AI programs have safety restrictions built in to prevent them from saying offensive or dangerous things. It doesn’t always work.