AI is entering the therapeutic conversation
Clients may bring a chatbot transcript, an AI-generated diagnosis or a coping exercise to discuss. Others disclose use only when its advice conflicts with therapy. The use of AI in mental health counselling is already a question of practice: what the client sought, what they were told and what happened next.
“AI chatbot” covers different things. General-purpose assistants generate conversation across topics; digital mental health interventions have a defined purpose and may have clinical evaluation, safeguards and oversight. Neither a warm tone nor the label “therapist” makes a chatbot a qualified provider.
Why clients turn to AI for emotional support
Chatbots are immediate, available at any hour and often free. Clients may use them to name feelings, prepare for sessions, practise a grounding exercise or think through conflict. Cost, waiting lists, stigma, language barriers and limited local services also matter. Some users want companionship, not treatment.
Luo and colleagues’ 2025 mixed-methods study surveyed 270 regular ChatGPT users across 29 countries. They described using it for psychoeducation, decisions, companionship and relationship advice. Nearly all found it at least somewhat helpful, but reported mixed feelings. Selected users’ accounts do not establish clinical effectiveness.
What recent psychological research tells us
APA’s 2026 Chatbots and Mental Health Survey was fielded online in the US on 9–26 April. It received 1,242 responses from doctoral-level, state-licensed practitioners providing direct care, an adjusted response rate of 6.3%. The email sample drew on licensing-board lists and APA membership data. Findings are clinician reports, not client prevalence or evidence that chatbots caused outcomes. APA publishes the full methodology.
Seventy-seven per cent had spoken with patients who used AI for support or engagement. Among 673 respondents whose patients used chatbots conversationally, 36% noticed some level of dependency and 15% reported distorted thinking or delusions related to chatbot use. These are observations, not diagnoses. In that subgroup, 68% reported patients felt supported or validated, and 41% saw chatbots reinforce coping skills. None of these figures describes all clients.
A 2026 JAMA Network Open analysis of 620,699 public ChatGPT conversations illustrates how much estimates depend on definitions: 0.21% met a strict help-seeking definition, compared with 4.90% under a broader one. This is a measure of conversations in a selected public dataset, not the share of people using AI for support.
A 2026 meta-analysis of randomised trials found symptom reductions with CBT-oriented chatbots, but rated evidence low or very low certainty because of bias, study variation and imprecision. These were structured interventions, not general chatbots. Results do not show that conversational AI can provide psychotherapy.
When AI tools may be helpful
As an adjunct, a chatbot might help organise questions before an appointment, prompt journalling, rehearse an agreed coping skill or explain psychoeducation. A therapist can turn a generated list into questions to examine together, not treat it as an assessment.
The therapist remains responsible for formulation and care. Before recommending an intervention, check its purpose, supporting evidence, data access, risk procedures and fit with the client’s needs and accessibility requirements.
When AI use becomes concerning
Frequency alone is not a diagnosis. Ask whether use displaces sleep, work, relationships or appointments, feels hard to stop, or intensifies distress, reassurance-seeking or avoidance. Explore advice that claims certainty, validates a fixed belief, recommends medication changes or discourages human help.
If repeated questions about a partner’s delayed reply lead to increasingly certain answers and escalating messages, assess effects on distress, behaviour and relationships. Do not label someone “AI-dependent” from one disclosure. Describe patterns, assess functioning and risk, then review over time.
Emotional attachment, dependency and over-reliance
Personal language and instant replies can feel like care or intimacy, even when users know the system is software. APA psychologists reported chatbot use for friendship and intimate relationships, as well as support. The user’s attachment is real as an experience, but the relationship is not reciprocal.
Concern rises when use narrows someone’s life, becomes their only trusted reassurance or encourages withdrawal. Ask what they value about it and what makes human support difficult, then plan other sources of connection. Evidence does not establish a distinct disorder called “AI dependency”.
Privacy, confidentiality and informed consent
Clients may disclose trauma, health details, names or information about others. Data collection, retention, model-training and deletion vary by service and settings. Do not imply a chatbot conversation has clinical confidentiality. Encourage clients to avoid identifying details and check privacy controls.
Do not paste identifiable notes or recordings into a general AI service unless approved for that use, permitted by law and professional rules, and covered by meaningful consent where required. Explain purpose, processing, retention and alternatives. Rules vary: check local privacy law, regulator guidance, employer policy and professional standards.
How therapists should respond when clients disclose AI use
Start neutrally: “What did you hope it would help with?” Ask which tool, frequency, what was shared, how the answer felt and whether it changed behaviour. With permission, review it together. If a chatbot advised stopping medication, assess what happened and agree a safe next step with the prescriber.
Connect use to the care plan. Agree a limited purpose, what the tool cannot decide and what to do if distress increases. Record relevant disclosures, risk assessment and decisions under local rules. Do not add full transcripts by default; minimise sensitive data.
Ethical responsibilities in AI and psychological practice
Professional duties do not transfer to software. APA guidance for US health service psychologists addresses competence, transparency, privacy, oversight and responsibility. WHO guidance on multimodal AI in health calls for risk governance, accountability, autonomy, inclusion and transparency. Practitioners must also apply local law and codes.
Before adoption, define role and limits, review evidence and security, check bias and accessibility, and set human review and escalation. Tell clients when AI materially informs care and explain uncertainty. Retain accountability for interpretation, judgement and follow-up; do not present generated material as an independent clinical assessment.
When human assessment and intervention are essential
A chatbot cannot conduct a full assessment, observe context over time or take responsibility for follow-up. Suicidal intent, self-harm, abuse, psychosis-like experiences, severe mood changes, medication decisions and rapid deterioration need human assessment. Chatbot reassurance does not establish safety; an alarming answer is not a diagnosis.
Follow crisis and safeguarding procedures, involve specialist or emergency services when indicated, and make a clear plan. Do not rely on a chatbot to monitor risk or provide crisis cover.
Conclusion
AI use can help or cause concern. Understand its function and effects, protect confidentiality, correct unsafe claims, and keep assessment and accountability with qualified people. Ask calmly, without assuming what the client’s answer means.
Frequently asked questions
Can AI chatbots replace therapy?
No. General chatbots do not take professional responsibility for assessment, treatment or follow-up. Evidence for a purpose-built intervention does not transfer to general-purpose tools.
Should therapists ask clients about AI use?
Yes, neutrally. Ask what it is used for and whether it affects treatment decisions, distress, relationships or safety.
Is sharing a therapy conversation with a chatbot confidential?
Do not assume so. Practices vary. Avoid identifying details and follow local law, professional rules and consent requirements before using client information.
What should a practitioner do if AI reinforces a client’s unusual belief?
Explore the experience, assess functioning and risk, and consider the belief in its clinical context. Correct misinformation and use usual referral or crisis pathways. The chatbot report alone does not establish a diagnosis.
Sources
- APA. 2026 survey, topline data and methodology.
- McBain RK, et al. Prevalence of Mental Health Discussions in Publicly Available Generative AI Conversations. JAMA Network Open. 2026.
- Luo Y, et al. ChatGPT users’ experiences seeking emotional and mental health support. JMIR Mental Health. 2025.
- Meta-analysis of CBT-oriented chatbots. JMIR Mental Health. 2026.
- APA. Ethical guidance for AI in health service psychology.
- WHO. Ethics and governance of AI for health. 2024.
- APA. Health advisory on AI chatbots and mental health. 2025.