Firms including OpenAI have developed safety tools to constrain AI decision-making. But there’s a limit to what these tools can accomplish.