The trials so far, and there are a few now, find the structured ones, the ones built on an actual therapy manual, about as good as a self-help workbook and better than a waiting list, for mild to moderate cases. Not as good as a therapist. The open-ended chatbots haven't been trialled properly, because nobody owns them in a way you can trial. So the honest version of the proposition is: fine for the manual kind, unknown for the one everyone's actually using.