Narrow AI research can lead to discoveries of more general methods that then apply broadly across domains, and therefore working on narrow AI does present risks because successful solutions often generalize in unexpected ways.