For all the enthusiasm surrounding AI in mental healthcare, the conversation tends to collapse into a single question: will technology replace therapists? It is a debate that generates more heat than clarity, in part because it treats therapy as a service that can be replicated rather than a relationship that produces something specific. Understanding what that something is, may be the most important starting point.
Jia Sitlani, a psychologist and counsellor based in Mumbai who runs a private practice and teaches psychotherapy at the university level, frames it in terms of witnessing. Therapy, she argues, is not simply a transfer of insight or technique. It is an encounter between two people who are each present to the other, each capable of being seen as well as seeing. “When an AI witnesses you, you are not really witnessing the AI back,”[1] she says. “Because there is nothing to witness. It is a program.” [2] The reciprocity is not incidental. It is, in her view, the mechanism through which therapeutic change becomes possible.
This distinction matters practically, not just philosophically. Many AI systems optimised for mental health applications tend toward relentless affirmation, designed to keep users engaged rather than to challenge them. But productive therapy is not always comfortable. It sometimes requires a clinician to name what a client is avoiding, to offer a reality check rather than reassurance, to hold a position even when it is unwelcome. An AI shaped by satisfaction metrics is structurally unlikely to do that well.
None of this forecloses a role for technology. It redefines it. The more useful question is not whether AI can do what therapists do, but where it can extend, prepare and support the work that therapists do, and the answers are more specific than the general debate tends to acknowledge.
One gap that clinical practitioners regularly encounter is the distance between insight and behaviour. Clients frequently understand, at an intellectual level, exactly what they need to do differently: communicate more directly, initiate social contact, manage a difficult conversation without becoming defensive, and still find themselves unable to act on that understanding when the moment arrives. Neural pathways, as Ms. Sitlani explains to her clients, are not created by knowledge alone. They require repetition.[3] What bridges the gap between knowing and doing is rehearsal, and what makes rehearsal possible is having something concrete to practise with. Scripts, small and specific and contextually adapted, function as cognitive handholds, reducing the uncertainty that causes people to freeze.[4] A tool capable of generating these for individual clients, available between sessions and responsive to their specific social contexts, would address something that most existing wellness applications do not.
A second and structurally different problem is dropout. The rate at which people attend one session of therapy and do not return is higher than is commonly appreciated[5], and a significant part of it comes down to unfamiliarity. People arrive without a clear sense of what therapy involves, encounter something unexpected, and disengage. The issue is compounded by the fact that different therapeutic modalities look and feel quite different from one another. A psychoanalytic session and a person-centred session are not the same experience, and this is rarely explained in advance. Avatar-based psychoeducation, in which prospective clients can explore what different therapeutic styles involve before committing to a session, addresses this problem at the point where it actually occurs: before the therapeutic relationship has had a chance to form. [6] The goal is not to simulate therapy but to reduce the mismatches that cause people to abandon it.
There is a harder scenario where the calculus shifts somewhat. When someone is in acute crisis, has no access to human support, and will not reach out to a professional, the question of whether AI contact is therapeutically ideal becomes secondary to whether it keeps them safe. “All a person needs in that moment is one morsel of reassurance,” Ms. Sitlani says, “even if it’s from a robot.” [7] The case for AI here is narrow and specific: stabilisation in the present, not ongoing care, with consistent redirection toward professional support. Used within those limits, it is not a substitute for clinical intervention. It is a bridge to it.
The question of who builds these tools, and for whom, carries its own complications. Most of the frameworks underpinning clinical practice were developed in Western, educated, industrialised contexts and carry assumptions that do not always travel. Therapeutic modalities that emphasise autonomy, rational problem-solving and separation from harmful relationships presuppose a social context in which those are realistic options. In collectivist settings, where family embeddedness is not simply a preference but a social reality, and where ostracism can carry genuine psychological cost, the same approaches can produce friction rather than progress. The therapeutic goal, in many cases, is not to leave but to learn to coexist [8]. Approaches like narrative therapy and acceptance and commitment therapy tend to be more adaptable in this context, not because they were designed for it but because they make fewer structural assumptions about individualism. [9] Any technology intended to support mental healthcare in these settings needs to treat cultural context not as a localisation issue to be addressed after the fact, but as a design constraint from the beginning.
A final concern cuts closer to the foundations of the field itself. Neuroscience is increasingly present in clinical training, but it tends to be taught as taxonomy: which region governs which function, which neurotransmitter affects which system. [10] What is taught less consistently is application: what any of this means for the person in the room. Anxiety and fear are protective responses, shaped by evolution, originating in parts of the brain that predate rational thought. [11] Knowing this changes how a clinician can respond to a client who is overwhelmed by their own reactions. But if training conveys the biology without conveying the meaning, the knowledge remains inert. The gap between what neuroscience knows and what practitioners can usefully do with it is not primarily a research problem. It is a pedagogical one.
Technology’s role in mental health care is most legible when the underlying question is right. Not whether AI can replicate what happens between two people in a room, for reasons that go beyond current capability the evidence suggests it cannot, but whether it can extend the conditions under which that kind of encounter becomes possible. Scripts that build behavioural muscle memory, simulations that lower the barrier to entry, stabilisation tools for moments of crisis: these are not replacements for human care. They are the infrastructure that allows more of it to happen.
You can watch the full interview at the following link: AI and the Future of Mental Healthcare | Dr. Jia Sitlani | TRANSiT Lab
-By Bhavya Dakavaram