AI and Therapy: Where It Helps, Where It Doesn't, and Why I'm Not Worried

A client mentioned recently that they'd been using AI to look something up between our sessions — researching a concept we'd talked about, trying to understand it a bit better before we met again. I want to say plainly: I didn't mind at all. If anything, it told me they were engaging with the work outside the room, which is exactly what I'd want for any client.

It also got me looking properly into AI and therapy — and I quickly noticed that most of the coverage on this is American. Given how different our systems are (both relating to therapy and technology), I wanted to know what the picture actually looks like here in the UK.

It turns out there isn't just one story about AI and therapy. There are at least three, and they matter quite differently.

A tomato plant, staked and supported — not held up entirely, just given enough to grow well. Not a bad way to think about where AI actually fits.

Point one: AI is already built into the NHS itself

This was the part that genuinely surprised me. A UK company, Limbic, has built an AI chatbot that's already deployed inside NHS Talking Therapies. It's important to be precise about what it actually does: Limbic Access doesn't deliver therapy, and it doesn't make treatment decisions. It's a self-referral and assessment tool — it screens people using standard clinical questionnaires, helps them navigate the referral process, and flags anyone who needs urgent attention, so that a human clinician can assess and treat them faster. Think of it as a smarter, more human-feeling front door into NHS mental health services, not a therapist.

A 2024 study published in Nature Medicine tracked over 129,000 patients across 28 NHS sites and found that services using Limbic's self-referral chatbot saw referrals rise by 15%, compared to 6% in matched services without it — with particularly large increases among non-binary and ethnic-minority patients, groups that are often under-referred through conventional routes [1]. A separate tool, Limbic Care, offers AI-based support between sessions alongside a human therapist, rather than replacing one — a 2025 study found patients using it attended 42% more sessions and showed 25% higher recovery rates than historical comparators [2].

Wysa, another UK-based chatbot, has published its own evidence-informed studies in partnership with the NHS, and works in a broadly similar way — supporting people alongside human-led care, rather than instead of it [3]. Both Limbic tools have been through a genuine regulatory process — Limbic Access holds UKCA Class IIa medical device certification and has been independently reviewed by ORCHA, the NHS's own clinical app assessment body [5]. A UK Parliamentary briefing on AI and mental healthcare also points to NICE's own evaluation process for exactly these kinds of technologies, specifically aimed at addressing NHS waiting lists in mental health [4].

This is where AI and therapy can genuinely work well together: regulated, clearly scoped, and sitting inside a service with proper oversight — helping people reach a human clinician faster and stay engaged with treatment, not standing in for one.

Point two: Using general AI alongside therapy, informally

This is where my client's between-session research sits — and where most people are, if they're using AI at all in relation to their mental health. Looking something up, revisiting a concept, doing a bit of your own reading around a topic we've discussed. Used this way, I think it's genuinely comparable to someone reading a book I've recommended.

The charity Mental Health UK offers sensible, practical guidance here: check what AI tells you against reputable sources, and where possible use a dedicated mental health chatbot rather than a general-purpose one, since these tend to be trained specifically on mental health research and data [6]. That's a reasonable, low-effort habit worth building if you're going to use these tools at all.

Point three: Unregulated, general-purpose, and companion apps

This is where the real risk concentrates, and it's worth being honest about it rather than glossing over it. NHS England itself has cautioned against unregulated generalist chatbots specifically [5]. A widely reported study out of Brown University identified around 15 distinct ethical risks in AI-generated mental health responses from general-purpose tools, including mishandling of crises, reinforcing harmful beliefs, and what researchers termed "deceptive empathy" — language that mimics care without the clinical understanding behind it [7].

A survey of psychologists found the large majority believe general-purpose chatbots can inadvertently reinforce negative or delusional thinking, and most said today's tools can't handle complex emotional distress with appropriate nuance [8].

Adolescents are a particular concern here — they're among the heaviest users of general-purpose and companion apps for emotional support, often without any of this context, and without a professional anywhere in the loop [9].

Where I've actually landed

Holding all of this together, here's my honest view: AI can be genuinely helpful, in the right tier, used the right way. What it isn't, and shouldn't become, is a crutch — something that quietly stands in for human contact rather than sitting alongside it.

The tools with real NHS-backed evidence behind them can work precisely because they're designed to get people to human support faster, not to replace it.

The risk isn't the technology itself; it's what happens when something built to be a supplement starts being used as a substitute, especially by people who are struggling and might be reaching for whatever feels immediately available.

If you're using AI in any of these ways alongside our work together, I'd genuinely like to know. It's not something to hide from me, and it's not something I'll take as a slight. If you're ever unsure whether something you're using is helping or quietly becoming a replacement for the harder, more human parts of this work, that's exactly the kind of thing worth bringing into a session and thinking through together.

References

1. Nature Medicine (2024). Study of Limbic's AI self-referral chatbot across 28 NHS Talking Therapies sites, n=129,400.

2. JMIR (2025). Evaluation of Limbic Care, a generative AI therapy companion, on session attendance and recovery rates.

3. Simply Psychology (2026). AI Therapy Chatbots: What the 2026 Research Actually Shows — overview including Wysa's NHS-partnered published studies. https://www.simplypsychology.com/articles/ai-therapy-chatbots-research-review

4. UK Parliament POST (Parliamentary Office of Science and Technology). AI and mental healthcare: opportunities and delivery considerations. https://researchbriefings.files.parliament.uk/documents/POST-PN-0737/POST-PN-0737.pdf

5. AI CERTs News (2025). AI Chatbots Boost Public Health in NHS Mental Health Services — coverage of Limbic's UKCA Class IIa certification, ORCHA review, and NHS England's caution on unregulated generalist chatbots. https://www.aicerts.ai/news/ai-chatbots-boost-public-health-in-nhs-mental-health-services/

6. Mental Health UK. Artificial Intelligence and mental health. https://mentalhealth-uk.org/help-and-information/health-and-wellbeing/artificial-intelligence-and-mental-health/

7. Brown University, reported via Medical Daily (2026). Study identifying 15 distinct ethical risks in AI-generated mental health responses from general-purpose chatbots. https://www.medicaldaily.com/ai-chatbots-mental-health-therapy-safety-risks-psychiatrists-2026-476335

8. American Psychological Association (2026). Survey of psychologists on chatbot use among clients. https://www.apa.org/news/press/releases/2026/06/patients-chatbots-mental-health

9. National Academy of Medicine (2026). AI Chatbots For Mental Health – What Works, What Harms, and What's Next. https://nam.edu/news-and-insights/ai-chatbots-for-mental-health-what-works-what-harms-and-whats-next/