Intervention
What the sources say
- Partly supportedcites [14] Chan, Developing a natural language-based AI-chatbot for social work training: an illustrative case study, China J Soc Work sourceExperimental — accuracy not yet measured
The section ‘The chatbot can generate natural and continuous conversations’ reports that the chatbot responded to a counsellor and extended the conversation. But ‘Potential developments’ describes it as roleplaying a client for counselling training; the paper does not establish that it can maintain therapeutic conversations as a counsellor providing therapy. (LLM, conf=0.84)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“finding that using ChatGPT improved provider efficiency by 37.5%”
§ The Opportunities and Risks of Large Language Models in Mental Health > Applications of LLMs to Mental Health > Education
What the sources say
- Partly supportedcites [13] Barish G Marlotte L Drayton M Mogil C Lester P . Automatically enriching content for a behavioral health learning management system: a first look. Presented at: The 9th World Congress on Electrical Engineering and Computer Systems and Science; Aug 3-5, 2023; London, United Kingdom. [doi: 10.11159/cist23.125] sourceExperimental — accuracy not yet measured
The Initial results report that editing ChatGPT-generated overviews improved efficiency by 37.5% on average compared with writing overviews. The Methodology specifies that this was a content-production task involving four behavioral-health division employees, not a general measure of provider efficiency. (LLM, conf=0.86)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“ChatGPT has been shown to be able to explain why an individual was classified as experiencing stress or depressive symptoms [71]”
§ The Opportunities and Risks of Large Language Models in Mental Health > Risks Associated With Mental Health LLMs > Lack of Transparency and Explainability
What the sources say
- Supportedcites [71] Yang K Ji S Zhang T Xie Q Kuang Z Ananiadou S . Towards interpretable mental health analysis with large language models. arXiv. Preprint posted online onOct11, 2023. [doi: 10.48550/arXiv.2304.03347] source
The Additional information section shows ChatGPT explaining a severe-stress classification using emotional language and physical symptoms. The Supervised emotion-enhanced prompting section also gives an example in which it explains a classification of very severe depression. These are explanations of classifications, not proof that every explanation is accurate. (LLM, conf=0.92)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
47 more claims · 2 with their cited sources checked
Create a free account to see all 88 claims and to verify the 62 references still open.
- It may state two separate claims as one. 20
- Probably not a substantial claim. 21
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Recent reports suggest that interactions with large language model (LLM) chatbots can shape delusional content, amplify conviction, and entrench maladaptive safety behaviors [9,10].”
From Delusional Experiences Emerging From AI Chatbot Interactions or “AI Psychosis” (2025) · cites this paper as [9]
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“However, while users form relationships and experience emotional support from chatbots, interactions can also foster dependency, worsen mental health conditions, or have severe consequences [51] [52].”
From "I tell him everything that I do": An investigation of privacy and safety implications of AI companion usage (2025) · cites this paper as [51]
In the citing paper
“While early rule‐based systems such as Woebot prioritized safety through scripted dialogues, LLMs such as Replika now risk generating unvalidated advice due to their reliance on uncurated datasets 189 .”
From Charting the evolution of artificial intelligence mental health chatbots from rule‐based systems to large language models: a systematic review (2025) · cites this paper as [189]
In the citing paper
“While early rule‐based systems such as Woebot prioritized safety through scripted dialogues, LLMs such as Replika now risk generating unvalidated advice due to their reliance on uncurated datasets 189 .”
From Charting the evolution of artificial intelligence mental health chatbots from rule‐based systems to large language models: a systematic review (2025) · cites this paper as [189]
In the citing paper
“The standard of care requires LLMs to do these tasks [85], but we find no evidence of LLMs' specific capacities on them despite their widespread deployment as therapists.”
From Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers. (2025) · cites this paper as [85]
1 more claim about this paper
Create a free account to see every claim other papers make about this one.
Citations
Cites
2 retracted
Darcy, A., Beaudette, A., Chiauzzi, E., Daniels, J., Goodwin, K., Mariano, T. Y., Wicks, P., & Robinson, A. (2022) · Expert Review of Medical Devices · doi:10.1080/17434440.2022.2075726 · Cited by 40
Retracted(2023) · Expert Review of Medical Devices · doi:10.1080/17434440.2023.2267389 · Cited by 1
RetractedThe interpersonal theory of suicide.
Van Orden, K. A., Witte, T. K., Cukrowicz, K. C., Braithwaite, S. R., Selby, E. A., & Joiner, T. E. (2010) · Psychological Review · doi:10.1037/a0018697 · Cited by 5,292 · cited 1×
“These approaches illustrate ways that LLMs can support provider training and potentially enhance provider efficacy without providers becoming reliant on LLMs for in the moment critical thinking or decision-making.”
BackgroundLarge language models encode clinical knowledge
Singhal, K., Azizi, S., Tu, T., Mahdavi, S. S., Wei, J., Chung, H. W., Scales, N., Tanwani, A., Cole-Lewis, H., Pfohl, S., et al. (2023) · Nature · doi:10.1038/s41586-023-06291-2 · Cited by 4,083
Large Language Models Encode Clinical Knowledge
Singhal, K., Azizi, S., Tu, T., Mahdavi, S. S., Wei, J., Chung, H. W., Scales, N., Tanwani, A., Cole-Lewis, H., Pfohl, S., et al. (2022) · arXiv · doi:10.48550/arxiv.2212.13138 · Cited by 4,083
Large language models in medicine
Thirunavukarasu, A. J., Ting, D. S. J., Elangovan, K., Gutierrez, L., Tan, T. F., & Ting, D. S. W. (2023) · Nature Medicine · doi:10.1038/s41591-023-02448-8 · Cited by 3,944
Ayers, J. W., Poliak, A., Dredze, M., Leas, E. C., Zhu, Z., Kelley, J. B., Faix, D. J., Goodman, A. M., Longhurst, C. A., Hogarth, M., et al. (2023) · JAMA Internal Medicine · doi:10.1001/jamainternmed.2023.1838 · Cited by 2,613 · cited 1×
“One area of opportunity for LLMs in the mental health domain is to provide education about mental health (see Figure 1) [8].”
BackgroundJoiner, T. (2007) · Harvard University Press eBooks · doi:10.2307/j.ctvjghv2f · Cited by 2,425
Toward expert-level medical question answering with large language models
Singhal, K., Tu, T., Gottweis, J., Sayres, R., Wulczyn, E., Amin, M., Hou, L., Clark, K., Pfohl, S. R., Cole-Lewis, H., et al. (2025) · Nature Medicine · doi:10.1038/s41591-024-03423-7 · Cited by 1,734
The World Mental Health Report: Transforming mental health for all
Freeman, M. (2022) · World Psychiatry · doi:10.1002/wps.21018 · Cited by 1,472
Research works printed but not linked: 5
- Kim, Lang Learn Technol, 26
- Ma, AMIA Annu Symp Proc, 2023
- Primm, Prev Chronic Dis, 7
- Koocher GP Keith-Spiegel P . Ethics in Psychology and the Mental Health Professions: Standards and Cases. Oxford University Press; 2008.
- Wei J Wang X Schuurmans D et al. Chain-of-thought prompting elicits reasoning in large language models. In: Koyejo S Mohamed S Agarwal A , editors. Advances in Neural Information Processing Systems. Vol35. Curran Associates, Inc; 2022:24824-24837.
Cited by
Torous, J., Linardon, J., Goldberg, S. B., Sun, S., Bell, I., Nicholas, J., Hassan, L., Hua, Y., Milton, A., & Firth, J. (2025) · World Psychiatry · doi:10.1002/wps.21299 · Cited by 323
Moore, J., Grabb, D., Agnew, W., Klyman, K., Chancellor, S., Ong, D. C., & Haber, N. (2025) · Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency · doi:10.1145/3715275.3732039 · Cited by 204
Hua, Y., Siddals, S., Ma, Z., Galatzer‐Levy, I., Xia, W., Hau, C., Na, H., Flathers, M., Linardon, J., Ayubcha, C., et al. (2025) · World Psychiatry · doi:10.1002/wps.21352 · Cited by 117
The Applications of Large Language Models in Mental Health: Scoping Review
Jin, Y., Liu, J., Li, P., Wang, B., Yan, Y., Zhang, H., Ni, C., Wang, J., Li, Y., Bu, Y., et al. (2025) · Journal of Medical Internet Research · doi:10.2196/69284 · Cited by 94
Chow, J. C. L., & Li, K. (2025) · Information · doi:10.3390/info16070549 · Cited by 84
Metadata
- Authors
- Lawrence, H. R., Schneider, R. A., Rubin, S. B., Matarić, M. J., McDuff, D. J., & Jones Bell, M.
- Affiliations
- Google (United States) (US)
- ORCID
- Venue
- JMIR Mental Health
- Publisher
- JMIR Publications Inc.
- Published
- 2024-07-29
- Volume
- 11
- Pages
- e59479-e59479
- ISSN
- 2368-7959 (ISSN-L)
- PMC
- PMC11301767
- OpenAlex
- W4393177920
- Cited by
- 235
- References
- 85
- Citations by year
- 2026: 134 · 2025: 96 · 2024: 5
- FWCI
- 92.23
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- cc-by
- Locations
- JMIR Mental Health (publishedVersion, cc-by)
- PubMed Central (publishedVersion, cc-by)
- PubMed (publishedVersion)
- arXiv (Cornell University) (submittedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- arXiv (Cornell University) (cc-by)
- Topics
- Social Sciences › Psychology › Applied Psychology › Digital Mental Health Interventions
- Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
- Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
- Keywords
- large language models, mental health, mental health education, mental health assessment, mental health interventions, AI risks, ethical AI, mental health equity, lived experience, responsible AI deployment
- MeSH terms
- Mental Disorders, Humans, Language, Mental Health, Mental Health Services
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-05, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-05 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: The Opportunities and Risks of Large Language Models in Mental Health. https://app.cochat.ai/papers/10.2196/59479
- The paper
- Lawrence, H. R., Schneider, R. A., Rubin, S. B., Matarić, M. J., McDuff, D. J., & Jones Bell, M. (2024). The Opportunities and Risks of Large Language Models in Mental Health. JMIR Mental Health. https://doi.org/10.2196/59479
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.