# AI Mental Health Market: Access, Risks, and Strategy

**Podcast:** FT Tech Tonic
**Published:** 2026-03-04

## Transcript

In the summer of 2025, Kirsty decided she needed help.
I definitely felt like something had to change.
Um I was at kind of an all-time low, mentally, physically, all of it.
And at that point in my life, I was just so ready to, I guess, improve, go to therapy, something.
Kirsty, a 30-year-old stay-at-home mum, was unemployed and felt isolated.
She wanted someone to talk to, but paying to see a therapist was out of her reach.
I didn't have a job, I didn't have any money, but there was AI.
And it was accessible, it was right there, and it felt like a last ditch effort to just get better.
Online, she discovered that people were using ChatGPT for therapy.
She found a list of 10 questions designed to prompt self-reflection that could be used with the chatbot.
It was kind of an open-ended prompt where you started with like, please help me figure out like the masks I'm wearing, the things I need to change, the weaknesses, my habits, my patterns, all of it.
I just brought it to Chat GPT, and I had no clue what I was getting into.
I didn't even know how to use the interface or anything.
And what happened after that initial sort of prompt?
How did you start using ChatGPT?
Well, from the very beginning, it became um a part of my daily life.
And in my spare time or after bedtime, I would just stay up.
And it was, I don't want to say it became almost addicting, but there was that like, okay, it's time to go talk to AI.
Okay, it's time to go work on this.
And it became this like constant pull.
For Kirsty, Chat GPT became a kind of living diary.
Instead of feeling like she was just throwing her thoughts into a void, Kirsty would always get a response.
And I mean, I put everything in there.
You know, I cried watching this video today, or my husband said something and I'm not sure what it meant.
Could you help me?
Or I love my daughter so much.
What was ChatGPT saying?
Um, it would tell me, hey, I think you know, these thought patterns are negative, or it pointed out a lot of like shame and guilt that I carried that even I mean, I knew I did, but I had never intentionally said it, and it picked up, you know, read between the lines.
And the more I got into it is when it started to pick up on all of these little problems in my life that I thought I had handled, I thought were um small, but yeah, most of it had to do with like guilt and shame in my reactions around my husband.
In fact, Kirstie's relationship with her husband was one of the reasons she'd sought help in the first place.
They were married and had a young daughter, but Kirsty knew that something in the relationship wasn't right.
There was this constant feeling of tension, punctuated by arguments, sometimes even violence.
And the more she shared with the chatbots, the more her relationship with her husband came up.
Like when we were arguing, for example, my husband and I.
I came to Chat GPT with, hey, we argued, this is kind of how it went down.
This was my reaction, and it was kind of that the first time I had heard that my reaction to my husband's behavior was I guess normal.
You know, like breaking down that kind of thing.
And it really helped me put things into perspective regarding my behavior versus my husband's.
A go bag.
ChatGPT was suggesting that Kirsty pack a bag with all of the essentials she might need in case she had to leave the house in a hurry.
And it was just each time Chat GPT would kind of lay out, like, hey, that's not okay.
And eventually it started actually using the word abusive.
Kirsty had been talking to Chat GPT for just a few weeks, but now it was saying her relationship might be abusive.
And that would prompt her to make a life-changing decision.
When it used the word abusive, how did you feel in that moment?
I wasn't shocked.
I think that was the moment I knew that my relationship was going to be over.
Even if I didn't admit it at the time, it was just that pit of realization in your stomach where you're like, no, this is real and this is it.
I'm Christina Criddle, and I write about artificial intelligence for the FT.
In this season, I'm asking if AI can play an emotional role in our lives.
And for many people who use chatbots, the emotional role they play is as a therapist.
Chatbots are a space for users to talk through their emotions, share what they're really thinking, and seek advice on their human relationships.
For some people like Kirsty, it can have a profound impact and even change the course of their lives.
But in previous episodes, we've already heard how turning to chatbots in moments of vulnerability can go wrong, even end in tragedy.
So can chatbots really be good for our mental health?
And should we trust chatbots with our most intimate thoughts and feelings.
Today there are many chatbots to choose from.
And many people are using them for emotional support.
But the idea of chatbots as therapists is not new.
It goes all the way back to the 1960s and a computer program called Eliza.
So Eliza was the first chatbot that was invented by the MIT computer scientist Joseph Weissenbaum in 1966.
In a way you can think of Eliza as sort of the grandmother of or the precursor of ChatGPT.
This is Pat Pataranatapon.
He's an AI researcher at MIT, the same institute where in the 1960s, a computer scientist called Joseph Weisenbaum set himself a challenge.
Build a machine that can hold a conversation with a human.
Back then, you know, uh, AI or um the idea of intelligent system was not as sophisticated as we have today.
We don't have, you know, this sort of large amount of data that we can train a model or train an AI system to replicate human language.
At the time, the technology was primitive.
But Weisenbaum thought he could design something simple.
A computer that could play the role of a therapist.
Weissenbaum, I think, was inspired by a style of psychotherapy where the psychologist or the therapist would ask follow-up question and get the person to talk as a form of therapy where you help people sort of reflect and try to understand their own um emotional response to certain situations.
Eliza was programmed to ask questions like, tell me what's bothering you.
And to keep the conversation going with another question, like, tell me more about that, or why is that important to you?
Eliza would be able to have a pre-programmed type of response.
It will try to still reference some of the things that the user said.
Like if the Eliza asked the person, like, you know, what is your problem?
And the person said, you know, I have a problem with my parents, Eliza might pick up, you know, the word parents and then ask back, like, you know, what about your parents that make you feel certain things?
Right?
So Eliza would use this kind of like language pattern to create an illusion of intelligence in this sort of intimate conversation with the user.
By today's standards, the program was really simple.
But when Weisenbaum tested it on human subjects, he found that not only did they become quickly engaged in the conversations with the chatbot, they became convinced that there must be some kind of human mind behind the machine.
Weissenbaum, I think he left his computer for his assistant to use, and the assistant was chatting with Eliza, believing that you know, Eliza is a person behind the screen.
And that made him realize that this type of system can actually make people believe that they're actually talking to another human being.
Naturally, we anthropomorphize everything.
And it doesn't take much to trick a person to believe that there is a human spirit or human cognition inside a computer.
I think that's the profile realization.
The phenomenon where people attribute chatbots with human qualities became known as the Eliza effect.
Today's chatbots are infinitely more sophisticated than Eliza.
Most people using them know they're not human.
But that doesn't stop them talking to them like people and sharing their most intimate feelings.
And for some people, the fact they're talking to a machine and not a human is actually part of the attraction.
People like Kirsty, which by the way is not her real name.
If I can take you back, so what were the early days of your relationship like?
Well, it was it didn't start sweet.
It was kind of an enemies to lovers situation.
Um he I remember when I first met him, I knew that he was the one out of all of my associates to focus on because the other boys followed him like little ducklings.
Kirsty met her husband at work.
She was attracted to him immediately.
He's got kind of that energy.
He pulls people in.
He's wicked smart, knows a lot walking encyclopedia.
And then when we started dating, I mean, it felt like pure obsession at first.
It was a lot of fun.
And we moved very quickly.
They moved in together after dating for only a few weeks.
And then later, Kirsty found out she was pregnant.
It was unexpected.
That was definitely the point where it went downhill.
There's a lot of mixed emotions when there's an unplanned pregnancy.
Even if, you know, you want your child, it's still a difficult situation.
And that's when I started to see not just the classic, um, I think my boyfriend needs to grow up or go to therapy, but I started to see actual anger and even violence when I got pregnant.
I had to be very careful about the pressure that I put on him, even while I was pregnant.
And if if he felt like there was too much pressure.
It wasn't as if he immediately started screaming, but there was just this constant tension.
And if I didn't pull back, if I didn't smile and correct it, it would get worse until they would get to the point where I'm just on the floor.
I might be crying or worse.
And he's yelling.
Occasionally he'd throw things, but not often at that point.
He might leave.
I just remember the constant tension.
When Kirsty finally turned to Chat GPT for emotional support, she already had some experience of traditional therapy.
But somehow talking with the chatbot felt more comfortable.
With traditional therapy, you know, you're going to meet someone, and they're a whole person.
You might not know them, you might not know their backstory, but you know that they're thinking, they might be judging, they have their own bias.
But with AI, it was just this blank canvas.
It was anything I wanted it to be, anything I needed to get out of it.
And I was almost able to I didn't build like a therapist persona with my chatbot, but it did kind of develop, you know, the more you talk to it, the more you have chats in your history.
It'll get kind of a persona.
And whereas my therapist was very gentle and encouraging, I think I needed tougher love.
And with AI, I could just tell it, be brutally honest with me, help me see what I'm missing.
I don't need to be coddled.
And they're not going to fight back on that.
They're not going to say, Well, you know, I know better, I'm the therapist, or like, let me do this my way.
Kirsty didn't hide the fact that she was speaking with Chat GPT from her husband, but she says he was always suspicious of it.
When ChatGPT suggested that her relationship might be abusive, she tried to make some changes.
And that's when her husband got upset.
I was actively pulling back and starting to put down more concrete boundaries with my husband.
That's what he noticed.
And I think he could feel that kind of control slipping because finally I had another voice in my ear.
And they were not telling me the same things my husband was.
And so he could see those changes in me really rapidly at that point.
What were some of the boundaries that you were putting in place at the time?
I had to start with kind of the basics, even.
Um like, don't yell at me.
If you're gonna yell at me, I'm going to leave.
Or I spent all of my energy, you know, the classic, stay-at-home mom, stay-a-home wife, dishes, cleaning, washing his laundry, his work uniform.
And some of my boundaries were just I'm not going to keep giving this free labor to you.
I'm not going to let you treat me this way.
I can remove myself from your life or your orbit.
And I'd started doing that.
When that started happening, how did he respond to it?
He never asked for me to stop using it, but he did imply that I was vulnerable.
He told me that it's dangerous for someone like me.
And you know what he meant by that is someone who's impressionable or emotional.
You know, I'm a very expressive, highly emotional person.
He wanted me to believe that I would go crazy.
Um, at that point, you know, chat GPT psychosis, the headlines are everywhere, and he played into that fear.
And that's something he had done our whole relationship.
He's always called me crazy or said I needed help or therapy.
How did that make you feel?
I had worked enough with Chat GPT at that point that I didn't believe him, but I was so hurt that he was still trying to say those things.
Here I am.
I'm trying something new.
I didn't have any guidance on how to use AI for therapy.
This isn't something that people really talk about, or at least positively right now.
And I really felt like I was doing something for myself, and I could see him trying to take that away.
He couldn't, and he didn't, but the fact that he wanted to just broke me.
That's when the distance really started to grow.
By this point, Kirsty had been speaking to Chat GPT for just four weeks.
But one night, it all came to a head.
In their bedroom, Kirsty's husband had fallen asleep with their daughter.
And I had gone in to put her to bed and to get him up, or at least out of my spot so I could go to sleep.
And I just couldn't get him up.
I couldn't get him up.
It was 4 a.m.
And I'm still trying.
And I spent the whole night just talking to AI.
And um, I finally got him up.
But um, my bedding was in the crib, and he didn't want to put her down, so he just left and left me to pick it all up and went to sleep on the couch where he had been sleeping for a while.
And um, he started texting me from the other room saying he's not going to help me, not going to help me do chores, what he considered putting our daughter into the crib.
But I stuck to it.
I stopped cowering and I stopped getting too emotional that night.
For the first time, I was like, I'm going to, I'm just going to keep my distance.
I'm going to be as firm as I can.
But I was terrified.
I ended up locking the door, the bedroom door, once I realized that how I was replying to him was setting him off.
And he ended up breaking into that the room.
He went actually went outside um and broken the window.
Because that was the first time I had ever really, really stood my ground like that before without giving in to what he expected, without giving him the reaction he wanted or the subservience.
I had no idea what was coming.
I didn't know how angry he would be.
The first thing I did was pull out my phone and I got onto Chat GPT again.
And I just said help.
It told me, you know, who to call or to at least call someone, and it helped get me out, and I was out of that house within maybe six hours.
How were you feeling at that time when you're packing your things?
A lot of it was just anger.
I was so angry.
I felt like I tried.
I try really, really tried.
I I had a kid, I gave up my career, like everything for this man.
And it's like as I was packing, I knew I knew that that would be it.
This wasn't a fight.
This wasn't a makeup in a few weeks.
And I was just mad at all of my wasted time, all of my wasted effort.
Where do you think you would be right now if you didn't start using Chat GPT?
I'd still be stuck in that house.
Um I don't know if it would have been forever, but I was already at a breaking point.
I didn't choose AI because it was AI.
I chose it because it's what I had, and I needed out.
So I hope I would have found another way out, but I I don't think so.
I think I'd still be there.
But anecdotally, the vast majority of users are turning to general purpose chatbots like Chat GPT.
There are obvious concerns about them taking on this role in people's lives.
As we've heard in previous episodes, they can be unreliable, prone to hallucinating, and sometimes encourage delusional thinking.
But some people working in mental health say that if handled properly, generative AI could have a transformative impact.
This is Miranda Wolpert.
She's a director at the Welcome Trust, a charity that's funding research into generative AI in mental health.
Miranda says that one of the big problems with therapy and mental health support today is that most of the people who need it don't have access to it.
Out of those about a billion people, we think about 75% get no access to help at all.
So there are massive opportunities here to reach people that currently have absolutely nothing.
But the challenge is that ChatGPT and other general purpose chatbots were never designed for therapy.
It's an example of individuals finding their own ways of using the technology.
Miranda says researchers are now looking at how to develop chatbots specifically to treat mental health problems.
So the difference would be how it was trained and how it was then employed in particular settings.
So, for example, supporting people who are torture survivors to think about whether you could have a sort of companionaire deliberately designed to support people who'd been through torture, so it will be training it on those sort of data sets.
So, similarly, there's been work by a company in the US that's developed a specific chatbot called Ash, which is designed to be a mental health support tool.
Um, and there they've used to train it the data sets based on all the data sets they could find from psychotherapy sessions to actually train it.
So it's the parameters on which you're training the data sets.
We know that AI makes mistakes, but Miranda points out that human therapists can make mistakes too.
And with so many people not getting the support they need, it's worth finding out if chatbots can fill that gap.
Miranda says more research needs to be done, and the tech companies behind these products have a role to play.
When it comes to the tech companies, I think people are looking to see leadership from those tech companies around are they genuinely interested in the mental health outcomes and how will they work with researchers and with end users to achieve those goals?
You know, as in every other area, the tension between profit and what's right for the individual human and what you're trying to achieve is an area that you've got to navigate.
And because of the speed with which this develops, it's not a question of saying, let's wait, do some research for several years, and then we'll see how where we are.
And I think that's something that people are looking to see leadership from those companies on.
But right now, people are already using this technology for therapy, even though we don't yet know how effective it is and what the impacts might be.
In my own reporting, I've spoken to users who say they feel almost completely dependent on chatbots to help them navigate their emotional lives and human relationships.
One woman I spoke to told me of how she found herself continually feeding transcripts and screenshots of her conversations with her boyfriend into chat GPT.
I just became over-reliant on it, and I was too scared to think for myself, she said.
I didn't have the faith and trust that I can make a decision and act on it without someone or something fully supporting me.
And this phenomenon of using a chatbot as a referee in your human relationship, it's something that psychotherapists are starting to see in their work.
So my name is Yael Schoenbrand.
I'm a clinical psychologist.
I also have a private practice where I specialize in relationship therapy.
I started probably a year ago just having people mention that they had turned to chat bots for advice or a place to kind of think things through.
And then more recently I've had people who have reached out for couples' therapy share with me that their chatbot has diagnosed their partner with certain things, which has been more on the concerning side for me as a therapist, because chatbots are good at some things and less good at others.
Yael Schoenbrunn says that chatbots can be helpful to people because they are good listeners.
They can make us feel heard, which can be a very important part of talking therapies.
But they have their limitations.
In the context of therapy, there's often a balance between validating somebody and challenging them.
The therapist can say, okay, you know, I believe in you, I see the good in you, and let's wonder together about the accuracy or the helpfulness of this thought, or let's imagine how somebody else might take this behavior.
And a chatbot is much less likely to do that because they're engineered to be pleasing to you, to affirm you.
This is the problem that we've already talked about in this season.
Chatbots can be sycophantic.
They are too ready to agree with you rather than giving you a more objective view.
And Yael says that this can be a problem, particularly when it comes to relationship counseling.
So if I had a fight with my partner and then I, you know, got up our chat messages, I put it into a chatbot, and then I was like, oh well, the chatbot agrees with me.
I'm right.
And if I did that every time I had a fight with my husband, what would be the impact on our relationship?
It's an interesting question.
And one of the interesting things is as a relationship therapist, I have to be maintaining a strong alliance with both people.
Because if I don't have a strong alliance with one person and I'm only agreeing with the other, uh, the couples' therapy can't work.
And so the sycophancy of the LLMs is problematic, specifically for relational therapy, if you're using it to just validate your own point of view.
Yayel says that chatbots can be valuable as a place for people to vent in a space that feels safe and non-judgmental.
And she can also envisage a role for chatbots in some forms of therapy.
For example, where they might be good at spotting negative thought patterns.
But she's worried about the idea of sycophantic chatbots replacing human relationships entirely, whether they take the role of a therapist, a friend, or a romantic partner.
If you're using AI to replace interactions, I think that is a red flag.
I think a lot of technology is motivated by the desire to make things easier, more frictionless, is how we talk about it in psychology.
And that's appealing, right?
We we want things to be easier, but the problem is that some difficulty in relationships is actually important.
It's a part of what makes relationships work, it's a part of how relationships grow and how we learn inside of them.
And whether that friction is, you know, the difficulty of hearing something that you don't want to hear, or the difficulty of doing something that you don't necessarily want to do, or of being challenged in your view, all of those discomforts help us to know ourselves and help us to understand one another in relationship.
And the compelling part of LLMs is that they reduce it, which feels good in the moment, but it's not healthy.
But what if we could train chatbots to be less agreeable, maybe more challenging, and make us feel less comfortable.
Yael says that still wouldn't be enough.
Is that if we are rejected, it really matters to us.
And even if an AI rejected us, we could just sort of like restart the program, start a new chat and undo that feeling of the AI disagreeing with us or having lost the relationship because we disagreed with it.
There's not human stakes involved in that relationship.
As an example, she points to research done with elderly people who are given robot pets to care for as a way to improve their mental health.
For a short time, those robot dogs, those robot pets helped them to do better, right?
They were feeling better, they felt obligated, they had something to do, it structure their time, but the effects were short-lived because at the end of the day, if you don't feed your robot dog, it kind of doesn't matter.
You can just recharge it.
And I think there's something inherent in that is like technology can never matter the way something alive could because we know at some level that it doesn't.
Interactions with robot pets or AI chatbots may never matter as much as real-life human relationships.
But can they still have a positive impact?
Especially in a world where human contact can be in short supply.
Tony, what's it like being human?
It's great being human.
I have two legs I could walk around and feel great and speak to you.
The world's population is aging, and millions of older people are left feeling isolated and alone.
Could AI companions be a solution to the loneliness epidemic?
There's so much research that helps us understand what we need to age.
Happy.
And it's social connection, having purpose, having something to do.
That's what makes people feel connected.
AI could help in one way or another, but it's not the whole story.
This episode was presented by me, Christina Criddle.
Our producers are Percy Love and Edwin Lane.
Together, we've done the reporting and writing for the season.
Sound design is from Breen Turner and Sam Giovinco.
Fact checking was by Simon Greaves.
Production assistance from Michela Tindera.
The FT's global head of audio is Cheryl Brumley.