The problem with your AI therapist

You may have noticed some intriguing new language cropping up in conversations with friends and family recently. “Are you sure this isn’t your anxious voice talking?” a pal asked me a few months ago during an episode I will simply describe as a wobble (OK, minor meltdown). “I wonder what your honest voice might say instead.” Heh?

She suggested that I might not be speaking from my wisest place (she was correct). We discussed the possibility that I was just having a temporary freak-out (I was). She reminded me that, when we are feeling anxious, we can be pretty poor at decision-making and generally figuring out whether the thing that’s bothering us is a minor issue or, as the case usually seems at the time, an Extremely Major Issue that is going to make the rest of our lives miserable.

But it wasn’t until after our conversation, when she sent me a screenshot showing two columns full of delightfully straightforward bullet points, that I realised this rather curious “anxious voice” versus “honest voice” dichotomy had come directly from what is surely the world’s largest provider of mental health services: ChatGPT. The anxious voice “speaks from fear” and “wants one definitive answer”, the hive mind expounded, whereas the honest voice “speaks from values” — whatever that means — and “allows two truths to exist”.

It wasn’t an entirely misguided framework, although I found some of it a bit off the mark, including the very idea of an “honest voice”. (Wouldn’t “secure voice” have worked better?) To be fair to the chatbot (as one should always be), I definitely felt better and more level-headed after our conversation. But as I thought more about it, I realised that what had helped me was not so much the AI-generated psychobabble, but having a friend there to speak to me, with all her kindness, uncertainty and imperfection.

What bothered me was how attached my friend seemed to certain specific terms during our conversation, repeatedly deferring to them as some kind of authoritative backup for what she was saying. Because they had come from a machine and not a human being, they were delivered with none of the nuance or hesitancy that a real-life therapist might have spoken with and were therefore somehow being treated as the whole truth.

But the thing about words is that they can never fully capture the whole truth. There are usually several versions of the truth anyway, particularly when it comes to interpersonal dynamics, but the ever-obsequious chatbot forgets to tell you that, giving you too much confidence in your (and its) version. Further, there are certain experiences, feelings and energies that are ineffable and more elusive than the words that seek to tie them down.

And while language might be incredibly useful at helping us make sense of how we experience the world and ourselves and others, it starts to become unhelpful when we become so attached to words and labels that they begin to narrow and to warp our experience of reality. Language can become a barrier to genuine self-reflection and exploration, a way of keeping uncomfortable feelings safely locked up in the brain rather than encouraging further introspection and more embodied forms of healing.

Furthermore, words can themselves have a pathologising effect. Even before the dawn of chatbots, the explosion of interest in mental health issues and the sudden arrival into the mainstream of terms that would have once been reserved for the shrink’s couch had begun to change the way people understand themselves. No, that man who just broke up with you is into you, he just has an avoidant attachment style (and he’s probably a narcissist too). No, you’re not just disorganised, you have ADHD. Indeed, it is so commonplace to be “neurodivergent” these days that it sometimes seems more neurodivergent not to be.

When I went back to my friend to ask her whether she was still finding her chatbot therapist as helpful as before, I was quite struck by her response. Increasingly, she told me, her experience is that it’s “like eating cardboard covered in chocolate — it seems good but then I wake up after three hours and realise I’m not actually that well nourished”. When she speaks to her real-life therapist, she told me, she realises “how not alive the chatbot is, and that feeling alive in a dynamic really matters”.

In a study a year ago, 37 per cent of British adults had used an AI chatbot for mental health support. That figure is almost certainly higher now, and I’m sure it is helping many people. But it’s worth remembering that an AI therapist simply does not possess the kind of “honest voice” that “allows two truths to exist” — and anyway, it can never go beyond its words.

jemima.kelly@ft.com