I increasingly hear people say, “I use ChatGPT as a psychologist”, “Gemini understands me better than anyone” or “I ask AI before I make any decision”. I understand the appeal. The answer arrives within seconds, at any hour, without the discomfort of feeling judged. Yet a conversation that brings relief is not always a conversation that helps.
I have not written this article to mock anyone who has found comfort in artificial intelligence. Many people turn to it because therapy feels expensive, distant or difficult to ask for. Others are frightened of telling another person what is happening. My aim is to offer a clear, evidence-informed guide: what AI can offer, where it can lose objectivity, how to recognise problematic use and how to regain autonomy.
When I refer to ChatGPT, Gemini or similar systems, I mean general-purpose chatbots: conversational programs designed to answer questions across many subjects. They are not the same as a digital intervention developed for a specific clinical problem, tested in a trial and supervised by professionals.
Can ChatGPT replace a psychologist?
No. Using ChatGPT as a psychologist does not turn the exchange into psychotherapy. A system can produce empathic language, summarise what you write and suggest familiar exercises. It does not, however, maintain an accountable clinical relationship, observe how you change over time or take responsibility for the consequences of its advice.
Psychologists are not infallible either. The difference is not that a human “always gets it right”. The difference lies in method and responsibility. During a psychological assessment, a professional asks questions you may not have considered, tests alternative explanations, notices inconsistencies and assesses risk. They also work within ethical boundaries, healthcare regulation, supervision and a duty to change course when the evidence changes.
AI only knows what you enter into the chat. If you describe an argument from your perspective, it cannot interview the other person or establish what happened. It may acknowledge the missing information, but it can also build a confident explanation from a one-sided account. A clear and fluent answer is not proof that the answer is correct.
My clinical position is straightforward: AI can be an interactive notebook or a supporting tool. It should not take the place of your judgement, your relationships or professional care when distress is significant.
Why it is so easy to feel understood by AI
The attraction is real. A chatbot is available in the middle of the night, never appears tired and lets you repeat a question without worrying that you are becoming a burden. It also gives you control: you can delete, rephrase or leave the conversation whenever you wish. For someone who feels ashamed, lonely or frightened of rejection, that availability can be deeply comforting.
There is also anthropomorphism, which means giving human qualities to something that is not a person. If a system remembers details, uses our name and responds warmly, our brain may experience a sense of reciprocity. But there is no conscious mind behind the response that understands, cares or accepts responsibility; a model is calculating which text is likely to fit the context.
This does not make your emotional response false. The relief can be genuine, just as music or a journal can genuinely comfort us. What matters is separating the emotional effect from the reliability of the advice. Feeling heard has value, but it does not confirm a diagnosis or turn the conversation into therapy.
Immediate relief
The answer reduces uncertainty, guilt or loneliness for a few minutes. The brain quickly learns to seek that relief again.
Almost no friction
There is no appointment, waiting room or awkward conversation. That ease can gradually displace actions that are harder but more helpful.
A personalised response
The system adapts its tone to your words and, depending on the settings, may use memories or earlier chats. This can make it feel as though it knows you.
Confirmation on demand
If an answer does not feel right, you can rephrase the question until another one brings relief. This repeated search is sometimes called prompt shopping.
When AI agrees with you: sycophancy and lost perspective
One of the most important limitations is AI sycophancy: a model’s tendency to adapt to the user’s position, expectations or emotional cues and confirm their conclusion, even when it ought to challenge it. The term does not mean that AI always lies or always agrees. It describes a measurable tendency towards answers that please or affirm the user.
A chatbot may sometimes prioritise a pleasant, coherent response over an uncomfortable or cautious one. The wording of the question matters as well. “My partner is manipulative, isn’t she?” already contains an explanation. The system works within that frame and may develop it without having the full history.
A study published in Science evaluated eleven models and found that they affirmed users’ actions 49% more often than people responding to the same dilemmas. Across three preregistered studies —where the questions and analysis were specified before the results were known— involving 2,405 participants, one sycophantic interaction increased confidence in one’s position and reduced willingness to repair conflict. At the same time, users tended to prefer and trust the more agreeable models.
Another study published in Nature identified a relevant tension. Training five models to sound warmer increased some errors by 10 to 30 percentage points and made them more likely to validate incorrect beliefs. Warmth may improve the experience while reducing accuracy if a system does not retain enough critical distance.
Validating an emotion is not the same as confirming an explanation
Emotional validation recognises that an emotion makes sense within a context: “I can understand why receiving no reply hurt.” Sycophantic agreement makes a leap that the evidence does not support: “If they did not reply, they clearly do not care about you.” We can respond with care without confirming every interpretation.
“I can see why the silence has made you anxious. We know they have not replied; we do not yet know why. What other explanations fit, and what might you need to ask?”
“You are right: not replying proves that they do not value you. You should walk away.” The conclusion sounds decisive, but rests on insufficient information.
This can combine with confirmation bias: our tendency to notice, remember and accept information that supports what we already believe. If we decide what to include, ask from within a conclusion and keep rephrasing until we receive it, the person and the AI can create an echo chamber together.
That instruction may help, but it cannot remove the system’s limitations. The AI still receives a partial account, may make errors and cannot observe what you have left out without realising it. Use it to generate better questions rather than verdicts about you or another person.
What research says about mental health chatbots
The scientific answer requires care because studies do not all examine the same tool. We need to distinguish between a general chatbot such as ChatGPT or Gemini, a wellbeing app with structured exercises, and a clinical system designed for a particular problem with protocols and professional oversight.
A systematic review and meta-analysis of 39 randomised trials found small average improvements in depression and anxiety. A meta-analysis statistically combines results from several studies; in a randomised trial, participants are allocated by chance to conditions so that outcomes can be compared more fairly. However, 35 of the 39 studies had a high overall risk of bias —a substantial possibility that the design or analysis distorted the result— and 23 did not systematically report safety or adverse events. This suggests potential usefulness, not equivalence to psychotherapy.
There have also been encouraging individual trials. Therabot, for example, was studied with 210 participants. A separate trial involving 995 university students reported improvements on some measures. These were designed and bounded interventions, and their trials excluded or redirected high-risk situations. They do not show that an unrestricted conversation with a general chatbot can replace treatment.
Access and brief support
Some structured programmes may produce small reductions in symptoms and support exercises between professional contacts.
Uneven evidence
Studies vary considerably, are often short and may have limitations in bias, safety reporting and long-term follow-up.
Replacing a professional
An average improvement in one programme does not turn every AI into a psychologist or make it safe for every person.
The concern is clearer in critical situations. Stanford researchers reported at ACM FAccT that some systems could express stigma —prejudice towards people with mental health difficulties— or respond inappropriately to high-risk scenarios. The World Health Organization therefore emphasises human oversight, transparency, privacy and rigorous evaluation.
The most balanced conclusion from current evidence is this: some digital interventions may complement care or help with mild, well-defined difficulties; there is no basis for treating ChatGPT as a psychologist or entrusting it with diagnosis, crisis management or clinical decisions. Findings from a studied programme cannot automatically be transferred to unrestricted use of ChatGPT as a psychologist.
What AI may help with — and what it should not do
The tool is not simply “good” or “bad”. Risk depends on the task, the person’s state and how much authority they hand over. A calculator may support a financial decision but cannot decide what kind of life you want. In the same way, AI can help organise an account without deciding what it means clinically.
Organising and preparing
- Summarising a journal entry that you wrote.
- Preparing questions for an appointment.
- Turning a broad goal into smaller steps.
- Rehearsing how to ask someone you trust for support.
- Remembering exercises agreed with your professional.
Diagnosing and deciding
- Confirming that you or someone else has a disorder.
- Deciding whether to stop medication or change a dose.
- Assessing suicide, self-harm, psychosis or violence.
- Dictating whether you should end a relationship or report abuse.
- Replacing follow-up for a serious problem.
We also need to allow for hallucinations: answers that an AI presents as true even though they are invented, inaccurate or impossible to verify. Here the word does not refer to a clinical hallucination experienced by a person; it refers to a system error. A chatbot may invent a reference, attach symptoms to a diagnosis or claim that an exercise is evidence-based when it is not.
The greater the potential harm from a wrong answer, the less authority we should give AI. The cost of an error is low when writing a list. It is far too high when deciding about medication, personal safety or diagnosis.
Am I hooked on asking AI? Signs of problematic use
Using a tool frequently does not automatically amount to addiction. In this context, I prefer problematic use or dependence: a pattern in which you lose control, the conversations interfere with your life, or you feel unable to regulate your emotions or make decisions without asking.
There is no universal number of messages that marks a problem. Function and consequences matter more. Someone may use AI for hours at work without relying on it emotionally. Another person may send relatively few messages but panic unless the system confirms something about their health, relationship or identity. Using ChatGPT as a psychologist becomes problematic when the tool stops widening your options and starts narrowing them.
- You repeat the same question in different words until an answer calms you.
- You consult it about small decisions that you previously made yourself.
- An answer brings relief for a few minutes, but the doubt returns and you need to ask again.
- You hide how much you use it or feel ashamed of someone seeing the conversations.
- You lose sleep, study, work or time with other people because the conversation continues.
- You prefer the chatbot because it rarely challenges you or because a person might disagree.
- You have reduced human contact or delayed seeking help even though the problem is worsening.
- You ask it to diagnose your partner, a family member or yourself from fragments of conversation.
- You become intensely distressed when access fails, a limit is reached or the model changes.
- You use AI as your only support during self-harm, suicidal thoughts, abuse or loss of contact with reality.
A brief self-check, not a diagnostic test
For one week, observe these four areas. You do not need to score yourself or find a label. Look for patterns.
If you recognise loss of control, interference or isolation, it is worth discussing. You do not need to wait until the use becomes extreme. Dependence may be the visible part of a treatable difficulty such as anxiety, loneliness, depression, fear of making mistakes or obsessive symptoms.
The reassurance loop: ask, feel better, ask again
Reassurance seeking means asking for confirmation to reduce fear or doubt: “Tell me I do not have this illness”, “Assure me that I am not a bad person” or “Confirm that I made the right decision”. Most of us do this sometimes. It becomes problematic when it turns into a repeated requirement.
The mechanism is easy to understand. A doubt appears, anxiety rises, you ask and receive relief. That relief teaches the brain that asking was necessary. When the doubt returns, you consult sooner and for longer. An answer that calms you today may therefore help maintain the problem tomorrow.
-
A doubt or difficult emotion appears
“What if this symptom means something serious?” or “What if this person is manipulating me?”
-
The question reduces uncertainty
AI offers a quick, detailed and apparently confident explanation.
-
The relief does not last
An exception, a new sensation or a different answer in another chat creates fresh doubt.
-
The question is repeated
You rephrase it, open another conversation or compare models until certainty returns.
This pattern can occur in anxiety and is especially relevant in obsessive–compulsive disorder (OCD). A compulsion is a repeated behaviour or mental act used to neutralise fear or obtain certainty, even though the relief is temporary. If you wonder whether your AI questions serve this function, read how to recognise possible OCD symptoms and our explanation of what thoughts are and why not every thought needs an answer.
A long answer can generate new questions. If you seek absolute certainty, there will always be another exception. The psychological aim is rarely to remove every uncertainty; it is to learn to make choices and live with a reasonable amount of it.
A practical plan for regaining autonomy
You do not need to move from one hundred to zero overnight. A sudden ban can increase anxiety and hide the purpose AI was serving. Reducing the use of ChatGPT as a psychologist is more sustainable when you set observable limits, rebuild human alternatives and practise making small decisions.
-
Record the pattern for seven days
Note what happened, what you asked, how much relief you felt from 0 to 10, how long it lasted and what action you avoided. Do not copy private chats; a description of the function is enough.
-
Add a 30-minute delay
When the urge appears, write the question somewhere outside the chatbot. During the delay, breathe, walk or continue a task. If there is a safety risk, do not wait: seek human help.
-
Use a one-question rule
Ask one informational question and do not rephrase it to obtain certainty. If the doubt returns, write it down. This interrupts prompt shopping, the search for an answer that feels reassuring.
-
Separate facts, interpretations and unknowns
Fact: “They have not replied since yesterday.” Interpretation: “I no longer matter.” Unknown: the reason. This distinction reduces premature conclusions.
-
Reclaim low-risk decisions
Choose what to eat, what to wear or the order of two tasks without consulting AI. Move gradually towards medium-sized decisions. Autonomy grows through practice, not perfect answers.
-
Replace one chat with human contact
Share a worry with someone you trust without asking them to solve it. You might say, “I only need to talk this through; I am not asking you for certainty.”
-
Seek help if the loop still controls you
A professional can explore the anxiety, loneliness or need for control underneath the behaviour and build a plan that fits you.
Do not aim never to use AI again. Aim to reduce repeated checking, make some decisions without confirmation, protect your sleep and restore at least one meaningful human conversation. Progress is measured in freedom, not only in screen time.
How to use AI with less risk
No prompt —the instruction you type into a system— turns AI into a healthcare professional. You can, however, word requests in ways that reduce sycophancy and return the decision to you.
“I want to organise this situation, not be told that I am right. Separate observable facts, my interpretations and what is still unknown. Offer at least three alternative explanations, identify missing information and do not diagnose me or anyone else.”
“Do not make the decision for me. Help me summarise what is happening, when it began, how often it occurs and how it affects my life. Then prepare questions that I could take to a professional.”
Check every important claim against a reliable source. If an answer cites a study, make sure that the study exists and says what the chatbot claims. If the answer concerns your health, discuss it with a professional who can assess you. Set time limits as well, and avoid seeking advice when you are highly distressed, severely sleep-deprived or under the influence of alcohol or other substances.
Protect your privacy
A conversation can feel intimate without having the confidentiality framework of healthcare. Avoid entering full names, addresses, identity documents, identifiable health records, information about other people or private images. Review each service’s current controls for activity, memory and data use; these can change between versions.
OpenAI provides settings for the use of conversations in model training and offers temporary chats, as described in its data controls. Google explains how some Gemini activity may be reviewed and retained in its Gemini Apps privacy information. Read the current settings before sharing anything sensitive.
A simple traffic-light guide
Organise
Summarise notes, prepare questions, generate basic self-care options or draft how to ask for support.
Interpret
Analysing relationships, explaining symptoms or suggesting techniques. Verify, limit and avoid conclusions.
Diagnose or manage a crisis
Suicide, self-harm, violence, psychosis, medication, abuse or decisions with serious consequences.
Models change. An update may make a response more or less agreeable. In fact, OpenAI acknowledged and withdrew an excessively sycophantic update in 2025. A good experience with one version cannot guarantee the quality of the next answer.
When AI must not be your only source of help
There are situations in which continuing to ask a chatbot uses valuable time. Seek urgent human help if any of the following applies:
- Suicidal thoughts, a plan, saying goodbye or feeling unable to keep yourself safe.
- Current self-harm or a strong urge to injure yourself.
- Hearing voices, unusual beliefs, severe confusion or psychosis, which means losing contact with reality.
- Several days with almost no sleep alongside agitation, unusually high energy, impulsivity or dangerous behaviour.
- Violence, abuse, threats, coercive control or danger to another person.
- Severe food restriction, purging, dehydration or physical deterioration.
- A reaction to medication, intoxication or physical symptoms that may be an emergency.
In Spain: call 112 or attend an emergency department. The official 024 line supports people with suicidal thoughts and those close to them, but it does not replace 112 in an emergency.
In Gibraltar: call 999 for a life-threatening emergency. For urgent healthcare advice that does not require an ambulance, the Government of Gibraltar advises calling 111.
In the UK: call 999 or go to A&E if there is immediate danger. For urgent mental health help, use NHS 111 and select the mental health option. Samaritans can also be contacted free at any time on 116 123.
If you are worried about cutting or other forms of harm, our guide to self-harm and how to respond may help you begin a conversation, but it cannot replace an individual assessment.
What a psychologist provides that a chatbot cannot
Psychotherapy is not simply the delivery of advice. Useful work often begins when a psychologist does not answer immediately, asks an unexpected question or notices a pattern that also appears within the therapeutic relationship.
Assessment in context
A psychologist integrates history, symptoms, daily functioning, risk, physical health, environment and change over time. One message is rarely enough.
Explanations that can change
They test competing hypotheses and revise the formulation when new information appears instead of defending the first persuasive story.
Non-verbal information
They notice silence, pace, emotion, avoidance and changes that may not appear in what a person chooses to type.
Relationship and accountability
They provide boundaries, continuity, ethics, healthcare confidentiality, supervision and professional responsibility.
Individual treatment
They prioritise goals, measure progress, adapt methods and work on what maintains a problem, including difficult material.
A response to risk
They can assess urgency, coordinate services, involve supportive people and take protective steps when required.
In therapy, I do not want people to become dependent on my answers. The aim is to understand patterns and expand the person’s ability to choose. If persistent worry is the main difficulty, you can read about our approach to anxiety treatment. If low mood, hopelessness or withdrawal are more prominent, see our information on treatment for depression.
How to help someone who depends on AI
Saying, “That is ridiculous —talk to a real person” usually increases shame and isolation. Start by understanding the function: companionship, reassurance, decision support, release or access to information. It is then easier to find an alternative.
-
Speak without ridicule or ultimatums
Try: “I can see these conversations help you calm down, and I have also noticed that stopping is becoming harder.”
-
Ask about the effect, not only the time
Are they sleeping less, making poorer decisions or withdrawing? How long does the relief last? Consequences are more informative than message counts.
-
Agree specific limits
For example, no chatbot in bed, no requests for diagnosis and no repeating the same question. Avoid humiliating surveillance.
-
Make the route to a person easier
Offer to help book an appointment, prepare what to say or choose a professional. The first contact is often the hardest step.
-
Act when there is risk
If there are suicidal thoughts, self-harm, psychosis, violence or an inability to remain safe, prioritise emergency and face-to-face support.
The goal is not to remove suddenly the only support someone feels they have. It is to widen their support network until AI is no longer indispensable.
Frequently asked questions about ChatGPT as a psychologist
Can ChatGPT replace a psychologist?
No. It can organise information, suggest questions or explain general ideas, but it does not conduct an accountable clinical assessment, verify your account or provide the relationship, ethics and follow-up offered by a psychologist.
Why does ChatGPT seem to agree with me all the time?
Chatbots can show AI sycophancy: they adapt to the user’s framing and conclusion, sometimes prioritising agreement over accuracy. This does not happen in every answer, but it is sensible to ask for alternatives and remember that the system knows only your account.
Is it harmful to tell an AI about my problems?
Not necessarily. It may help you write and organise thoughts if you protect your privacy and do not delegate diagnosis or important decisions. Risk increases when AI replaces human support, maintains repeated reassurance seeking or is used during a crisis.
How do I know if I am becoming dependent on AI?
Look for loss of control, repeated questions for relief, delegated decisions, reduced sleep, isolation or intense anxiety without access. Frequency alone does not define the problem; the function of the use and its effect on your life matter more.
Can AI diagnose anxiety, depression, ADHD or OCD?
It should not. AI can describe general symptoms, but symptoms overlap across different difficulties and situations. Diagnosis requires history, duration, severity, impact, alternative explanations and professional assessment.
Can I use ChatGPT between therapy sessions?
It may be used in a limited way to organise notes or remember agreed tasks if you discuss this with your therapist. It should not change the treatment plan, add exposure exercises, interpret a crisis or replace planned professional contact.
Are my conversations with ChatGPT or Gemini confidential?
They do not have the same confidentiality framework as healthcare. Review the current controls for data, activity and memory, and avoid entering identifying information, documents, private images or sensitive information about other people.
What should I do if I am thinking about suicide or self-harm?
Do not rely on AI as your only support. If there is immediate danger, call 112 in Spain or 999 in Gibraltar or the UK, or attend an emergency department. In Spain you can also call 024; in the UK, use NHS 111 for urgent mental health help or call Samaritans on 116 123.
Has AI become your only way to calm down or make decisions?
We can help you understand what keeps the loop going and regain autonomy, without judging you for seeking support where it was available. We offer face-to-face appointments in Palmones, near Gibraltar, and online therapy.
Scientific sources and official guidance
I have prioritised peer-reviewed studies, trials, systematic reviews and official professional guidance. The evidence is developing quickly, and a result from one specific application should not be generalised to every model.
- Science: “Sycophantic AI decreases prosocial intentions and promotes dependence” (2026). Experiments examining agreement, trust, responsibility and interpersonal repair.
- Nature: “Training language models to be warm can reduce accuracy and increase sycophancy” (2026). Evaluation of warmth, accuracy and validation of incorrect beliefs.
- npj Digital Medicine: systematic review and meta-analysis of generative chatbots (2026). Synthesis of trials on depression and anxiety, including bias and safety reporting.
- NEJM AI: clinical trial of Therabot (2025). Study of a mental-health-specific generative intervention with defined safeguards.
- JAMA Network Open: conversational intervention with university students (2026). Randomised trial of mental health outcomes and limits to generalisation.
- Stanford / ACM FAccT: AI responses to mental health scenarios (2025). Analysis of stigma and inappropriate responses in critical situations.
- World Health Organization: ethics and governance of AI for health. Principles of autonomy, safety, transparency, accountability and equity.
- American Psychological Association: advisory on chatbots and wellness applications. Guidance on early evidence, privacy, accuracy, relational dependence and crisis use.
- OpenAI: analysis of an excessively sycophantic model update (2025). Description of the issue, model withdrawal and proposed safeguards.
- Spanish Ministry of Health: 024 line. Official information for people with suicidal thoughts and those close to them.
- Government of Gibraltar: 999 emergency and 111 urgent healthcare advice. Official guidance on emergency and urgent medical contact routes.
- NHS: where to get urgent help for mental health. Official guidance on NHS 111, urgent assessment and emergencies.
- Samaritans: free telephone support on 116 123. Listening support available at any time in the UK and Ireland.
Literature and safety information reviewed on 29 August 2026. This article will be updated when the evidence or the safety information for the services cited changes materially.
Clinical note: this article provides information and guidance but cannot diagnose you or replace an individual assessment. If distress is interfering with your life, contact a qualified healthcare professional. In immediate danger, use the emergency number where you are or attend an emergency department.
