Can ‘super intelligence’ and a non-binding safety pact solve AI’s image problem?