Some AI research outputs are information hazards - broadly harmful if disseminated, regardless of intent. Who decides what counts, and should refusal policies for dual-use biology be standardized across labs?