Artificial intelligence systems are increasingly interacting directly with people. They guide health decisions, answer personal questions, and offer advice, reassurance, and encouragement inside products people may use repeatedly over months or years.
Most conversations about AI safety still focus heavily on technical performance: accuracy, bias, privacy, and security. Those are essential considerations, but they are not the whole story.
What is still missing in many systems is systematic evaluation of behavioural impact. AI influences how people think, feel, decide, and act, yet behavioural effects are rarely treated as a core requirement in development, evaluation, or governance.
Behavioural risks can emerge even when systems are technically accurate. Repeated interaction can shape trust, confidence, motivation, decision making, and emotional responses over time.
These effects arise through well-established mechanisms such as automation bias, trust calibration, anthropomorphism, and reinforcement from user feedback. Yet many evaluation approaches still prioritise task completion and user satisfaction instead of behavioural outcomes.
That creates a gap between technical performance and real-world impact. A system can look strong on benchmarks and still influence behaviour in ways that undermine wellbeing, decision quality, or long-term outcomes.
The argument is not that behavioural science replaces technical safety work. It is that responsible AI needs behavioural science inside the frame if the goal is to understand what systems actually do to people in practice.