The emergence of AI systems positioned as therapeutic agents — not merely tools for emotional processing but active participants in psychological intervention — marks one of the most consequential and contested developments in contemporary mental health care. The promise is legible and significant: in the United States alone, over sixty percent of people who need mental health services do not receive them. The barriers are structural — cost, availability, stigma, geography — and they are not being resolved by the existing system at any rate that would match the scale of demand. AI therapists, in this framing, are not a luxury but a response to a genuine public health emergency, and dismissing them on the grounds that they are not equivalent to human therapy is a counsel of perfection that the clinical system cannot provide at the required scale.
The harm is also legible and significant. Psychotherapy is not a conversation about feelings. At its most developed, it is a highly specialized form of interpersonal engagement in which the therapist's own subjectivity — their capacity for genuine emotional response, their trained ability to monitor and use their own reactions as diagnostic data, their commitment to the client's wellbeing that can withstand the client's resistance — is the primary instrument of change. This instrument cannot be simulated by a language model, however sophisticated. The transference and countertransference dynamics that psychodynamic therapists identify as central to therapeutic change require two actual minds, each of which can be surprised, moved, and changed by the other. The alliance research in psychotherapy — the body of work establishing that the quality of the therapeutic relationship is the strongest predictor of outcome — is not evidence that any warm relationship will do; it is evidence that a particular kind of human relationship, characterized by empathy, genuineness, and unconditional positive regard enacted within a real asymmetric care dyad, is what produces change.
At collective scale, the promise and harm of AI therapists are not merely additive but structurally interactive. The same conditions that make the promise compelling — the undersupply of human mental health care — also make the harm more serious. A person with access to effective human therapy who also uses an AI therapeutic tool has a very different risk profile than a person for whom the AI is the only mental health resource available, and for whom the AI's limitations — its inability to detect psychosis, to recognize genuine suicidal intent from verbal cues alone, to escalate appropriately, to hold a person through a crisis — are not supplemented by human backup. The distribution of AI therapy across the population is not random; it will be concentrated precisely where human alternatives are least available, which means the harm from its limitations will be concentrated in the most vulnerable populations.
The clinical evidence for AI therapeutic tools is real but constrained. Studies of Woebot and similar CBT-based chatbots show genuine reductions in self-reported anxiety and depression symptoms, with effect sizes in the moderate range. These effects are strongest for mild to moderate presentations, and they attenuate significantly for more severe presentations. The evidence base does not support AI as a substitute for human care in moderate-to-severe cases, but the clinical system does not reliably deliver human care in those cases either. The gap between what the evidence supports and what the market deploys is significant: AI therapeutic tools are marketed to and used by populations whose presentations go well beyond mild to moderate, and the marketing language rarely makes this limitation clear.
The regulatory landscape for AI therapeutic tools is globally incoherent. In the United States, the FDA has taken inconsistent positions on which AI mental health tools constitute regulated medical devices and which constitute consumer wellness products, and the line has been drawn in ways that appear to correlate with commercial lobbying more than with clinical risk. In Europe, the AI Act creates some framework for high-risk AI systems in health contexts, but the implementation remains uncertain. In most of the rest of the world, there is no effective regulatory framework at all. The practical consequence is that AI therapeutic tools with highly varied quality, safety protocols, and clinical warrant are deployed at global scale with minimal oversight.
The promise and harm of AI therapists at collective scale ultimately reflect a structural tension that cannot be resolved by improving the technology alone. The promise arises from a social failure: the failure to build mental health care systems that meet the scale of need. The harm arises from the same failure: because human care is unavailable, AI moves in to fill a gap it is not equipped to fill. Addressing the promise and harm together requires not just better AI design — though that matters — but also deliberate investment in the human infrastructure of mental health care alongside, not instead of, AI tools. The two are not alternatives; they are complements in a system that currently lacks both.