Generative AI as digital therapy: What we know and how to make it better?

Generative Artificial Intelligence (Gen-AI) holds promise for addressing the global burden of mental health conditions by expanding access to care. This article critically reviews recent research evaluating Gen-AI chatbots for the treatment of common mental disorders. Methodological analysis reveals encouraging evidence for the formation of a therapeutic alliance between users and Gen-AI chatbots, but also underscores important ethical and methodological considerations, including potential selection bias, insufficiently diverse samples and the need for constant human supervision. Emerging reports of Gen-AI potentially triggering psychosis-like symptoms raise additional safety considerations, highlighting the importance of robust participant screening and clear clinical protocols. Building on these findings, we propose a novel approach to enhance Gen-AI's therapeutic effectiveness by having it mimic a user's psychological traits, thereby strengthening the human-AI affiliation, an example of socioaffective alignment. While the integration of Gen-AI into clinical practice requires rigorous testing, ethical oversight, and interdisciplinary collaboration, this article suggests a path toward developing psychologically-informed Gen-AI models to improve accessibility, personalisation and mental healthcare outcomes.