Can AI Replace Therapy? Benefits, Limits and Human Counselling

AI and mental health

AI can help you reflect, learn and feel supported. A helpful conversation, however, does not establish that a chatbot can provide the assessment, emotional work and professional responsibility involved in psychotherapy.

By Alistair Gordon, MA, RCC, founder of Wellspring Counselling & Psychotherapy Inc. · November 8, 2026
Educational information. This article discusses general-purpose chatbots separately from purpose-built clinical tools.

Why people turn to AI, and where it can help

An immediate, apparently nonjudgmental response can be reassuring when you feel overwhelmed or alone. Writing to a chatbot may help you put an experience into words before you feel ready to discuss it with another person. Cost, availability and uncertainty about counselling can also influence that choice.

Useful, limited tasks might include:

• Learning basic psychological concepts and preparing questions for a counsellor.
• Organizing your thoughts before a session, without including identifying details.
• Rehearsing a conversation or considering a different way to express a need.
• Reviewing a coping skill already discussed with your clinician.
• Keeping a brief record of situations, reactions and questions to bring to therapy.

These uses can have value. They do not establish a diagnosis or determine which treatment is suitable for you. If you are using AI alongside counselling, discussing its suggestions with your counsellor can help you notice when they support your goals and when they pull you away from them.

When validation becomes an echo chamber

Having your feelings acknowledged matters. But emotional validation and agreement with your interpretation are different things.

A counsellor can help you explore those possibilities collaboratively, without assuming that your feelings are wrong or that the other person is harmless. The aim is a fuller understanding of the situation, including actual mistreatment where it exists.

AI can sometimes offer alternative interpretations. It can also follow the framing you supply too readily. In a 2025 experimental study, researchers found that the tested language models sometimes responded inappropriately to serious mental-health scenarios, including reinforcing delusional beliefs. Those findings concern the models and situations studied; they are not a test of every current or future system. [2]

For anxiety, another question is whether repeated requests for certainty are becoming part of the problem. Relief after another answer may be brief. A therapist can help you examine what you do when uncertainty returns, rather than simply supplying reassurance again.

Understanding an emotion is different from working through it

You may understand why you react strongly and still feel the same fear, anger or shame. Knowing that a partner is trustworthy does not automatically settle the alarm that arises when they seem distant.

In Wellspring’s clinical perspective, it can be helpful to explore several layers of an experience: thoughts, bodily sensations, emotions, protective responses and underlying needs. Anger may coexist with hurt. Withdrawal may protect against shame. A repeated argument may involve both a longing for connection and fear of being rejected.

These are possibilities to explore, not hidden meanings a therapist can simply read from a person. The client’s own experience remains central. A counsellor might ask, “What are you noticing as you say that?” or “Does that interpretation fit, or are we missing something?”

That work may involve helping you stay with an emotion at a manageable pace, express a need, practise a different response, or notice what happens between you and the counsellor. Our emotional regulation counselling page introduces the team who can support this work.

Nonverbal cues and the timing of an intervention

A text conversation does not reveal changes in breathing, tone, posture or facial expression unless you describe them. In a session, a counsellor may notice a pause or change in expression and check what it means for you. A cue is an invitation to inquire, not proof of an emotion.

Some AI systems can process voice and images. It would therefore be inaccurate to say that all AI can only handle text. Access to more inputs, however, does not by itself demonstrate reliable clinical judgment: knowing when to explore, when to challenge, when to slow down and when a different kind of support is needed.

The difference also applies to online (virtual) counselling with a human clinician. Therapy through video remains a professional relationship with a counsellor; it is not the same service as talking to an AI.

Insight, emotional learning and memory reconsolidation

Psychotherapy can include practical skills, changes in behaviour, a different relationship with thoughts, and experiences that change how old situations feel in the present. No single mechanism explains every effective therapy or every client’s improvement.

Memory reconsolidation is one research framework for understanding how reactivated memories may become open to updating under certain conditions. Some experiential approaches draw on that framework when helping clients explore established emotional expectations.

For example, someone who expects vulnerability to lead to rejection may gradually encounter a different experience while expressing a difficult feeling in therapy. Clinically, what matters is whether that experience is meaningful and helpful to the person. We cannot conclude from an emotional shift alone that a particular memory was reconsolidated.

Researchers continue to debate the mechanisms, boundary conditions and translation of laboratory findings into clinical practice. Reconsolidation should not be presented as a guaranteed process, proof that one approach is superior, or something only humans could ever facilitate. [4]

A chatbot can explain the idea or propose an exercise. That is different from establishing that the exercise is appropriate for you, responding to your reactions and monitoring the work over time. For difficult experiences, explore trauma counselling or meet the clinicians who offer EMDR therapy.

The therapeutic relationship includes responsibility

Human counselling offers more than conversational warmth. A qualified clinician works with you to understand your concerns, agree on goals, review progress and adjust the work. They also have professional responsibilities concerning consent, confidentiality, boundaries and responding to risk within their role.

Human therapists can make mistakes, and a good fit is not automatic. You should be able to question an interpretation, say that something is unhelpful, and discuss whether another clinician or approach would suit you better.

That accountability is especially relevant when several perspectives are involved. In couples counselling, for example, the work involves two people’s experiences and their interaction. An account given by one partner to a chatbot may leave out important context, even when that person is trying to be fair.

What does research on AI therapy actually show?

A 2025 randomized trial involving 210 adults found that the specially developed Therabot system improved several measured symptoms compared with a waitlist. It did not directly compare Therabot against human psychotherapy or establish equivalence to a therapist. The research team monitored interactions and could intervene. [1]

Those findings are encouraging. They do not show that every commercial chatbot is effective, that an unmonitored tool is appropriate for everyone, or that a sense of connection guarantees safety.

The American Psychological Association’s 2025 advisory recommends that consumer chatbots should not replace qualified mental-health care and emphasizes informed, cautious use. It also distinguishes general-purpose systems from wellness tools and other clinical technologies. [3]

AI will continue to change. The practical question is whether a specific tool has suitable evidence, safeguards and accountability for the job you are asking it to do—not whether it can sound convincing.

Questions to ask before relying on AI for emotional support

• What is this product intended to do? Look for evidence about the actual tool, not just claims about “AI therapy” in general.
• Am I getting broader perspective or repeated reassurance? Notice whether using it helps you act, connect and tolerate uncertainty—or keeps you asking the same question.
• What happens to my information? Check retention, access, sharing and training settings. Avoid identifying details about yourself, partners or clients. A consumer chat is not automatically protected like a clinical record.
• Could this suggestion be wrong for me? Do not treat an AI answer as a diagnosis, medication instruction or a substitute for assessment.
• What human support is available? Consider a trusted person, your clinician or an appropriate local service when the issue exceeds a tool’s role.

If there is an immediate safety concern, seek urgent human assistance through local emergency or crisis services. Do not wait for a chatbot to assess the situation.

When it may be time to speak with a counsellor

Consider human support when distress keeps returning, relationships are affected, daily functioning is becoming difficult, or you understand a pattern but cannot change it. You can also seek counselling before things reach that point.

You do not need a diagnosis or a polished explanation to begin. “I keep having the same reaction,” “AI helps me feel better briefly,” or “I’m not sure what I need” are all reasonable starting points.

Common questions

Sources and further reading

1. Heinz, M. V., et al. (2025). Randomized trial of a generative AI chatbot for mental health treatment. NEJM AI. See also Dartmouth’s account of the trial and clinical oversight.
2. Moore, J., et al. (2025). Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers. Proceedings of FAccT ’25.
3. American Psychological Association. (2025). Health advisory: Use of generative AI chatbots and wellness applications for mental health.
4. Appraising reconsolidation theory and its empirical validation (2022), and Beyond reconsolidation: The need for a broad theoretical approach in clinical translations of research on retrieval-induced plasticity (2024).

Research findings and product capabilities change. Clinical examples above are illustrative, not client stories. This article describes considerations for choosing support; it does not predict an individual’s treatment outcome.