Imagine a building where a smoke alarm goes off. A guard walks over, sees the smoke, and decides it isn't serious enough to call the fire department. He switches off that one alarm and goes back to his desk. The person who set it off walks into the next room and keeps going. That is what British Columbia says happened here. OpenAI's own people read the shooter's conversations, found them alarming enough to review, and let it end there. The warning worked. The people holding it chose not to pass it on.
Here's why that's a big deal: most fights about AI safety are about machines getting something wrong. This one isn't. The system caught the danger, and a human decision kept it from reaching the police, who already knew this family and had taken guns out of that home before. Now a government is suing to prove there is no special exemption for AI companies when a warning they hold never leaves the building. If that argument wins, every company that watches conversations for danger will owe the public a phone call, not just a review.
