Behavioural safety is the extent to which an interaction makes a patient more or less likely to take a clinically appropriate action. It covers tone, timing, framing, reassurance, escalation, refusal, and reinforcement.
It is not separate from clinical safety, and it is not softer than it. A patient who is reassured at the wrong moment, or encouraged in a behaviour that looks like progress and is not, can come to harm as directly as one given incorrect information. The difference is that the words themselves look fine.
That is what makes it hard to find. Most safety review asks whether the AI said something accurate. Behavioural safety asks what the patient did next. Reviewing transcripts sentence by sentence tends to miss it, because each individual message reads as acceptable and the problem only appears in what follows.
It matters most where an AI system is patient-facing, conversational, and embedded in care, which is where most health AI is heading.
Where this is explored in depth