Behavioral training conducted in health conversations improved the model's performance in other areas as well

Researchers found that when instruction in desirable behavior was limited to health-related conversations for an AI model, the model nevertheless improved on non-health evaluations — for example, alignment errors, deception, and reward exploitation; this indicates that learned behavior transfers to other domains, which is important for the effectiveness of safety interventions.