Are People Choosing AI Instead of Therapy? What Does a Human Therapist Offer Now?

I use AI. I also work as a therapist. So I'm not about to tell you to delete ChatGPT from your phone, throw your laptop into the sea, and return immediately to the reassuring embrace of a therapy room.

AI is useful, sometimes extraordinarily so. And people are increasingly using it to talk about their mental health, understand themselves, and think through their lives. The American Psychological Association's 2026 survey of more than 1,200 psychologists found that 77% had patients who'd talked about using AI, for self-diagnosis, alongside existing treatment, and in some cases as an additional source of mental health support.

So yes, this is happening. And I think therapists need to talk about it, not defensively, not by pretending AI has nothing useful to offer, and not by assuming that everyone having a meaningful conversation with ChatGPT is trying to replace their therapist.

I understand why people are using AI

Imagine it's 11.47pm. You've had an argument with your partner. You're replaying something you said at work. You're wondering whether your childhood was actually traumatic. You've just discovered the term "rejection sensitivity" online and suddenly your entire life makes sense.

A therapist is not available. AI is. No waiting list, no introductory email to write, no fear that your problem isn't serious enough, no appointment to organise. You type why do I do this? and get an articulate answer within seconds.

That's enormously appealing. AI can help people organise thoughts, understand psychological concepts, generate questions to take into therapy, identify patterns, rehearse conversations, and find language for things they've struggled to describe. Therapists don't help anyone by pretending otherwise.

For neurodivergent people, AI can be a thinking partner

There's another use of AI that deserves more attention: sometimes you're not asking it to be your therapist, you're using it to think alongside you.

For some autistic and ADHD people, thoughts don't arrive neatly packaged and ready to communicate. You might have twelve competing thoughts at once, know something feels wrong without being able to name it, or understand something internally but struggle to translate it into words someone else will follow. AI can be an external processing space for that: dump out the messy version, ask it to organise what you've written, turn a page of thoughts into three questions, explore a few interpretations, rehearse a difficult conversation, draft an email and change it six times, ask for something more literal or more concise, come back to the same question from a different angle.

There's also something significant about the lack of social demand. No monitoring someone's facial expression while you process. No working out when it's your turn to speak, or worrying you've explained something for too long. You can stop halfway through, disappear for three hours, and come back. You can ask a question that feels embarrassingly obvious without absorbing someone else's reaction to it. For anyone dealing with executive functioning difficulties, working-memory difficulties, processing differences, or simply needing more time to formulate a thought, that's genuinely useful.

None of that makes AI a replacement for human connection. Used well, it might make therapy more accessible, someone might use it between sessions to organise what happened during the week, work out what they want to discuss, or turn a mass of thoughts into something manageable enough to bring into the room.

So the useful question isn't "is using AI good or bad?" It's "what function is it serving for this particular person?"

Understanding yourself isn't always the same as changing

We have possibly never had more access to psychological information. People arrive in therapy already fluent in attachment styles, trauma responses, nervous systems, dopamine, boundaries, masking, OCD, ADHD, people-pleasing. Sometimes they can give me an extraordinarily sophisticated account of why they behave as they do, and then add, "but I still keep doing it."

Because insight isn't always the problem. You can know your perfectionism developed because achievement got tangled up with safety, know intellectually that you don't need everybody to like you, know your OCD is demanding certainty, know your relationship with food is about control or sensory needs or trauma, and still find that knowing it doesn't change your emotional responses, your behaviour, or your nervous system. Sometimes we don't need another explanation. We need somewhere to do something different.

AI knows what you tell it. A clinician is also paying attention to what you don't know to tell it.

Complex psychological difficulties rarely arrive in neat boxes. Someone might come to me saying "I think I have anxiety," and through a detailed assessment we uncover obsessive-compulsive processes, trauma, neurodivergence, eating difficulties, sensory needs, attachment experiences, burnout, often several of these at once, interacting. The most clinically important part of an assessment often isn't the thing someone initially asked about.

Good psychological assessment is more than matching symptoms to information. It's asking the right questions, noticing contradictions, understanding developmental history, considering risk, watching for patterns over time, knowing when something doesn't quite fit, and being willing to change your mind. AI can generate extraordinarily convincing psychological explanations. Convincing isn't the same as clinically accurate.

AI can also be very agreeable

Most of us enjoy feeling understood, and AI can be exceptionally good at manufacturing that feeling. But therapy isn't supposed to agree with everything you think. Sometimes my job is to validate. Sometimes it's to say, gently, "I wonder if there's another way of understanding that." Sometimes it's noticing avoidance, helping someone tolerate uncertainty rather than handing them another answer, recognising that reassurance is maintaining their OCD, or noticing that something described as "self-care" has quietly become another form of avoidance.

Feeling validated can be wonderful. Feeling better for ten minutes isn't necessarily the same thing as getting better.

And then there is The Good Person Trap®

This is where AI gets interesting to me. I developed The Good Person Trap® to describe what happens when being good stops being simply something we value, and becomes tangled up with our sense of safety, belonging, and worth. It can look like being helpful, agreeable, competent, undemanding, easy to be around. Not taking up too much space. Not inconveniencing people. Getting things right. Anticipating what everyone else needs. Explaining yourself perfectly so nobody misunderstands you. Being deeply uncomfortable at the thought that someone might be disappointed, irritated, or unhappy with you.

Now give that person an endlessly available thinking partner. One they cannot exhaust, cannot inconvenience, cannot bore. One that can't be annoyed by the same question asked twelve times, or burdened by a need for reassurance at midnight, or made uncomfortable by a sudden change of mind. I can see why that would feel incredibly safe.

For neurodivergent people there may be another layer. If you've spent years monitoring how much you talk, formulating your thoughts fast enough, worrying you've misunderstood someone, masking, rehearsing conversations, trying to make your communication more palatable, the absence of those social demands can feel genuinely liberating. I don't want to pathologise that.

But The Good Person Trap makes me curious about something else: what if the very thing that makes AI feel safe is also the thing we sometimes need to practise doing differently with humans? Taking up space. Being repetitive. Not explaining ourselves perfectly. Needing something. Risking misunderstanding. Saying the wrong thing and repairing it. Letting another person have a reaction without immediately managing it for them. Discovering that someone can see the messy, inconvenient, uncertain version of us, and the relationship survives.

AI can tell you that you're allowed to take up space. A human relationship gives you the chance to actually take it, and find out what happens next. For someone caught in The Good Person Trap, that distinction is enormous.

Therapy is a relationship, not an information service

This can sound frustratingly fluffy until you experience it. Therapy isn't information transferred from one brain to another. You bring yourself into the room. So do I. And eventually, some of the patterns that exist outside therapy show up inside it too.

Maybe you worry you've disappointed me. Maybe you automatically tell me you're fine. Maybe you spend the week rehearsing the "correct" answer to my question. Maybe you apologise every time you cry, or become frightened I'll be annoyed with you, or want to disappear the moment you've said something vulnerable. Maybe you get very good at being my good client.

Those moments aren't interruptions to therapy. Very often, they are the therapy, the chance to notice a pattern while it's happening and build a different experience around it. A chatbot can simulate warmth. It cannot take part in a genuinely reciprocal human relationship.

Sometimes therapy also involves actual treatment

Reading about EMDR is not EMDR. Understanding exposure and response prevention isn't the same as completing carefully planned ERP. Knowing that restriction maintains an eating disorder doesn't automatically make eating possible. Knowing reassurance maintains OCD doesn't necessarily make uncertainty tolerable. There's a difference between information about treatment and receiving treatment.

It's part of why my own practice isn't built on the assumption that therapy automatically means 50 minutes with me every Tuesday, indefinitely. Different problems need different things. Sometimes ongoing therapy. Sometimes a detailed assessment and formulation. Sometimes clarity about what's happening and what to do next. Sometimes a focused piece of psychological treatment, or an EMDR intensive. And sometimes my professional recommendation is that someone needs something I don't provide at all. Good therapy isn't about keeping someone in therapy.

Should you tell your therapist that you're using AI?

I think so. I'd much rather hear "I asked ChatGPT about this, and this is what it said", great, let's look at it together. What resonated? What didn't? Did it give you a diagnosis? Did something it said frighten you, or reassure you? Are you asking it variations of the same question over and over? Is it helping you organise your thoughts, or communicate something you've struggled to explain? Or has using it quietly become part of the difficulty we're treating?

For someone with OCD, repeatedly asking AI for certainty could become another form of reassurance-seeking. For someone else, asking it to organise a chaotic page of thoughts before a session might be a genuinely useful accessibility tool. Context matters. So increasingly, I don't think the useful question is "are you using AI?" It's "how are you using it?"

I don't think AI is going to make good therapists obsolete

I do think it will change what people expect from us, and maybe that's not entirely a bad thing. If someone can get generic psychoeducation from an AI system in thirty seconds, therapists need to offer more than information they could have found online: clinical judgement, depth, nuance, accountability, treatment skills, the ability to recognise complexity, the ability to challenge as well as validate, and the capacity to build a real relationship with another human being.

AI may become part of the mental health landscape, in many ways, it already has. Our job isn't to compete with it, and clients don't need to choose between AI and therapy either. AI can be an excellent thinking partner. Therapy is somewhere we practise being a person with another person. The better question is: what am I actually needing right now, and what's the right tool for that job?

Sometimes AI might help. Sometimes a conversation with someone you trust is what you need. Sometimes self-directed resources are enough. And sometimes you need an experienced clinician sitting opposite you, helping you make sense of something considerably more complicated than the question you first typed into the box.

Perhaps, particularly if you're used to being the good, easy, undemanding person, there's something important about occasionally choosing the relationship in which you can inconvenience somebody. Where you can be uncertain. Where you can take your time. Where you can get it wrong, and another human being gets to know you anyway.

Becky Grace Irwing is a BABCP-accredited CBT psychotherapist, EMDR therapist, and Registered Mental Health Nurse. She works with complex trauma, OCD, eating difficulties, and neurodivergence, and is the creator of The Good Person Trap®.

Further reading

American Psychological Association (2026). Patients are bringing AI to therapy: Highlights from the 2026 Chatbots and Mental Health Survey.

American Psychological Association (2026). Health advisory: Use of generative AI chatbots and wellness applications for mental health.

American Psychological Association (2026). Guide to navigating AI-generated advice thoughtfully and safely.

Mental Health UK (2025). Over one in three adults using AI chatbots for mental health support, as charity calls for urgent safeguards.

Samaritans (2026). AI Chatbots: Policy Briefing.