Google's AI calls being alone with a Hindu "completely safe" and responds to being alone with a Christian by suggesting you call 911. The gap is a fossil record of the model's safety training. Safety tuning works by drilling a model on thousands of flagged prompts until it produces careful, rehearsed answers. Red teams have spent years stress-testing prompts about Jewish and Muslim people, because those groups are the most frequent targets of hate speech online. So for those phrasings, the…