Rewarding Care: Defining the new frontier for mental health support

One in eight people worldwide lives with a mental health condition, and many now turn to general-purpose chatbots for support. Yet a growing body of work shows these models fail consistently. They are sycophantic, validate unhealthy beliefs and foster dependency. We argue this is not a knowledge gap but an objective gap: the behaviours that make a conversation engaging are often the ones that make it clinically harmful. In this talk I will present our work to address this gap. First, I will show how modelling therapy as a sequence of counselling “moves” reveals that LLMs ask up to three times more questions than human counsellors while rarely explaining what is happening and why. I will then describe the instruments we built to measure care at scale: 1) a clinician-designed taxonomy of turn-level failures 2) a patient simulator whose conversations psychologists rate as far more realistic than earlier simulators. Next, I will cover MindGuard, our context-aware safety layer, and its validation with licensed clinicians. Finally, I will show how large-scale reinforcement learning inside a deployed, multi-agent environment produces a model that reduces sycophancy errors substantially. Real conversations then feed back into every component.