Concept Safety

We can think of concepts and ontologies as tools for achieving our values. How does this influence our thinking on AI safety?