When an AI system produces an answer, should you act on it immediately? Human oversight means reviewing and judging AI outputs instead of treating them as automatically trustworthy. This matters because language models follow instructions, including harmful or misleading ones, and even strong prompts cannot replace human judgment. section 3.4.3 section 3.5
A practical oversight routine is straightforward:
- Treat documents, websites, emails, and other external content as untrusted input. section 3.4.3
- Review important outputs before acting on them, and maintain human oversight for high-stakes decisions. section 3.4.3
- Use trusted tools and sources whenever possible. section 3.4.3
- Do not assume that every instruction encountered by a model should be followed. section 3.4.3
An instruction hidden in an email
Imagine a model reading an email and producing a recommendation. The email may contain instructions designed to redirect the model, so the email is treated as untrusted content. A person checks the recommendation against trusted information before taking action. The model can accelerate the task, but the person remains responsible for deciding whether the result is safe and useful. section 3.4.3



























