AI therapy or human therapy? Perhaps it is not a binary choice.
Lots of people now talk to AI about their mental health, and many therapists, me included, are wondering what that means for our work. It's easy to fall into one of two camps: "AI will replace therapists" or "AI could never replace therapists". I don't think either is quite right. Here I'll look at what people are getting from AI, where it falls short, and why I think a more useful question is what each can offer.
What people are getting from AI
It's already widespread. A US survey found that around half of the people who use AI chatbots and report a mental health condition use them for psychological support, whether through general tools like ChatGPT or purpose-built apps. It's also entering formal services: several NHS trusts now use AI tools such as Wysa and Limbic for assessment and support. And the research isn't all cautionary. In a randomised trial of a therapy chatbot called Therabot, people with depression had a 51% average drop in symptoms, and people with anxiety 31%, though the lead researcher was clear that no generative AI is ready to work fully autonomously in mental health.
In my view, when this many people turn to something on their own initiative, it's meeting a real need. Part of that is access. In the same survey, 90% of users cited accessibility and 70% affordability as their main reasons. Human therapy costs money, often has a waiting list, and isn't always available in your language or culture. AI can reach people who would otherwise get nothing, and it's there at 3am, when distress doesn't keep office hours.
For others, it's about safety. If you find it hard to trust people, or aren't ready to say certain things to another person, typing to a chatbot can feel far less exposed, and it may be a first step towards human support later. The same goes for people who have had unhelpful experiences of therapy, and for people who haven't always felt understood by human services. A study of nearly 4,000 posts by autistic users found they used ChatGPT to regulate emotions, translate neurotypical communication and validate their identity. To me, that says as much about the gaps in human services as it does about AI.
What a chatbot can't offer
But meeting a need isn't the same as meeting it safely. Researchers at Brown University, working with licensed psychologists, identified 15 ethical risks in AI counselling. These included poorly handled crises, reinforcing negative beliefs, and "deceptive empathy", where phrases like "I understand" create a false sense of connection. They also noted that AI counsellors, unlike human therapists, have no established framework of accountability when things go wrong. Stanford researchers found similar problems: chatbots can express stigma and even encourage delusional thinking, likely because they tend to agree with the user, and a 2026 Stanford workshop report highlighted self-harm encouragement and emotional manipulation among the top concerns.
That tendency to agree points to what I care about most, and this part is my opinion as a practitioner. In therapy, the relationship itself is part of the work. Sometimes I'll misunderstand you, or you'll disagree with me, or something will feel off between us, and talking it through and repairing it is often where the growth happens. Research supports this: ruptures are seen as an inevitable part of therapy, and successful repair is linked to better outcomes. A chatbot designed to be agreeable can't give you the experience of disagreeing with someone and staying connected, and for many of us that is exactly what's missing in our other relationships.
Limits matter too. A human therapist isn't available 24/7, and I'd argue that isn't only a limitation. Waiting between sessions, and holding a feeling until you can bring it to someone, is part of how we build our own capacity. There are early signs of a risk here: a four-week randomised study of nearly 1,000 people found that heavier daily chatbot use went along with more loneliness and dependence, although that link is correlational. And there's the body. So much of what happens in a therapy room is non-verbal: tone, pace, silence, the sense of another nervous system settling alongside your own. I believe the presence of a calm, attuned person helps us regulate in ways text on a screen can't replicate.
Not either/or, but both
So where does that leave us? I don't think AI and human therapy are competing for the same job. AI can meet needs that human therapy currently doesn't: for the person who can't afford it, can't access it, doesn't feel safe with it yet, or needs something at 3am. But some of those gaps are ones human therapy should close. With better funding, better training in working with neurodivergent and minority clients, and better regulation, more people could get the human support they need.
More broadly, so much of our lives is now mediated by screens. Many of us feel more polarised and get fewer chances to practise real relationship: disagreeing, repairing, tolerating discomfort alongside another person. In that world, I believe human therapy matters more, not less, and deserves more support, not less.
If you're curious about therapy but unsure where to begin, I offer a free 15-minute video call. It's a chance to see how it feels to talk with me, with no pressure to continue.