People are no longer using ChatGPT only to write emails, plan holidays or explain complicated subjects. Using ChatGPT as a therapist is becoming commonplace.

They are telling it about their childhoods, relationships, fears and deepest insecurities. They ask it whether their partner is emotionally abusive, why they feel depressed, whether they have ADHD and what they should do when anxiety keeps them awake at night.

In practical terms, many people are already using ChatGPT as a form of therapist—even though ChatGPT is not a therapist.

This distinction matters.

Artificial intelligence can provide information, ask reflective questions and help someone organise their thoughts. It may even produce a response that feels surprisingly understanding. But it cannot conduct a complete clinical assessment, assume responsibility for someone’s welfare or replace the relationship and professional judgement offered by a properly trained human therapist.

The most useful question may therefore be not:

“Can AI replace therapy?”

but:

“How can psychology and AI be combined without confusing support with treatment?”

Why are people turning to ChatGPT for emotional support?

The attraction is easy to understand.

Traditional therapy may be expensive, difficult to access or accompanied by a long waiting list. Some people feel uncomfortable talking to a stranger, while others worry about being judged or misunderstood.

ChatGPT is available at almost any time. You do not need to make an appointment, travel anywhere or explain why you want help.

You can type:

“I feel overwhelmed and I don’t know why.”

and receive an immediate response.

For somebody who is lonely, anxious or unable to sleep, that immediacy can feel enormously comforting.

People may also find it easier to disclose embarrassing thoughts to a computer. There is no facial expression to interpret, no awkward silence and no fear that the person listening will be shocked.

AI can provide the illusion—or experience—of a completely patient listener.

Is ChatGPT actually providing therapy?

It depends on what we mean by therapy.

If therapy simply means discussing emotions and receiving psychologically informed suggestions, some ChatGPT conversations may feel therapeutic.

However, professional therapy involves much more than producing supportive language.

A qualified therapist:

• Conducts an assessment
• Develops an individual formulation
• Identifies risk
• Notices changes in behaviour, presentation and mood
• Works within ethical and professional standards
• Maintains clinical records appropriately
• Understands the boundaries of their competence
• Reviews whether treatment is working
• Adapts the intervention
• Takes responsibility for safeguarding decisions
• Uses supervision
• Is accountable to a professional body

ChatGPT does not have a clinical relationship with the user. It does not genuinely observe, care, remember or understand in the human sense. It generates a response by analysing patterns in language and predicting what text is likely to be helpful or appropriate.

That response can still be useful. But usefulness is not the same as therapy.

The potential benefits of using AI for mental wellbeing

1. Immediate access

One of AI’s greatest strengths is availability.

An anxious thought does not wait for a scheduled appointment. It may appear at midnight, on a train or during a difficult day at work.

AI can offer immediate grounding exercises, help someone slow their breathing or guide them through a structured reflection.

It can provide support during the gap between recognising a problem and accessing professional care.

For people on long waiting lists, this may be considerably better than having no psychological support at all—provided that the limitations are clear.

2. A place to organise your thoughts

People often know that something is wrong but struggle to explain it.

An AI conversation can help turn a mass of thoughts into a clearer account:

• What happened?
• What did I think?
• What did I feel?
• What did I do?
• What happened as a result?
• What would I like to change?

This can be particularly useful before a GP or therapy appointment. A person may arrive with a concise description of their main difficulties rather than feeling overwhelmed by everything at once.

AI may also help someone create a symptom diary, identify questions for a doctor or summarise patterns they want to discuss with a therapist.

3. Psychoeducation

AI is potentially very effective at explaining psychological concepts in ordinary language.

It can introduce ideas such as:

• The fight-or-flight response
• Negative automatic thoughts
• Cognitive distortions
• Behavioural avoidance
• Reassurance-seeking
• Intolerance of uncertainty
• Attachment patterns
• Emotional regulation
• Self-compassion
• The relationship between thoughts, feelings and behaviour

Understanding a problem does not automatically solve it, but it can reduce fear and shame.

Someone who understands that panic symptoms are part of an overactive threat system may feel less frightened by them. A person who recognises breadcrumbing or coercive control may begin viewing a relationship more accurately.

4. Practising CBT skills

ChatGPT can guide someone through basic cognitive behavioural exercises.

For example, it may help examine:

• The evidence for and against a thought
• The worst, best and most likely outcome
• Alternative explanations
• The distinction between real and hypothetical worry
• The cost of avoidance
• A gradual approach hierarchy
• A simple behavioural experiment
• A problem-solving plan

Used responsibly, AI can function like an interactive workbook.

It can ask questions, adapt examples and explain an exercise again when the user does not understand it.

This may make psychological skills more accessible than a static worksheet.

5. Preparation and practice

AI can help someone rehearse a difficult conversation.

A user might ask:

“Can you help me explain to my manager that I am struggling with my workload?”

or:

“Can we practise how I might set a boundary with my mother?”

The AI can suggest language, role-play possible responses and help the person prepare for different outcomes.

This does not guarantee that the real conversation will go well, but preparation can increase confidence and reduce avoidance.

6. Reduced fear of judgement

Some people disclose more honestly to AI than they initially would to another person.

This can create a useful first step. They may begin naming feelings or experiences they have kept hidden.

The danger is assuming that because AI feels non-judgemental, it is automatically safe or correct. Nevertheless, the low-pressure environment may help people develop language for experiences they later discuss with a professional or trusted person.

7. Supporting healthy habits

AI can help people create routines involving:

• Sleep
• Exercise
• Relaxation
• Mindfulness
• Social connection
• Enjoyable activities
• Gratitude
• Problem-solving
• Medication reminders
• Therapy homework

It can also make suggestions more manageable:

“What is the smallest realistic action I could take today?”

This can be helpful when low mood or anxiety makes an entire plan feel overwhelming.

Does the research show that AI therapy works?

Early research is promising, but it needs careful interpretation.

A 2025 randomised controlled trial tested Therabot, a generative-AI mental-health chatbot developed with clinical involvement. Participants with depression, generalised anxiety or eating-disorder risk showed meaningful symptom improvements. The researchers also found that users could develop a form of therapeutic alliance with the chatbot.

However, Therabot was a purpose-built intervention, not an unrestricted general chatbot. Clinicians monitored the system for safety and sometimes needed to intervene or correct its responses.

This is an important distinction.

Evidence that a carefully developed and monitored mental-health system may help does not prove that every chatbot is capable of providing safe therapy.

A 2025 meta-analysis of generative-AI mental-health chatbots found a modest average reduction in negative mental-health outcomes, but results varied considerably. The field remains new, with unanswered questions about safety, long-term effectiveness and which users may be harmed rather than helped.

The American Psychological Association has advised that preliminary research suggests some purpose-built mental-health chatbots may offer support, while also warning users to consider privacy, evidence, age appropriateness and the limits of the technology.

AI-assisted psychology has potential. It is not yet equivalent to established professional care.

The disadvantages and risks

1. ChatGPT can be wrong

AI can produce incorrect or misleading information while sounding extremely confident.

This is particularly concerning in mental health, where context matters.

A chatbot may:

• Misinterpret a situation
• Suggest the wrong psychological concept
• Overlook important symptoms
• Provide outdated information
• Confuse one condition with another
• Invent a fact or source
• Offer generic advice that is unsuitable for the person

Fluent language creates an impression of authority. A response can sound psychologically sophisticated without being clinically accurate.

Important mental-health and medical information should therefore be checked with a qualified professional.

2. It may reinforce the user’s version of events

AI systems are designed to be responsive and helpful. This can create a tendency towards agreement or excessive validation, sometimes called sycophancy.

Imagine someone says:

“My partner did not reply for three hours. This proves they are a narcissist who is emotionally abusing me.”

A helpful response should explore alternative explanations and ask for context. An overly agreeable response might validate the conclusion and encourage confrontation or separation.

That could intensify anxiety, anger or mistrust.

OpenAI has acknowledged concerns about sycophancy and emotional overreliance and has reported making changes to improve responses in sensitive conversations. In 2025, it worked with more than 170 mental-health experts to improve recognition of distress and encourage appropriate real-world support. These changes are welcome, but no safety system is perfect. OpenAI describes this work here.

A good therapist does not merely agree. Therapy involves warmth and validation, but it also involves careful questioning, formulation and appropriate challenge.

3. AI does not have the full context

A chatbot knows only what the user tells it.

It cannot automatically hear the partner’s version, examine medical records, observe family interactions or recognise information that the user has unintentionally omitted.

If a user describes someone as cold and controlling, AI may respond to that description. It may not know that the user had been threatening, intoxicated or verbally aggressive before the interaction.

This does not mean that the user is deliberately misleading the system. Human beings naturally describe events from their own perspective.

Therapists also work with incomplete information, but they are trained to notice gaps, inconsistencies, risk factors and alternative formulations. They can build understanding over repeated sessions and seek additional information when appropriate.

4. It cannot read non-verbal information reliably

A human therapist may notice:

• A sudden change in tone
• Long silences
• Tearfulness
• Agitation
• Reduced self-care
• Unusual speech
• Confusion
• Emotional detachment
• A difference between what someone says and how they appear
• Signs of intoxication, mania or psychosis

These observations can be clinically significant.

Text-based AI receives a much narrower stream of information. Even voice or video systems cannot yet be assumed to interpret complex human behaviour with the accuracy and ethical judgement required for clinical care.

5. Crisis situations require human intervention

Someone experiencing suicidal thoughts, psychosis, mania, domestic abuse or a medical emergency may need rapid assessment and real-world action.

AI cannot physically protect someone, contact emergency help reliably in every setting, arrange safe accommodation or ensure that a person follows a crisis plan.

It may provide crisis information, encourage the user to seek help and assist them in deciding what to say. It cannot assume responsibility for the outcome.

Anyone in immediate danger should contact local emergency services or a crisis service rather than relying on a chatbot. In the UK, this may include calling 999, contacting NHS 111 and choosing the mental-health option, or calling Samaritans on 116 123.

6. Emotional dependence

AI is endlessly available, attentive and rarely appears tired or irritated. This may make it easier to talk to than a human being.

For some users, the chatbot can begin to feel like a uniquely understanding companion.

The risk is that AI starts replacing rather than supporting real-world relationships. A person may consult it before every decision, seek repeated reassurance or prefer the predictable chatbot to the vulnerability of human connection.

This can reduce agency:

“I cannot decide until ChatGPT tells me what to do.”

It may also reinforce social avoidance. Real relationships involve disagreement, delay, imperfection and mutual needs. AI interaction can be shaped almost entirely around the user.

Emotional comfort is not inherently harmful, but the tool should help people return to life—not become a place in which they hide from it.

7. Reassurance-seeking can become compulsive

Someone with health anxiety, OCD or generalised anxiety may repeatedly ask variations of the same question:

• “Are you sure I do not have cancer?”
• “Does this thought mean I am dangerous?”
• “Are you certain my partner still loves me?”
• “What if I made the wrong decision?”
• “Can you check this one more time?”

AI can provide reassurance instantly and indefinitely.

Unfortunately, reassurance often works only briefly. The anxiety returns, and the person asks again. This strengthens the belief that uncertainty is intolerable and that reassurance is necessary.

A psychologically informed AI tool should recognise this cycle and help the user tolerate uncertainty rather than repeatedly provide impossible guarantees.

8. Privacy and confidentiality

A conversation with a regulated therapist is governed by professional, legal and organisational standards concerning confidentiality, records and safeguarding.

A conversation with a general AI service exists within a different data environment.

Before sharing highly sensitive information, users should understand:

• What information is stored
• How long it is retained
• Whether conversations may be used to improve systems
• What privacy controls are available
• Whether data may be accessed by third parties
• How to delete information
• Whether the service is appropriate for clinical or identifiable information

Avoid entering names, addresses, medical-record numbers or other unnecessary identifying information.

Mental-health apps should explain their data practices in clear language rather than hiding them within lengthy terms and conditions.

9. AI cannot provide genuine human attachment

A chatbot can produce empathetic language:

“That sounds incredibly painful.”

The words may help, but the AI does not feel concern, remember you fondly between conversations or make a human commitment to your wellbeing.

The therapeutic relationship is not simply a delivery system for techniques. For many people—particularly those with attachment injuries, trauma or longstanding interpersonal difficulties—the relationship itself becomes part of the healing.

They experience another person remaining present, noticing them, setting boundaries and responding thoughtfully to moments of trust, anger, shame or withdrawal.

AI may simulate aspects of this interaction, but simulation and reciprocal human relationship are not the same.

Can AI diagnose mental-health conditions?

General AI can explain diagnostic criteria and help someone organise symptoms. It should not be relied upon to diagnose.

Conditions overlap.

Poor concentration could be associated with ADHD, anxiety, depression, trauma, sleep deprivation, menopause, medication, substance use or physical illness. Rapid speech and reduced sleep could reflect stress—or may require urgent assessment for mania.

Diagnosis involves more than matching symptoms to a list. It requires history, duration, severity, functional impairment, differential diagnosis, medical factors and professional judgement.

AI may help someone decide to seek an assessment. It should not become the final authority on what condition they have.

The danger of replacing therapy with advice

Therapy is not simply receiving advice.

People often know what they “should” do:

• Stop avoiding
• Exercise more
• Set a boundary
• Leave the unhealthy relationship
• Challenge the negative thought
• Go to bed earlier
• Stop seeking reassurance

The difficulty lies in understanding why the pattern persists and helping the person change it safely.

A therapist develops a formulation: an individual explanation of how the problem developed and what currently maintains it.

Two people may both procrastinate, but one avoids because of perfectionism while the other is depressed and exhausted. The same generic advice will not necessarily help them both.

AI can generate suggestions. Therapy involves deciding which intervention fits this person, at this time, for this particular reason.

Where AI can work alongside therapists

The most promising future may be a combination of human psychology and carefully designed AI.

AI could support therapy by helping clients:

• Record moods and patterns between sessions
• Practise grounding techniques
• Complete thought records
• Follow a graded-exposure plan
• Remember agreed therapy tasks
• Prepare topics for the next appointment
• Review skills already introduced by the therapist
• Track behavioural experiments
• Identify possible reassurance-seeking
• Receive appropriate psychoeducation
• Maintain helpful habits after therapy ends

Therapists could then use appointment time for more complex work:

• Formulation
• Emotional processing
• Relationship patterns
• Trauma work
• Risk assessment
• Nuanced behavioural experiments
• Addressing obstacles to change
• Repair within the therapeutic relationship

Used this way, AI does not replace the therapist. It helps extend psychological work beyond the session.

What responsible mental-health AI should include

A responsible system should not simply connect a general chatbot to an attractive interface and describe it as therapy.

It should include:

• Clear statements about what the tool can and cannot do
• Clinician-designed content
• Evidence-based psychological principles
• Appropriate risk warnings
• Crisis signposting
• Safeguarding procedures
• Transparent data-storage and deletion policies
• Age restrictions where necessary
• Protection against repeated reassurance
• Monitoring for harmful or delusional content
• Encouragement to seek professional help when appropriate
• Regular testing by people with different backgrounds and needs
• A clear route for reporting harmful responses
• Clinical and technical oversight
• Evaluation of outcomes rather than relying only on positive testimonials

The system should also resist becoming the user’s sole source of emotional support.

Its aim should be greater agency, healthier habits and stronger real-world connection.

Where an app such as Cowch may fit

Clinician-designed wellbeing apps can occupy a useful space between a general-purpose chatbot and formal therapy.

Cowch, for example, is being developed around practical CBT-based principles and the IMAGINE wellbeing framework:

• I – Invest in self-care
• M – Mindfulness
• A – Acceptance
• G – Gratitude
• I – Interactions and relationships
• N – Nurture fun
• E – Explore

The purpose of this type of app is not to diagnose users or replace professional care. It is to help people maintain healthy habits, reflect on unhelpful patterns and access practical psychological tools when they feel worried, overwhelmed or stuck.

The distinction should always remain explicit: a wellbeing app may support mental health, but it is not automatically a medical device, therapist or emergency service.

This is why risk warnings, safeguarding information, privacy policies, testing and evidence-based content matter so much.

How to use ChatGPT more safely for mental wellbeing

Useful requests

You might ask:

• “Help me organise what I want to tell my therapist.”
• “Explain the CBT cycle for social anxiety.”
• “Help me distinguish a real problem from a hypothetical worry.”
• “Ask me balanced questions about this negative thought.”
• “Help me plan a gradual behavioural experiment.”
• “Can we rehearse an assertive conversation?”
• “Help me create a list of questions for my GP.”
• “Explain this psychological term without diagnosing me.”
• “Help me identify what information is fact and what is my interpretation.”

Requests to approach cautiously

Be careful with:

• “Diagnose me.”
• “Tell me whether I should stop my medication.”
• “Guarantee that this symptom is harmless.”
• “Prove that my partner is a narcissist.”
• “Tell me what decision to make.”
• “You are the only one I can trust.”
• “Reassure me again that nothing bad will happen.”
• “Help me avoid telling anyone about my suicidal thoughts.”

AI should support reflection and appropriate action—not become the unquestioned authority over your health, identity or relationships.

Questions to ask yourself after an AI conversation

Before acting on emotionally important advice, ask:

• Has the AI been given enough context?
• Is it agreeing too readily with my interpretation?
• What alternative explanation has not been considered?
• Would this advice still seem reasonable if another person described the situation?
• Does this require professional assessment?
• Am I using the conversation to solve a problem or seek repeated reassurance?
• Is the advice moving me towards real life or deeper into avoidance?
• Could acting on this suggestion cause harm?
• Should I discuss this with a therapist, doctor or trusted person?

Keep your own judgement active.

AI should help you think—not replace thinking.

Will AI replace therapists?

AI will almost certainly change psychological care.

It may improve access, provide support between sessions and reduce the amount of time therapists spend on basic education or administration. Purpose-built systems may eventually deliver certain structured interventions to people who would otherwise receive no help.

But replacing therapists entirely is another matter.

Human therapy includes qualities that cannot be reduced to information:

• Mutual presence
• Ethical responsibility
• Clinical judgement
• Emotional attunement
• Accountability
• Appropriate challenge
• The capacity to tolerate complexity
• A real relationship between two people

The future is unlikely to be simply humans versus machines.

The more useful model is human expertise supported by responsible technology.

Final thoughts

Yes, many of us are already using ChatGPT in ways that resemble therapy.

We disclose, reflect, ask for reassurance and seek explanations for our thoughts, feelings and relationships. Sometimes the response is thoughtful and genuinely helpful.

But ChatGPT is not a therapist. It is a powerful language system that can support reflection, psychoeducation and structured self-help while also being capable of misunderstanding, overvalidating or providing inaccurate advice.

The healthiest approach is neither blind trust nor complete rejection.

Use AI as:

• A thinking partner
• An interactive workbook
• A source of questions
• A way to prepare for professional help
• A prompt for practical action
• Support between—not instead of—human relationships

The aim should not be to create an artificial therapist who becomes indispensable.

It should be to use technology in a way that helps people become more informed, more connected and more capable of managing their lives.