A
Mental Health System Under Pressure

Mental health care is facing a demand crisis that existing
services cannot resolve through conventional means. Waiting lists for
counselling and psychiatric support have stretched to months or years in many
countries. Rates of anxiety, depression, and trauma-related conditions have
risen sharply in the years since the pandemic, and the workforce of trained
therapists has not kept pace. Into this gap, a new category of tool has arrived:
AI-powered applications designed to deliver structured psychological support,
in particular cognitive behavioural therapy, at scale and at low
cost.

The appeal is real and immediate. For someone waiting three months
to see a therapist, an AI app that can walk through the core techniques of
CBT at any hour of the day or night addresses a genuine need. For a
healthcare system struggling to allocate finite clinical resources,
automating the more routine elements of structured therapy offers a practical
route to extending reach without proportionally expanding the workforce. The
question is whether the trade-off is worth making, and for whom the costs are
highest.

What the Evidence Says About Effectiveness

A meta-analysis published in Nature Digital Medicine examined AI
conversational agents deployed to address symptoms of depression and
psychological distress. The review found evidence of potential benefit,
particularly when these tools were deployed in mobile, multimodal formats
that allowed users to engage frequently and consistently (Nature Digital
Medicine). For individuals experiencing mild to moderate distress
who have limited access to professional care, AI tools can provide structured
support that is meaningfully better than no support at all.

A systematic review on ResearchGate examined studies of
AI-delivered CBT interventions specifically for anxiety and depressive
symptoms. The findings were more qualified: modest effectiveness for anxiety
reduction, particularly among younger users, but less consistent results for
older adults and for depression (ResearchGate). The
conclusion was that AI therapy can help, but it cannot yet match the
consistency or depth of outcomes associated with human-led therapy,
especially for individuals with complex or longstanding
conditions.

The most revealing study came from an arXiv paper in which
licensed psychologists compared AI-driven CBT sessions against sessions led
by peer counsellors with similar training levels. The AI adhered to CBT
protocols with considerable precision, often outperforming novice therapists
in structural fidelity. But it failed repeatedly in areas that required
cultural sensitivity, emotional attunement, and the kind of collaborative
formulation that develops over multiple sessions with a human who remembers
the previous conversation in context. The conclusion was unambiguous:
technically precise, emotionally hollow.

The Empathy Problem

Empathy is not incidental to therapy. It is central. Decades of
research on what makes psychotherapy effective consistently identifies the
therapeutic alliance as one of the strongest predictors of outcome, in some
studies more predictive than the specific technique being used. The
therapeutic alliance is built on trust, rapport, and the experience of being
genuinely understood by another mind. AI systems, however fluent and
responsive, do not have minds in this sense. They simulate understanding,
sometimes convincingly, but they do not experience it.

Clinicians at Wildflower Mental Health have articulated this
directly, noting that AI chatbots do not do empathy in any meaningful sense
(Wildflower Mental
Health). They can reproduce phrases that resemble empathetic
responses. They can identify emotional language in user input and mirror it
back in appropriately supportive ways. But they cannot feel the discomfort of
sitting with someone in genuine distress, cannot notice the slight hesitation
that precedes a disclosure the user is not sure they want to make, and cannot
offer the wordless reassurance of a shared silence. For some users,
particularly those with complex trauma or relational difficulties, the
absence of these qualities is not merely an inconvenience. It is a barrier to
the therapeutic work itself.

Ethical Risks of Displacement

The Stanford Institute for Human-Centered AI has warned that
positioning AI as a full replacement for human therapists introduces risks
that go beyond questions of effectiveness (Stanford HAI). These
include the risk that individuals experiencing genuine crises are
inadequately screened and directed toward an AI tool that cannot safely
manage their level of need, the risk that hallucinations or algorithmically
generated misinterpretations of a user’s account lead to inappropriate or
harmful guidance, and the structural risk that investment in AI therapy tools
is used to justify reducing investment in human clinical
services.

That last risk is the most insidious. If healthcare commissioners
use AI adoption as grounds for not training more therapists or reducing
funding for specialist services, the net effect for the people most severely
affected by mental illness may be negative, even if the aggregate statistics
on access appear to improve. Concerns
about AI’s effects on mental health more broadly are already
receiving serious attention from clinicians and researchers, which makes the
question of how AI is deployed in therapeutic contexts particularly
sensitive.

Augmentation as the Productive Model

The evidence points toward a model in which AI augments rather
than replaces human therapeutic work. This is not a consolation position. It
is a genuinely useful and scalable approach if implemented carefully. An AI
application can deliver structured CBT exercises between sessions with a
human therapist, reinforcing techniques that have been introduced in person.
It can provide a low-stakes space for someone to practise the cognitive
skills of the therapy before bringing their reflections to the clinical
relationship. It can flag patterns in a user’s responses that a therapist
might want to explore, offering a richer picture than a once-weekly session
alone can provide.

This kind of integration requires human oversight at every step.
The AI is not making decisions about the user’s care; it is supporting the
human professional who is. That distinction preserves accountability, ensures
that complex needs are identified and escalated, and keeps the therapeutic
relationship, with all its human imperfection and genuine connection, at the
centre of the intervention. It also requires investment in the infrastructure
that connects AI tools to clinical services, investment that healthcare
systems will need to make deliberately rather than assuming that technology
alone will resolve the resource gap.

Access, Authenticity, and the Stakes of Getting It
Wrong

There are parts of the world where AI therapy, for all its
limitations, represents a genuine improvement over the alternative of no
support at all. In countries with severe shortages of mental health
professionals, a structured AI tool that can reach rural and underserved
populations may prevent some deterioration and provide a bridge to eventual
human care. In those contexts, the accessibility argument for AI therapy has
real moral weight, and dismissing it entirely would be its own kind of
harm.

But accessibility and adequacy are different standards, and
conflating them risks normalising a second tier of mental health support for
people who already face disproportionate disadvantage. The
regulatory conversation around AI in healthcare needs to engage
directly with the question of whether AI therapy tools are being deployed as
a complement to adequate care or as a substitute for it. That distinction
will determine whether AI in mental health represents a genuine expansion of
human wellbeing or a more efficient management of its
absence.

About the Author

By Stuart Kerr, Technology Correspondent, LiveAIWire. Stuart
covers artificial intelligence, health, and the ethical frameworks shaping
how technology interacts with human vulnerability. About
LiveAIWire.