AI, Anxiety and the Future of Calm

Originally published October 2025 · Updated May 2026 with the March 2025 Dartmouth Therabot RCT, the May 2025 JMIR NHS CBT comparison study, and the November 2025 CBT chatbot narrative review.

Mental health demand has outpaced clinical supply for years. In the UK, 55% of adults with mental illness access no treatment. NHS waiting times for talking therapy can stretch to months. Private therapy costs make consistent support inaccessible for many people.

AI-powered mental health tools have emerged as the most discussed potential response to this gap. The evidence for their effectiveness — limited and inconsistent as recently as 2023 — has advanced considerably in the past eighteen months. The picture is more nuanced than either the enthusiasts or the sceptics suggest. Here is what the 2025 research actually shows.

TL;DR

• A March 2025 Dartmouth RCT of the Therabot AI chatbot found clinically significant reductions in major depressive disorder, generalised anxiety disorder, and eating disorder risk — with effect sizes approaching those of first-line psychotherapy.
• A May 2025 JMIR study found an AI-combined digital programme achieved outcomes comparable to face-to-face NHS CBT for generalised anxiety, while significantly reducing clinician time per participant.
• A November 2025 narrative review of 14 studies found CBT-based chatbots consistently demonstrated short-term reductions in depressive symptoms, with mixed but often positive results for anxiety.
• A November 2025 meta-analysis found significant improvements in depression, anxiety, stress, and psychosomatic symptoms in adolescents and young adults using AI chatbots.
• The evidence supports AI tools as effective for mild to moderate anxiety and depression — particularly as a first point of contact, a complement to human therapy, and a resource during waiting periods.
• AI tools cannot replicate the therapeutic alliance — the strongest single predictor of therapy outcomes for complex conditions.

The 2025 Therabot Trial: The Most Significant Finding

The most consequential piece of AI mental health research published in 2025 is a randomised controlled trial from Dartmouth College, published in NEJM AI in March 2025.

The study tested Therabot — a generative AI therapy chatbot developed at Dr Nicholas Jacobson''s AI and Mental Health Lab at Dartmouth — against waitlist controls in patients diagnosed with major depressive disorder, generalised anxiety disorder, and clinically high risk for eating disorders.

The results were striking. Therabot produced clinically significant reductions across all three conditions. The effect sizes for depression and anxiety exceeded those commonly reported for antidepressant medications in clinical trials, and approached or matched the effect sizes observed for first-line psychotherapy. These are not marginal improvements — they represent a meaningful therapeutic signal from a fully digital, AI-delivered intervention.

For context: mental health smartphone applications typically show effect sizes of around 0.28 for depression and 0.26 for anxiety. Antidepressant medications show approximately 0.31. First-line psychotherapy shows approximately 0.85. The Therabot trial produced effect sizes within a range that, if replicated at scale, would represent a genuinely significant advance in accessible mental health care.

The study is not without limitations. It was a single trial from the lab that developed the tool, with a waitlist rather than active treatment comparator. Replication by independent research groups is essential. But its publication in NEJM AI signals that AI mental health research has reached a level of methodological credibility that warrants serious clinical attention.

The NHS Comparison: Does AI Match Face-to-Face CBT?

The second landmark 2025 study addresses the most practically important question for the UK context: can an AI-combined digital programme match the outcomes of NHS-delivered face-to-face CBT?

A May 2025 study published in JMIR evaluated a digital programme combining AI and human support for generalised anxiety against two external comparator groups — NHS Talking Therapies typed CBT and face-to-face CBT from an NHS Foundation Trust. The programme used an AI-driven conversational agent to deliver clinician-written content with clinician oversight and user support.

The digital programme achieved clinical outcomes comparable to both human-delivered comparators, as measured by reduction in GAD-7 anxiety scores. Crucially, it also significantly reduced clinician time per participant — meaning comparable outcomes were achieved with considerably less demand on the clinical workforce.

This finding has direct implications for the NHS treatment gap. If an AI-combined programme can produce comparable anxiety outcomes to face-to-face CBT while reducing clinician time, it represents a genuine scalability solution — not a compromise but a pathway to reaching the 55% of people with mental illness who currently access no treatment.

The Broader Evidence Base

Beyond these two landmark studies, the 2025 evidence base for AI mental health tools has accumulated considerably.

A November 2025 narrative review of 14 peer-reviewed studies on CBT-based chatbots found they consistently demonstrated short-term reductions in depressive symptoms, with anxiety outcomes more variable — some studies showing significant improvements, others showing non-significant effects. The review noted that fully automated chatbots produced smaller effects than those with some human oversight — a finding consistent across the literature.

A November 2025 meta-analysis specifically examining AI chatbots for adolescents and young adults found significant improvements in depression, anxiety, stress, and psychosomatic symptoms — with standardised mean differences of 0.43 for depression, 0.37 for anxiety, and 0.41 for stress. These are moderate effect sizes, consistent with the broader digital mental health literature.

A May 2025 meta-analysis of AI-driven conversational agents for young people found significant effectiveness for reducing anxiety and depression symptoms, with engagement and personalisation emerging as key moderators — tools that adapted to the user and maintained engagement produced stronger results than static or rigid programmes.

What Makes AI Mental Health Tools Work — When They Work

The research on what determines whether an AI mental health tool produces meaningful outcomes points consistently to the same factors.

Evidence-based content. Only 48% of mental health apps deliver content based on CBT principles — the gold-standard evidence-based approach for anxiety and depression. Tools built on structured, evidence-based therapeutic frameworks produce meaningfully stronger results than those built on general wellness content or conversational AI without clinical grounding.

Engagement and personalisation. Tools that adapt to the user, maintain conversational engagement, and provide a sense of being understood — rather than simply presenting information passively — produce better outcomes. The therapeutic relationship that makes human therapy so effective has partial analogues in digital tools: consistency, responsiveness, and the sense of being accompanied rather than informed.

Human oversight. The consistent finding across the 2025 literature is that AI tools with some level of human support or oversight produce larger effect sizes than fully automated tools. The combination of AI scalability and human clinical judgment appears to optimise both reach and effectiveness.

Appropriate indication matching. AI tools work best for mild to moderate anxiety and depression in people who are motivated to engage. Evidence for severe mental illness, complex trauma, and personality disorders is considerably weaker. Appropriate triage — matching people to the right level of care — is as important as the tool itself.

What AI Tools Cannot Replace

The honest assessment of AI mental health tools requires being clear about their limits — not to dismiss them, but because overselling produces harm when people who need human therapeutic care delay or avoid it.

The therapeutic alliance — the quality of the relationship between therapist and client — is the strongest single predictor of therapy outcomes across all modalities. It predicts outcomes better than the specific therapy type. It predicts outcomes better than treatment adherence. Human attunement — the neurobiological experience of being genuinely understood by another person — produces specific physiological effects including parasympathetic activation, oxytocin release, and co-regulation of arousal that AI systems cannot replicate.

For complex conditions — severe depression, PTSD, trauma, personality disorders, psychosis — human therapeutic relationship is not just preferable. It is part of the mechanism of change. AI tools used as a substitute for human therapy in these contexts can delay appropriate care.

The appropriate framing for AI mental health tools is not "as good as therapy" but "accessible, effective, and complementary to therapy for mild to moderate conditions in people who cannot otherwise access care." That is a genuinely valuable and large indication — not a compromise position.

The Anxiety Epidemic Context

Understanding why AI mental health tools matter requires understanding the scale of the problem they are addressing.

Anxiety disorders are the most common mental health condition globally, affecting an estimated 284 million people. In the UK, anxiety and depression account for the majority of the 8 million people on NHS mental health waiting lists. The gap between need and provision is not a resource problem that can be solved by training more therapists — the training pipeline is too slow and the prevalence too high.

Digital and AI tools are the only intervention that scales to match the scale of the problem. The question is not whether they should be part of the solution — they must be — but which tools, for which people, at what level of care, with what human oversight.

The 2025 evidence provides a clearer answer to that question than was available in previous years. AI tools with CBT-grounded content, human oversight, and appropriate engagement design are effective for mild to moderate anxiety and depression. They work best as a first point of contact, a resource during waiting periods, and a complement to human therapy rather than a replacement for it.

The Reset Companion: Where Vitae Fits

The Reset Companion is designed within this evidence context — as a complement to professional support, not a substitute for it.

It is not a clinical tool and does not diagnose, treat, or replace therapy. It is an AI-powered support tool designed to help people build the daily habits, reflection practices, and nervous system regulation skills that the evidence consistently identifies as beneficial alongside — and between — professional mental health support.

The Reset Companion works best alongside the Reset Series™ guides — providing the daily touchpoint and reflection structure that supports the habit formation and self-awareness that both therapy and lifestyle change require. For people on waiting lists, it provides accessible support in the gap. For people in therapy, it provides between-session continuity. For people who are not in crisis but want to actively support their mental wellbeing, it provides a structured, evidence-informed framework.

Frequently Asked Questions

Do AI therapy chatbots actually work?

The 2025 evidence suggests yes — with important qualifications. A March 2025 Dartmouth RCT found an AI chatbot produced effect sizes for anxiety and depression approaching first-line psychotherapy. A May 2025 JMIR study found AI-combined digital therapy comparable to NHS face-to-face CBT for generalised anxiety. A November 2025 meta-analysis found significant improvements in depression, anxiety, and stress from AI chatbots in young people. The evidence is strongest for mild to moderate anxiety and depression in people motivated to engage with structured, CBT-grounded tools.

Can AI replace therapy?

No — and this distinction matters clinically. The therapeutic alliance — the quality of the human relationship — is the strongest predictor of therapy outcomes and cannot be replicated by AI systems. Human attunement produces specific neurobiological effects that AI cannot reproduce. For mild to moderate anxiety and depression, AI tools can produce meaningful outcomes comparable to some forms of human-delivered care. For complex conditions, severe presentations, trauma, and personality disorders, human therapeutic care is not substitutable.

What makes an AI mental health tool effective?

The consistent findings across the 2025 literature point to: evidence-based content grounded in CBT principles; adaptive personalisation that responds to the user; some level of human oversight rather than full automation; and appropriate matching to the right indication — mild to moderate anxiety and depression rather than complex or severe conditions.

Is the Vitae Reset Companion a therapy tool?

No. The Reset Companion is an AI-powered support tool designed to complement professional mental health support — not replace it. It is not a clinical tool and does not diagnose or treat any condition. It works best alongside the Reset Series™ guides as a daily reflection and habit-support tool, particularly useful during waiting periods, between therapy sessions, or for people actively managing their mental wellbeing as a lifestyle practice.

How do I know if I need an AI tool or a real therapist?

AI tools are appropriate for mild to moderate anxiety and depression, general wellbeing support, stress management, and building daily mental health habits. A human therapist is appropriate for significant anxiety or depression that is impairing daily functioning, trauma, complex mental health conditions, suicidal ideation, or any situation where symptoms have persisted for more than a few weeks despite self-directed support. If in doubt, a GP assessment is the appropriate first step.

What is the NHS doing about AI mental health tools?

The NHS is increasingly incorporating digital mental health tools into its care pathways. NHS Talking Therapies — the primary access route for CBT in England — includes digital programme options alongside face-to-face delivery. The May 2025 JMIR study demonstrating comparable outcomes to face-to-face CBT from an AI-combined digital programme was conducted using NHS comparators and has direct implications for NHS commissioning of digital mental health tools.

The Bottom Line

The evidence for AI mental health tools has crossed a meaningful threshold in 2025. The Dartmouth Therabot RCT, the NHS CBT comparison study, and the accumulated meta-analyses consistently show that well-designed, CBT-grounded, appropriately supported AI tools can produce meaningful reductions in anxiety and depression at a scale that human-delivered care cannot match alone.

The honest framing: AI tools are effective complements to human care, not replacements for it. They work best for mild to moderate presentations, with human oversight, using evidence-based content. They cannot replicate the therapeutic relationship that makes human therapy most powerful for complex conditions. What they can do — and what the 2025 evidence confirms they are doing — is reach the millions of people who currently access no support at all. That is not a small thing.

For structured daily support for anxiety and stress, the Anxiety Reset and Stress Reset from the Reset Series™ provide the evidence-based frameworks the Reset Companion works best alongside.

Related reading

• Why Therapy Works: The Neuroscience of Talking and How Digital Support Is Changing Everything
• Micro-Anxiety: The Tiny Stressors That Are Quietly Draining Your Mental Health
• Cortisol Explained — and How to Reduce It Without Making Things Worse

Popular Articles