A good start would be forcing them to preface their every response with: “This response is from an AI that can easily get things wrong: …”