The AI models' pain signal was triggered by harm directed at the model itself, such as insults, repeated work rejection, or shutdown threats, but not by descriptions of user suffering.