By
Stuart Kerr, Technology Correspondent,
LiveAIWire
A randomised controlled trial
published in the New England Journal of Medicine AI in March 2025 found that
Therabot, a generative AI therapy chatbot developed at Dartmouth College,
produced clinically significant reductions in symptoms among patients
diagnosed with Major Depressive Disorder, Generalised Anxiety Disorder, and
those at clinical high risk for eating disorders. The Therabot
study is the most rigorous clinical trial of a generative AI mental
health tool published to date, and its findings shifted the conversation from
whether AI therapy can work to how, when and for whom it works best. The
answers are more nuanced than either enthusiasts or sceptics have tended to
acknowledge.
Mental health services globally face a
structural access problem that no realistic expansion of human therapist capacity
will resolve in the near term. In England, the NHS waiting list for talking
therapies routinely exceeds twelve weeks. In the United States, 150 million
people live in federally designated mental health professional shortage
areas. In lower-income countries, the treatment gap is still larger. AI
therapy tools are not emerging because clinicians think algorithms make
better therapists. They are emerging because the alternative, for much of the
world, is no treatment at all. The relevant question is not whether an AI
therapist matches a human therapist in head-to-head comparison. It is whether
an AI tool is better than a twelve-week wait or no access at
all.
For anyone considering whether to use an AI therapy
tool, or wondering whether to recommend one to someone they care about,
understanding what the clinical evidence actually shows, and where its limits
sit, is the right starting point.
What the Randomised
Trial Evidence Shows
The NEJM AI Therabot study is the
landmark data point, but it sits within a growing body of evidence. A meta-analysis
published in the Journal of Medical Internet Research in November
2025 synthesised 31 randomised controlled trials covering 29,637
participants, examining AI chatbot effectiveness for mental health outcomes
in adolescents and young adults aged 15 to 39. The analysis found that
chatbots were more effective for psychosomatic symptoms in clinical
populations, those with more severe baseline symptoms, than in non-clinical
groups, a finding that holds across the broader evidence base. Standalone
chatbot apps were more effective for anxiety than web-integrated chatbots,
suggesting that deployment format affects outcomes. The certainty of evidence
across most outcomes was rated as very low to low, reflecting the early stage
of the research base rather than evidence of harm. The authors called for
more rigorous study, which is the appropriate scientific response to
promising but preliminary results.
The effect sizes
reported in the Therabot trial were meaningful, not marginal. Participants
experienced clinically significant symptom reductions in depression and
anxiety measures, comparable in some metrics to outcomes reported for brief
human-delivered cognitive behavioural therapy. The trial compared Therabot
against a waitlist control, not against active human therapy, which is the
most relevant comparison for the access gap problem but not the one that
answers whether AI therapy is as good as human therapy. That comparison has
not yet been adequately tested at scale.
What AI Therapy
Cannot Do
The limits of AI therapy tools are as important
to understand as their capabilities, particularly for users who might turn to
them in situations that exceed their design parameters. A study published in
Psychiatric Services in August 2025 assessed popular chatbots’s ability to
identify when a user was at risk of suicide, using 30 queries of varying risk
levels. The results highlighted serious inconsistencies in crisis recognition
across the most widely used platforms, which is a clinical safety concern of
the highest order. AI therapy tools are not equipped to provide the level of
risk assessment and crisis intervention that trained clinicians can offer.
For anyone experiencing suicidal thoughts or acute mental health crises, a
human clinician or crisis service is the appropriate response, and no AI tool
should substitute for that pathway.
Beyond crisis
situations, the therapeutic relationship remains an open question. Decades of
psychotherapy research establish that the quality of the relationship between
therapist and patient, the alliance, is one of the strongest predictors of
treatment outcomes across modalities. AI tools can simulate relational warmth
and provide consistent, non-judgmental responses. Whether they can build the
kind of genuine therapeutic alliance that characterises effective human
therapy is a research question that the current evidence base cannot yet
answer. The absence of evidence on this point is not evidence of absence, but
it is a reason for caution about claims that AI therapy is equivalent to
human therapy for all purposes.
The Privacy and Data
Questions You Should Ask
AI therapy apps collect unusually
sensitive data. Conversations about depression, anxiety, relationship
difficulties, trauma, and medication are among the most personal information
a person can share. The privacy protections governing that data vary
significantly across apps and jurisdictions, and the data is commercially
valuable in ways that create incentives misaligned with therapeutic
confidentiality. Before using any AI mental health tool, it is worth reading
the privacy policy with attention to how conversation data is stored, whether
it is used to train models, who has access to it, and whether it can be
shared with third parties or disclosed to authorities in legal contexts.
These questions have different answers depending on the app and the
jurisdiction, and the answers matter.
For the broader
picture of how AI health applications are being regulated and what data they
collect, the situation in AI
health monitoring on smartphones shares many of the same structural
tensions. And the evidence base for AI-assisted
approaches to trauma therapy adds additional context for specific
therapeutic applications where the stakes are highest.
The
Honest Verdict
AI therapy tools have demonstrated clinical
efficacy for depression and anxiety symptoms in controlled trials. They
provide access to structured mental health support for populations who would
otherwise wait months or receive nothing. They are most effective for people
with moderate clinical symptoms rather than either mild wellness concerns or
severe acute crises. They have significant unresolved questions around
therapeutic alliance, crisis safety, long-term outcomes, and data privacy.
They are not a replacement for human therapists and should not be positioned
as one.
For anyone navigating a mental health challenge
right now, the honest advice is: an AI therapy tool is likely to be helpful
if you have mild to moderate depression or anxiety, no acute crisis
indicators, and no reliable access to human-delivered therapy in the near
term. If you are in crisis or experiencing suicidal thoughts, please contact
a crisis service directly. If human therapy is accessible to you, it remains
the evidence-based first choice for most presentations. The AI tools are
filling a gap, not closing one. That is a meaningful contribution, and it
deserves honest acknowledgment alongside honest recognition of what they
cannot yet do. For a broader view of how AI
is being positioned as a therapeutic companion across different
delivery models, the landscape is more varied than the headline claims on
either side suggest.
The regulatory environment for AI
therapy tools is also evolving. The FDA’s January 2026 guidance on wellness
devices clarified the oversight boundary between wellness applications and
regulated medical devices, but the therapeutic chatbot category sits in a
grey zone where the clinical evidence is stronger than the regulatory
clarity. A tool producing clinical-grade symptom reductions, as Therabot
demonstrated, arguably warrants a clearer regulatory pathway than the current
framework provides. Several developers are voluntarily pursuing clinical
validation rather than relying on the wellness category, which is the right
direction for building the long-term credibility that genuine therapeutic
tools require.
About the Author
Stuart
Kerr is Technology Correspondent at LiveAIWire, covering artificial
intelligence, cybersecurity, and the social impact of emerging technology. He
publishes daily at LiveAIWire.com.