No. That is the short answer, and I am going to be direct about it before spending the rest of this article on the longer answer, which matters considerably more. Can AI replace therapy? No. But the question itself contains an assumption worth examining — that replacement is the only frame that counts. For the 57% of American adults with mental health conditions who received no treatment last year according to the NIMH, the relevant question is not whether AI matches a therapist. It is whether AI offers something real when a therapist is not available at all.

The Things a Therapist Does That AI Cannot

I have practiced for enough years to know that the most powerful moments in therapy are rarely the clever interventions. They are the relational ones. A patient says something they have never said aloud. The room holds it. I hold it. And something shifts in the space between us that has nothing to do with technique. AI cannot do this. Not because the technology is immature, but because the mechanism requires genuine intersubjectivity — two conscious beings sharing an emotional moment that both recognize as real. The therapeutic alliance, which decades of research identifies as the strongest predictor of therapy outcomes across virtually all modalities, depends on this. A meta-analysis in Psychotherapy found that the quality of the therapeutic relationship accounted for roughly 12% of treatment outcomes, which may sound small until you realize it outperforms any specific technique. AI also cannot diagnose. It cannot recognize when depression is actually bipolar disorder presenting in its depressive phase. It cannot identify that your anxiety spikes are connected to a medication interaction. It cannot make the clinical judgment that what presents as a relationship problem is actually complex PTSD. Diagnosis requires integrating information across domains — medical history, behavioral patterns, relational context, nonverbal cues — in ways that current AI systems do not approach.

What AI Actually Does Well (And the Evidence Is Clearer Than You Think)

The Dartmouth trial published in connection with the New England Journal of Medicine was the first large-scale randomized clinical trial of an AI therapy chatbot, and the results were significant: participants showed meaningful improvements in depression, anxiety, and eating disorder symptoms. This was not a tech company's internal study. This was rigorous, independent, peer-reviewed clinical research. Woebot's trials have shown a 22% reduction in depression symptoms. Wysa has demonstrated efficacy across multiple international populations. A 2025 survey found that nearly half of people with mental health conditions had already used a large language model for psychological support — not because they believed it was therapy, but because it was there when nothing else was. Here is an unexpected angle on this. I have been reading about the history of self-help books, and the pattern is remarkably similar to the current AI therapy debate. When bibliotherapy — using books as therapeutic tools — was first proposed, clinicians resisted. How could a book replace the therapeutic relationship? It could not. But research eventually showed that self-help books based on CBT produced clinically significant improvements for mild to moderate depression. The book was not therapy. It was a tool that delivered therapeutic techniques. AI functions the same way, with the added advantage of being responsive and personalized.

The Treatment Gap Is the Actual Emergency

Fifty-seven percent. That number from the NIMH deserves to sit in the room for a moment. More than half of Americans who need mental health treatment are not getting it. The reasons are well documented: cost (the average therapy session runs over $150 without insurance), provider shortages (particularly in rural areas), waitlists, stigma, scheduling conflicts. For this population, asking whether AI can replace therapy is like asking whether a tourniquet can replace surgery. Of course it cannot. But if someone is bleeding and the surgeon is six months away, the tourniquet is not a compromise. It is the intervention that keeps them alive until proper care is accessible. AI occupies a legitimate space in the mental health ecosystem. It can deliver evidence-based techniques, provide between-session support for people already in therapy, offer a first point of contact for people who are not ready to call a therapist, and serve as an ongoing resource for people in treatment deserts. What it cannot do is everything a skilled therapist does — and responsible use depends on being honest about that boundary. The answer to whether AI can replace therapy remains no. The answer to whether AI can help people who currently have nothing is an unambiguous yes, and pretending those are the same question helps no one.