A Question I Get Asked More Than Almost Any Other

Barely a week goes by without a client, a colleague, or someone who has just discovered Dzeny asking me some version of the same question: "Should I try an AI chatbot for my mental health? Does it actually work?"

I understand why people ask. The landscape has changed fast. There are now dozens of apps and platforms offering AI-guided support for anxiety, low mood, sleep problems, and stress. Some are polished and evidence-backed. Others are, frankly, marketing dressed up as therapy. And somewhere in between sits a genuinely curious public trying to figure out what is real.

So let me tell you what I know — not as a pitch, but as a clinical psychologist who has been reading the trial literature carefully and thinking hard about where these tools fit into responsible care.

What the Clinical Trials Have Actually Found

The honest answer is: it depends enormously on the problem, the person, and how the tool is used.

The strongest evidence we have in 2026 still clusters around mild to moderate anxiety and depression. A meta-analysis published in *npj Digital Medicine* in 2023 pooled 15 randomised controlled trials and found that AI-based conversational agents significantly reduced depressive symptoms (Hedge's g 0.64). A larger and more recent synthesis in the same journal, covering 48 randomised trials and just over 28,000 participants, found smaller but consistent effects — standardised mean differences of −0.27 for depression, −0.20 for anxiety and −0.26 for stress. That is meaningful. It is also limited. Waitlist controls are a low bar. The question is not just "better than nothing" but "better than, or complementary to, something else."

The Woebot trials, which remain the most cited work in this space, showed reductions in PHQ-9 depression scores over two weeks in college students — a result that held up in a later study with a more diverse adult sample. What matters as much as the effect sizes is that people kept using these tools — retention is where most digital mental health products fail, and it is a real-world data point, not a laboratory one.

What the research does not show is effectiveness for complex presentations. There are no credible trials demonstrating that AI chatbots meaningfully help people with post-traumatic stress disorder, psychosis, active suicidality, severe eating disorders, or personality disorders involving emotional dysregulation. This is not a gap that better prompts will fill. These conditions require the kind of relational attunement, clinical risk management, and individualised formulation that a language model cannot replicate.

Where Chatbots Actually Help

I try to be specific with my clients, so let me be specific here.

AI tools seem most useful when they function as structured habit support between sessions — prompting you to notice mood patterns, log your sleep, practise a breathing exercise, or revisit a CBT thought record you learned with your therapist. In that role, they are essentially a sophisticated, non-judgmental notebook with a reminder function. That is genuinely useful for a lot of people.

They are also genuinely helpful for reducing the barrier of first contact. I have spoken with multiple people who told me they practised naming their feelings in an AI chat before they felt ready to say the same words to a human clinician. That is not trivial. Shame and stigma are real barriers to help-seeking, and if a chatbot lowers the threshold, that matters clinically.

Other areas where evidence is accumulating:

  • Psychoeducation — explaining what generalised anxiety is, what CBT involves, what panic attacks feel like physiologically
  • Structured skills delivery — guided muscle relaxation, behavioural activation scheduling, worry postponement exercises
  • Between-session support for people already engaged in human-led therapy
  • Crisis signposting — quality apps will always direct users to emergency services when risk indicators appear

Where Chatbots Fall Short — And Why That Is Not a Failure

A language model does not have a nervous system. It cannot notice that your voice tightened when you mentioned your mother, or that you laughed in a way that did not match your words. It cannot hold the full context of your life across months and years with the kind of meaning-making that a skilled therapist brings to that history. It cannot be genuinely present with you in distress.

These are not bugs to be patched. They are structural features of what AI is. A hammer is not a failed screwdriver — it is simply a different tool. The problem comes when marketing suggests otherwise.

I am also concerned about a subtler risk: substitution without equivalence. If cost or access leads someone with moderate depression to use an AI app instead of ever connecting with a human clinician, and the app keeps them just functional enough to not seek more help, that is not a good outcome. Managed adequately, AI support can be a bridge. Used as a permanent replacement, it may prevent people from getting the depth of care they need.

How to Use an AI Tool Without Expecting the Wrong Things From It

If you are considering an AI mental health tool, here is how I would encourage you to approach it practically.

Before you start:

  • Be honest with yourself about symptom severity. If you are having thoughts of self-harm, please contact a crisis line or a clinician first — not an app.
  • Choose a platform that is transparent about its evidence base and has a clear safeguarding protocol. Look for clinical advisory involvement, not just a "built with mental health experts" tagline.

While you use it:

  • Treat it as a tool, not a relationship. Notice what it helps you do — and what it cannot.
  • Use it to build skills you can take elsewhere. CBT thought records, breathing techniques, mood tracking — these are yours, not the app's.
  • If you are in therapy, tell your therapist you are using it. The best outcomes come when digital tools are integrated into care, not hidden from it.

Know when to step up:

  • If your symptoms are worsening over two to three weeks, that is a signal to seek human support.
  • If you find yourself relying on the chatbot in moments of acute distress as a primary coping strategy, that is worth examining with a professional.

What I Think This Means for You

I built Dzeny because I believe that access to good psychological support should not be determined by postcode or income. Digital tools are part of how we close that gap. But closing the gap honestly means being clear about what we are offering.

AI-assisted tools, used thoughtfully, can extend reach, build skills, reduce shame, and support people between sessions. They are not interchangeable with therapy. They are not going to process your childhood, repair attachment wounds, or hold risk. What they can do, done well, is meaningful — and right now, meaningful matters.

If you are not sure where to start, Dzeny offers an initial consultation with a human clinician who can help you figure out what kind of support actually fits your needs. That conversation costs nothing, and it will give you a clearer map than any algorithm can.