Imagine that every week you write to your therapist, and someone writes back, warm, attentive, and helpful. Could you tell whether it was an AI? And if you could not, would it matter? We put both questions to the test.

Most people who need psychological treatment never receive it, so this is not a thought experiment for long: when the clinic is closed, people already turn to whatever answers. I take three common objections in turn. That it does not work: across many controlled trials, symptoms come down modestly but reliably. That it is a cold machine: in our trial of internet-based CBT for social anxiety, participants trusted humans more beforehand, yet the bond formed with the AI was at least as strong, and neither who actually wrote the feedback nor what people were told about it changed how much they improved. That it is dangerous: under professional review it was not, while unsupervised companion bots are a different matter entirely.

A second trial, in which the therapy was delivered by AI alone, found benefits that continued to grow after treatment ended. But short trials with willing volunteers and supervised systems support a narrow claim, and no difference found is not the same as proven equal. The useful question is not replacement but combination: reach from the machine, judgment and responsibility from the human.