The other half of AI safety