AI and mental health: “it could help revolutionise treatments”
Professor Miranda Wolpert is Director of Mental Health at Wellcome. Here, she gives her insights into the opportunities and risks of AI in mental health – and why we must approach the potential of AI with curiosity and not assumption.
Credit:
Licence: All Rights Reservedpocketlight via Getty Images
11-minute read
11-minute read
Elapsed time:00:00Total time:00:00
Pressing play on the video above will set a third-party cookie. Please read our cookie policy for more information.
Watch our explainer video, where we dive deep into the conversation around AI therapy – joined by Professor Miranda Wolpert, Dr Joyce Nakatumba-Nabende and Shuranjeet Singh.
Okay, so… would you use an AI therapist?
Artificial intelligence and therapy actually goes way back to the 1960s. Meet ELIZA, the world’s first ever AI chatbot. Written with just 420 lines of code, it was designed by this guy to mimic the language of a psychotherapist. Sixty years later, chatbots have come a long, long way. Now they can edit passive aggressive emails to sound more friendly, plan holidays in seconds, and DJ your party on command, as well as acting as your 24/7 therapist.
But let’s be clear. AI therapy is just one of the many stages in mental health’s digital revolution. So much has already happened. There are algorithms that analyse speech to spot early signs of psychosis. Apps that use a simple game of Tetris to help your brain reprocess a traumatic memory. We’ve even got digital avatars that help patients process the voices they hear.
Today, the brains behind AI are promising less admin for therapists, improved training, and perhaps most promisingly, easier access to mental health care. Sounds impressive, right? Well, when it comes to AI and therapy, there’s a lot of different views out there.
So most therapy is a conversation. And conversation is something AI is getting pretty good at. The more powerful models are trained on tonnes of existing written material, identifying common ways each word is used, using that data to piece together a reply to your message. But not every mental health chatbot works the same. One might scan for super‑specific words, like a symptom you typed out, to trigger a reply. One might be slightly more rigid, giving you a list of messages to pick from, each with a prewritten response.
When used for therapy, some people have mentioned how they like the warmth, companionship, and the advice of the chatbot. One described their conversation as a safe space. This study found that an AI model trained on medical data has the potential to help people reframe their negative thoughts. Another concluded that chatbots are just as effective as human therapists for those with depression, anxiety, and eating disorders. However, not everyone has experiences like this. Some research found that others are more frustrated, saying their conversations feel empty, generic, and not knowledgeable enough.
Those with lived experience of mental health conditions have already questioned whether replacing a human therapist with a chatbot could increase isolation, which is exactly what one study by researchers at MIT and OpenAI found. To add to that, most AI chatbots are trained on English language data, often from Western contexts. This raises questions about how well they serve people from different cultures, or those who don’t speak English as a first language.
If training data has a narrow cultural view, AI might ignore important nuances and contexts in your messages, or reinforce harmful stereotypes around race, gender, and sexuality, which could widen existing inequalities. And of course, it’s not just chatbots. Your human therapist can hold these biases too. Either way, whether it’s AI or face to face, those developing the therapy must always work closely with the communities they want to help. They also need to collaborate with ethics experts to make sure the training is inclusive and the support is safe.
If we do this properly, AI could be part of the solution to offering a more effective, consistent approach to mental health care, like one chatbot currently used by the UK’s National Health Service. So far, it’s already making a difference, with faster referral times and nearly every user finding the chatbot helpful. Joyce is actually working on something similar, focusing on speech recognition, powering hospital call centres with AI so more patients can get the care they need.
Although, yes, AI might be capable of speeding up assessments, staff training, note‑taking, or even the waitlist for human therapy, a lot of people have concerns about how all this data is being used and protected. So, who’s regulating the chatbot therapists? And what policies do we need to protect everyone? Or if things really do go wrong while someone is under the care of AI, what happens to them?
Clearly, the role of artificial intelligence in mental health poses lots of questions. We should approach it with caution. But at the same time, there’s this massive potential to expand healthcare access for those who need it most. To make sure that potential is realised in the right way, breakthroughs in technology need to be paired with the right regulation, involving those with lived experience at every stage.
So, one more time… would you use an AI therapist?
Mental health problems affect one in two people and are projected to be the cause of the world’s biggest health burden by 2030. The scale of the challenge means we need a seismic shift in how we address these problems.
Artificial intelligence is already being used to speed up or improve many parts of the complex path that leads from identifying potential treatments to ensuring people are benefitting from them. In particular, generative AI (genAI), which are models of artificial intelligence that learn patterns from existing data and generate new content, is starting to transform many aspects of our lives. Now, people are asking: can genAI play a leading part in transforming mental health treatments?
For many in the mental health community, the answer to this is a hard no. There are concerns that interventions involving genAI are, at worst, harmful and, at best, inferior to current treatments. There are concerns about genAI widening health inequalities as it learns from skewed and biased datasets, general or generic models being inappropriately deployed, leading to false information and unhelpful or even harmful interactions. There are also concerns about privacy, data protection and accountability.
While we need to remain alive to these important concerns, we should keep an open mind to exploring scientifically the role genAI might play as part of a wider revolution in mental health solutions that is taking place.
"Even the best human delivered therapy does not help everyone. Is it right to apply higher standards to AI?"
What are the challenges of human interventions in mental health?
Many non-pharmacological interventions for mental health problems are delivered through language. They are reliant on the capabilities of the human brain, interviews and conversations between people.
As a former clinician, I know the training that therapists go through to deliver talking therapies. I also know that even the best delivered therapy with the best trained human does not help everyone. As a mental health researcher, I am also aware of the issues of cultural competence, bias, misunderstanding and fatigue that affect the effectiveness of interventions delivered by humans. This is irrespective of how caring the humans involved may be.
Is it right to apply higher standards to AI than we apply to humans?
What is the potential of AI in mental health?
As a system grounded in language, genAI could potentially help revolutionise mental health treatments in several ways – from those that are low risk to those that may have higher risk. Below are some examples ranging from the least controversial to the most:
- Automating routine tasks and supporting human interactions such as note-taking, creating reports tailored to different audiences and providing reminders between sessions in order to aid efficiency and reduce waiting times.
- Providing scalable training data for new therapists or AI-generated roleplay to help train those who need to speak to people in a mental health emergency.
- Helping individuals learn new skills that may have therapeutic benefits. For example helping train people in cognitive reframing which is a technique that helps individuals change negative thought patterns.
- Powering fully automated therapeutic chatbots which seek to coach or support individuals with their mental health challenges.
Could AI transform mental health treatments?
We need to consider genAI as part of a wider revolution happening in mental health therapy – from a digital therapy to help reduce the distress that people who hear voices can experience to singing therapy for postnatal depression.
These new treatments are potentially transformative of the mental health landscape in that they focus on specific symptoms that hold people back and are underpinned by cutting-edge science focused on understanding the mechanisms of action. They address issues that are a priority for those with lived experience of mental health problems and, in many cases, are co-designed with lived experience expertise input. They also present new options for scalable solutions. I believe that genAI innovations can exemplify these features as well as helping many of these other new treatments go to even greater scale.
Why we’re choosing curiosity over fear of AI
At Wellcome, we are focused on solutions to the urgent health challenges facing everyone – including mental health. We believe science is crucial to achieve these solutions. Our vision is of a world where no one is held back by mental health problems, and we want to understand the role generative AI can play in making this possible.
That’s why we’re bringing together researchers, healthcare professionals, developers, ethicists and people with lived experience to help explore foundational aspects of genAI’s in terms of potential to aid mental health.
We must not neglect consideration of the dangers in the use of genAI. We are aware of the valid concerns around continuing or even widening health inequities by deploying models trained on biased data, limiting access to those with data resources and the potential harm from the deployment of inappropriate models. Excessive automation of mental health care may need to be guarded against if this is not what best meets the needs of those seeking help.
But for Wellcome the question of whether AI could and will transform mental health outcomes is ultimately an empirical question. It must be answered by investment in science to explore potential including exploring potential risks and harms rather than solely relying on opinions or beliefs. By supporting research into efficacy that also engages with the perspectives of those with lived experience and the complex ethical questions surrounding the use of AI, we can establish what tools are useful, in what context and for whom.
Professor Miranda Wolpert MBE is the Director of Mental Health at Wellcome.
Connect with Miranda:
We're funding teams to carry out fundamental genAI research
We’re funding fundamental research on generative AI and mental health. The goal is to improve measurement or treatment of anxiety, depression and psychosis.
This award spans two phases: applicants first participated in a four-month accelerator delivered by MEXA and could then apply to our funding call.
Successful teams have been awarded up to £3 million each for up to two years and with access to resources and support from experts at Google.
Led by Prof Vikram Patel, Harvard Medical School is developing and testing an AI‑enabled supervision tool to strengthen psychological treatment quality delivered by non‑specialist counsellors in India.
Prof Lauren K White, leading this project at the Children’s Hospital of Philadelphia, is using generative AI to detect how anxiety signals move between parents and children by analysing multimodal behavioural and emotional cues.
New York University, led by Prof Joao Sedoc, is creating controllable, AI‑powered digital patient simulations to train clinicians using realistic psychiatric presentations.
Kids Help Phone, led by Dr Lydia Sequira, is building a generative‑AI conversational simulator to train frontline workers and volunteers to better support youth with anxiety and depression.
Led by Dr Isabelle Scott, University of Melbourne is developing SensAI, a generative‑AI collaborator that supports personalised, collaborative care for young people with anxiety and depression.
Baylor College of Medicine, in collaboration with DeliberateAI, is combining passive sensing with voice‑based generative AI check‑ins to generate more accurate, real‑time measurements of depressive symptoms. This project is led by Prof Eric Alan Storch.
Led by Dr Antonio Fernandez Pardinas, Cardiff University is designing an LLM‑based system, informed by lived experience, to produce proxy measurements of psychotic symptoms using multimodal data.
The University of Toronto, led by Prof Syed Ishtiaque, is creating culturally aligned generative‑AI tools to support para‑counsellors providing mental health care in Bangladesh.
Led by Dr Dominic Oliver, the University of Oxford is developing a voice‑based generative‑AI interview system to identify people at high risk for psychosis and produce tailored clinical reports.
The Slum and Rural Health Initiative is building and validating a culturally grounded generative‑AI platform to enhance depression and anxiety screening for major language groups in Nigeria. This project is led by Dr Isaac Olufadewa.
Led by Prof Lekhansh Shukla, the National Institute of Mental Health and Neuro Science is developing a multilingual AI‑assisted pipeline for accurate depression diagnosis and severity scoring across diverse clinical populations.
The African Population and Health Research Centre is creating a bilingual generative‑AI model for culturally adapted screening of depression, anxiety and psychosis in Kenya and Tanzania. This will be led by Dr Tatenda Duncan Kavu.
Freedom from Torture, led by Dr Jacqui Gratton, is developing a generative AI therapeutic companion for torture survivors with PTSD, providing grounding support to clients, generating clinical reports and adapting strategies based on therapist guidance.
Led by Dr Christine Wasanga, the Shamiri Institute is developing shamiriAI, an AI tool that processes session audio, using multilingual Automatic Speech Recognition (ASR) and prosodic analysis to generate structured feedback for supervisors.
Tags
