Who is at fault when systems powered by AI put humans in danger?