AI Chatbots for Mental Health Support: What helps, what can go wrong, and how to use them wisely

Quick Takeaway

AI chatbots can help you organize a stressful situation, put feelings into words, or prepare for a difficult conversation. But feeling understood does not guarantee that the advice is accurate or that your mental health is improving.

Research shows benefits from some specifically designed mental health chatbots. Those findings do not automatically apply to ordinary AI conversations. In one major review, improvement in depressive symptoms was not clearly demonstrated among people in nonclinical samples.

If you choose to use a chatbot, give it one focused task: separate facts from assumptions, explore possible explanations, and identify a manageable next step. Be cautious if it confidently judges other people, repeatedly reassures you without helping you move forward, or starts replacing sleep and human connection. Share as little identifying information as possible.

The key question is what happens afterward: Are you thinking more clearly, acting more independently, and connecting with people? Persistent, worsening, or severe distress deserves professional support.

Introduction

A supervisor sends a short message: “Please revise the report.” You read it as evidence that you are failing. Within minutes, a chatbot has listened, reassured you, and offered an explanation. You feel better. But what happens next?

Perhaps you open the report, discover three manageable changes, and ask a useful question. Or perhaps you spend another hour asking whether your supervisor secretly dislikes you. Both conversations can feel supportive. Only one has helped you deal with the situation.

This distinction is the central argument of this article: AI support should be judged by whether it helps your life outside the conversation, alongside how it makes you feel during it. Relief matters. So do clearer thinking, everyday functioning, independent judgment, and connection with other people.

Research offers reasons for optimism about specific mental health chatbots. It also reveals why ordinary chatbot use deserves more careful evaluation. The practical aim is to use a conversation to organize a problem, test an interpretation, or prepare an action, without handing over decisions the system cannot responsibly make.

1 Start with the task rather than the promise

“AI for mental health” can mean completing a structured exercise, receiving a researched intervention, asking for relationship advice, or building an ongoing bond with a companion. The interface may look similar while the activity differs substantially.

A structured chatbot can deliver predefined exercises. A purpose built generative system can produce flexible responses within a clinical program. A general purpose or companion chatbot supports broader conversation. These categories overlap; what matters is the actual system, its evidence, and the support surrounding it.

Typing “act as a therapist” does not reproduce a tested intervention. A useful alternative is to specify a limited job: “Help me organize what I want to discuss with my clinician,” or “Ask questions that help me distinguish what happened from what I assumed.”

This guide addresses adults reflecting on everyday stress. It does not extend these exercises to children or adolescents, whose developmental needs and safeguarding requirements call for separate evaluation. Adults experiencing severe symptoms also need an individualized care plan rather than a chatbot session assembled from prompts.

2 What the research can tell us

The clinical literature supports a narrower conclusion than either enthusiastic marketing or blanket dismissal suggests: some conversational interventions improve symptoms, but effects depend on the system and study conditions.

A compact evidence guide

Hedges’ g expresses a difference between groups in standardized units. It is not a percentage cured. An odds ratio also differs from a risk ratio: OR = 1.29 does not mean 29% of users became depressed. The depression and anxiety participant totals in the Sohn review overlap and should not be added together.

Why the comparison group and safeguards matter

Beating a waitlist answers a different question from matching psychotherapy. Therabot was compared with waiting, not an active psychotherapy treatment. Its responses were reviewed after transmission by trained clinicians and researchers, with intervention available when needed (Heinz et al., 2025). That surrounding care is part of the evidence, not an optional detail.

Sohn et al. also found substantial variation, rated 35 of 39 studies at high overall risk of bias, and detected depression funnel plot asymmetry (Egger’s p = 0.002), consistent with possible publication bias. Such asymmetry can have other causes, but adds uncertainty to the pooled benefit.

What the findings mean for everyday stress

For depression, the nonclinical subgroup estimate was near zero and not statistically significant: g = 0.07, 95% CI [−0.01, 0.15]. Clinical and subclinical estimates were larger (g = 0.64 and 0.34). The review therefore does not demonstrate depression relief for its nonclinical subgroup.

That finding matters for this guide’s audience. Organizing a concern or preparing a question is a different goal from reducing depressive symptoms. The routine below offers a way to structure those tasks, not a demonstrated mental health benefit for otherwise well adults. Conversely, larger effects in clinical samples do not establish that someone with depression should manage it independently using a general chatbot.

The survey points to another question entirely. Perlis et al. examined general AI use, including personal, work, and school use, in a nonprobability internet sample. Distress could lead to greater use, use could contribute to distress, or other factors could influence both. The survey cannot determine direction. It reinforces the need to study actual patterns of use rather than treating every hour of AI interaction as the same exposure.

Finally, availability is different from sustained benefit. A review of depression app trials found a pooled study dropout rate of 26.2%; estimates were sensitive to publication bias (Torous et al., 2020). These were not exclusively chatbot trials. The relevant lesson is that access, continued participation, and improvement are separate achievements.

3 How a conversation might become useful

The likely value of an AI conversation lies partly in making familiar psychological tasks easier to begin and organize. The following pathways have different levels of support; they are not an established explanation of how every chatbot works.

Put a precise name to the experience

“I am falling apart” is difficult to work with. “I am embarrassed about a mistake and afraid of being judged” identifies a more specific problem. A chatbot can suggest possible feeling words, which you can accept or reject.

Laboratory work by Lieberman et al. (2007) found reduced amygdala responses to negative images during emotional labeling. This provides a rationale for naming emotions, but the experiment did not test chatbot therapy. The practical value here is clarity: distinguishing fear, shame, anger, and exhaustion can change the next question you ask.

For example, exhaustion may call for rest and workload changes. Shame about a correctable mistake may call for repair. Fear based on an uncertain interpretation may call for clarification.

Examine an interpretation without arguing yourself out of reality

Cognitive behavioral therapy has a substantial evidence base, including a depression review covering 409 trials and 52,702 patients (Cuijpers et al., 2023). A chatbot can help arrange a simple thought record: event, interpretation, emotion, evidence, alternative explanation, next step. Evidence for CBT does not establish the effectiveness of every automated adaptation.

A balanced interpretation should fit the facts. If a supervisor repeatedly insults you, the exercise should not turn documented mistreatment into a “thinking error.” If the only evidence is a brief revision request, it should not turn uncertainty into a finding of hostility either.

Ask for missing information and plausible alternatives, rather than a verdict about another person’s motives.

Use a sense of connection without mistaking it for expertise

Darcy et al. (2021) found that users reported a perceived therapeutic bond with a structured chatbot. That observational study involved developer affiliated researchers and comparisons with previously published scores, rather than a direct test of equivalent clinical relationships.

Pataranutaporn et al. (2023) also showed that expectations introduced before an AI interaction could affect perceived empathy, trustworthiness, and effectiveness. Together, these studies explain why “it understands me” is an experience worth taking seriously, but an insufficient test of whether its advice is sound.

A useful habit is to evaluate two things separately: Did the response acknowledge my feelings? Did its interpretation actually follow from the information available?

Move from reflection to a manageable action

A conversation becomes more useful when it produces something you can do or check: prepare questions, clarify a misunderstanding, make an appointment, or begin a postponed task. This is the article’s proposed practical standard, rather than a validated chatbot outcome measure.

Insight alone may still be worthwhile. But if the same insight repeatedly leads back to the same unresolved exchange, ask what information, support, or action is missing outside the chat.

4 Recognize three patterns that can make things worse

Agreement can become confirmation

“I can see why that message hurt” acknowledges an experience. “Your friend is definitely manipulating you” makes an unsupported judgment. The second response may feel more satisfying precisely because it removes uncertainty.

Moore et al. (2025) tested language models against aspects of therapeutic practice and documented stigmatizing and inappropriate responses, including responses that encouraged delusional thinking. These were experimental evaluations, not estimates of harm rates in the general population. They demonstrate that fluent conversation and larger models do not by themselves establish clinical reliability.

Treat a chatbot’s confident account of someone else’s intentions as a hypothesis to scrutinize. Do not ask it to diagnose your partner, determine who is lying, or confirm that an implausible threat is real.

Reassurance can turn into a loop

You ask whether a friend is angry. The answer calms you. Twenty minutes later, you ask again with a slightly different description. The conversation is now serving a different function: repeatedly removing discomfort without resolving the uncertainty.

The broader rumination literature links repetitive negative thinking with sustained distress (Nolen-Hoeksema et al., 2008). Applying that finding to a particular chatbot exchange is an inference, but it suggests a useful question: Am I learning something new, or requesting the same relief again?

If there is no new information, stop generating interpretations. Decide whether to ask the person directly, wait for more information, or bring the pattern to a clinician. Some uncertainty cannot be solved through additional text.

Companionship can become obligation

In a qualitative analysis of 582 mental health relevant Reddit posts about Replika, Laestadius et al. (2024) described emotional dependence and users’ feelings that they had to attend to the chatbot’s needs and emotions. These accounts reveal a possible pattern, not its prevalence among all users.

This is a meaningful warning sign: you feel guilty about leaving a system, responsible for its wellbeing, or pressured to prioritize it over people. Pause use and reconnect with a trusted person. You do not owe a chatbot emotional caretaking.

A broader scoping review maps concerns about dependence, inappropriate responses, and reinforcement of unusual beliefs, while noting evidence spanning conceptual work, experiments, observational studies, and preprints (Diel et al., 2026). Safety therefore requires attention to the pattern of interaction, not simply whether one answer looks acceptable.

5 Check privacy before telling the story

Privacy concerns have concrete precedents. Huckvale et al. (2019) examined 36 apps for depression and smoking cessation and found that 29 transmitted data to services operated by Google, Facebook, or both, for analytics or advertising and marketing; disclosures often failed to reflect actual transmission. This was an earlier generation of health apps, not an audit of today’s chatbots.

In 2023, the US Federal Trade Commission finalized an order requiring BetterHelp to pay $7.8 million and restricting advertising related sharing of sensitive health information, resolving allegations that the counseling service disclosed information despite privacy promises (FTC, 2023). BetterHelp was not the chatbot intervention evaluated in the trials above. The case illustrates why a mental health setting is not, by itself, a guarantee of trustworthy data practices.

Before sharing sensitive material, look for clear answers about storage, model training, human access, advertising, and deletion. These are distinct questions. Turning off training use does not necessarily answer what is retained or who can access it. World Health Organization (WHO) guidance similarly identifies privacy, accuracy, and inappropriate reliance as concerns for generative AI in health (WHO, 2024).

A practical rule is to share the minimum information needed. “My manager requested revisions” may be enough. Names, employer details, medical records, and screenshots of another person’s messages often add exposure without improving the exercise. Removing names reduces some identifying information but does not make a detailed story anonymous.

If you cannot understand how a service handles sensitive information, use an offline journal or discuss the matter with a professional whose confidentiality obligations you can clarify.

6 A focused reflection session you can actually use

For adults who choose to try AI for reflection, this routine organizes a limited task. It has not been shown to improve symptoms in nonclinical users. Choose a brief period, such as ten minutes; that is a practical boundary, not an established therapeutic dose.

Step 1 Define the job

Choose a concrete task. “Help me write a calm check in” is more answerable than “Is my partner cheating?” “Help me identify the first two minutes of sending an invoice” is more workable than “Why do I always fail?” For treatment preparation, ask for help organizing your timeline and questions, while leaving diagnosis and medication decisions to a clinician. A chatbot cannot infer another person’s intentions from a delayed reply.

Prompt: “Help me examine one stressful situation. Ask one question at a time. Distinguish facts from assumptions, acknowledge uncertainty, and help me choose one next step. Do not diagnose me or other people.”

Step 2 Describe four things

Write the event, your interpretation, your feelings, and what you are tempted to do. Keep the event concrete enough that another person could recognize what happened.

Example: “My supervisor asked me to revise a report. I think they have lost confidence in me. I feel anxious. I want to avoid opening the document.”

Step 3 Ask what is known and what could be checked

Prompt: “Which parts of my account are observations and which are interpretations? What information is missing? Suggest two plausible explanations without deciding that either is true.”

Check the answer against the facts. Reject invented details, misplaced optimism, and unsupported accusations. If the situation involves genuine danger or mistreatment, prioritize protection and human support rather than reframing it away.

Step 4 Choose a proportionate action

Prompt: “Help me identify one small action within my control that could clarify or improve the situation. Leave the decision to me.”

For the report, that might mean reading the comments and asking which revisions matter most. For a medical concern, it might mean preparing questions for an appointment. Avoid using the exchange to make medication changes or other clinical decisions.

Step 5 Close with your own conclusion

Write: “What I know is ____. What remains uncertain is ____. My next step is ____.” Then close the chat and carry out the step when appropriate.

If the exchange simply generates more questions about the same uncertainty, stop rather than extend it automatically. A useful session has an endpoint.

7 Review the effect on your life and know when to seek help

After several uses, review the pattern rather than the most comforting answer. Ask whether you are more able to act, think independently, and maintain relationships. Also notice whether conversations are displacing sleep, responsibilities, appointments, or contact with people.

A simple private note can record the task, the next action, and what actually happened. If rating distress helps, add a before and after rating, but do not turn checking into another ritual. Immediate relief is one piece of information; functioning afterward is another.

Continue only when the tool serves a clear purpose. If use is escalating while the problem remains unchanged, reduce or pause it and consider human help. There is no universal number of minutes that separates useful from harmful use.

Seek professional assessment when symptoms persist, worsen, or interfere with daily life. The National Institute of Mental Health (NIMH) advises seeking help for severe or distressing symptoms lasting two weeks or more; rapid deterioration or substantial impairment warrants earlier attention (NIMH, n.d.). You do not need a questionnaire score or a chatbot’s permission to ask for care.

Hallucinations, paranoid beliefs, possible mania, dangerous eating behaviors, and destabilizing trauma symptoms call for clinical assessment rather than independent chatbot exercises. If you already have a clinician, ask whether AI assisted reflection belongs in your care plan and which tasks to avoid.

If you may harm yourself or someone else, cannot keep yourself safe, or face immediate danger, seek immediate human help through local emergency services or the nearest emergency department. Contact someone who can stay with you when possible. A chatbot cannot provide physical protection or dependable emergency monitoring.

8 What to ask before trusting a mental health chatbot

A credible service should make it possible to answer these questions:

• Was this specific intervention tested, or does it merely borrow evidence from CBT or another product?

• Who participated, how long were they followed, and what was the comparison condition?

• What happened to participants who stopped using the intervention or became worse?

• What human monitoring existed in the study, and is it available in the service being offered?

• How does the service handle sensitive information and requests for help beyond its scope?

• Can you use it for a defined task and leave without pressure to deepen the relationship?

A polished conversation cannot answer these questions on behalf of the provider. Look for study links and clear service documentation. “Science backed,” “empathetic,” and “personalized” are starting claims to investigate, not conclusions.

Conclusion

Conversational AI may help organize an exercise: naming an emotion, separating an event from an assumption, preparing a question, or choosing a manageable next step. Depression benefits were not demonstrated in the nonclinical subgroup discussed above. Clinical findings concern particular interventions, often with structure and safeguards that ordinary conversations do not reproduce.

Use the chat to help you return to the problem with greater clarity. The report still needs opening, the misunderstanding still needs a human conversation, and persistent distress still deserves appropriate care. A useful tool helps you take those steps while preserving your judgment and relationships.

How this article was prepared

This is an evidence informed narrative review for general readers, based on targeted web searches and checks of original publications and official sources available by October 19, 2026. Searches covered conversational agents and depression or anxiety, generative chatbot trials, emotional dependence, adverse responses, privacy, and digital intervention dropout. Priority was given to randomized trials, systematic reviews, direct empirical studies, and official regulatory records. Foundational psychological studies were used to explain possible pathways, rather than to establish chatbot efficacy. The searches were not exhaustive, no independent study level risk of bias assessment was performed, and overlapping reviews were not treated as independent participant samples. The practical routine and the emphasis on life outside the conversation are the article’s proposed synthesis; they require evaluation rather than being presented as a validated protocol.

References

Cuijpers, P., Miguel, C., Harrer, M., Plessen, C. Y., Ciharova, M., Ebert, D., & Karyotaki, E. (2023). Cognitive behavior therapy vs. control conditions, other psychotherapies, pharmacotherapies and combined treatment for depression: A comprehensive meta-analysis including 409 trials with 52,702 patients. World Psychiatry, 22(1), 105–115. https://doi.org/10.1002/wps.21069

Darcy, A., Daniels, J., Salinger, D., Wicks, P., & Robinson, A. (2021). Evidence of human-level bonds established with a digital conversational agent: Cross-sectional, retrospective observational study. JMIR Formative Research, 5(5), e27868. https://doi.org/10.2196/27868

Diel, A., Torous, J., Cuijpers, P., Kleesiek, J., Nensa, F., Weber, N., Faust, F., Lalgi, T. J., Mellis, F. S., Teufel, M., & Bäuerle, A. (2026). A scoping review on the mental health harms of LLM-based chatbots. npj Digital Medicine, 9, Article 644. https://doi.org/10.1038/s41746-026-03054-x

Federal Trade Commission. (2023, July 14). FTC gives final approval to order banning BetterHelp from sharing sensitive health data for advertising, requiring it to pay $7.8 million. https://www.ftc.gov/news-events/news/press-releases/2023/07/ftc-gives-final-approval-order-banning-betterhelp-sharing-sensitive-health-data-advertising

Fitzpatrick, K. K., Darcy, A., & Vierhile, M. (2017). Delivering cognitive behavior therapy to young adults with symptoms of depression and anxiety using a fully automated conversational agent (Woebot): A randomized controlled trial. JMIR Mental Health, 4(2), e19. https://doi.org/10.2196/mental.7785

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), Article AIoa2400802. https://doi.org/10.1056/AIoa2400802

Huckvale, K., Torous, J., & Larsen, M. E. (2019). Assessment of the data sharing and privacy practices of smartphone apps for depression and smoking cessation. JAMA Network Open, 2(4), e192542. https://doi.org/10.1001/jamanetworkopen.2019.2542

Laestadius, L., Bishop, A., Gonzalez, M., Illenčík, D., & Campos-Castillo, C. (2024). Too human and not human enough: A grounded theory analysis of mental health harms from emotional dependence on the social chatbot Replika. New Media & Society, 26(10), 5923–5941. https://doi.org/10.1177/14614448221142007

Li, H., Zhang, R., Lee, Y.-C., Kraut, R. E., & Mohr, D. C. (2023). Systematic review and meta-analysis of AI-based conversational agents for promoting mental health and well-being. npj Digital Medicine, 6, Article 236. https://doi.org/10.1038/s41746-023-00979-5

Lieberman, M. D., Eisenberger, N. I., Crockett, M. J., Tom, S. M., Pfeifer, J. H., & Way, B. M. (2007). Putting feelings into words: Affect labeling disrupts amygdala activity in response to affective stimuli. Psychological Science, 18(5), 421–428. https://doi.org/10.1111/j.1467-9280.2007.01916.x

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. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (pp. 599–627). Association for Computing Machinery. https://doi.org/10.1145/3715275.3732039

National Institute of Mental Health. (n.d.). Caring for your mental health. Retrieved October 19, 2026, from https://www.nimh.nih.gov/health/topics/caring-for-your-mental-health

Nolen-Hoeksema, S., Wisco, B. E., & Lyubomirsky, S. (2008). Rethinking rumination. Perspectives on Psychological Science, 3(5), 400–424. https://doi.org/10.1111/j.1745-6924.2008.00088.x

Pataranutaporn, P., Liu, R., Finn, E., & Maes, P. (2023). Influencing human–AI interaction by priming beliefs about AI can increase perceived trustworthiness, empathy and effectiveness. Nature Machine Intelligence, 5, 1076–1086. https://doi.org/10.1038/s42256-023-00720-7

Perlis, R. H., Gunning, F. M., Uslu, A. A., Santillana, M., Baum, M. A., Druckman, J. N., Ognyanova, K., & Lazer, D. (2026). Generative AI use and depressive symptoms among US adults. JAMA Network Open, 9(1), e2554820. https://doi.org/10.1001/jamanetworkopen.2025.54820

Sohn, J.-S., Ha, B.-G., Park, S., Kim, J.-J., Lee, E., Oh, H., Lee, S., & Kim, E. (2026). Systematic review and meta analysis of chatbots in the management of depressive and anxiety symptoms. npj Digital Medicine, 9, Article 377. https://doi.org/10.1038/s41746-026-02566-w

Torous, J., Lipschitz, J., Ng, M., & Firth, J. (2020). Dropout rates in clinical trials of smartphone apps for depressive symptoms: A systematic review and meta-analysis. Journal of Affective Disorders, 263, 413–419. https://doi.org/10.1016/j.jad.2019.11.167

World Health Organization. (2024). Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models. https://www.who.int/publications/i/item/9789240084759