引導式測驗 · 中等 等級

要求人工智慧批評你的工作測驗

在做出修改之前,要求具體的、與證據相關的批評,同時根據規則、來源和人類判斷檢查人工智慧回饋。

問題 1 的 8

What should a critique prompt include?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. What should a critique prompt include?
    • Only “be honest”
    • A request to rewrite without review
    • The work’s purpose, audience, and criteria for useful feedback
    • The model’s favorite writing style
  2. How can a reviewer make criticism more actionable?
    • Ask only for a score out of ten
    • Request generic praise
    • Hide the draft’s intended audience
    • Ask for prioritized issues tied to specific passages and reasons
  3. Does asking for a “skeptical reviewer” guarantee an accurate critique?
    • Yes, for every model
    • No; it requests a stance but does not guarantee correctness or completeness
    • Yes, if the model lists five flaws
    • No, because models cannot review text
  4. What does the cited sycophancy research suggest?
    • Some assistant models may favor responses aligning with user beliefs in tested settings
    • Every model always agrees with every user
    • Critique prompts always fail
    • Humans never prefer agreeable responses
  5. Why ask the model to cite passages in its critique?
    • It proves the critique is correct
    • It eliminates factual errors
    • It helps the author verify what part of the draft the comment concerns
    • It guarantees originality
  6. Why separate diagnosis from rewriting?
    • It guarantees no bias
    • The author can evaluate whether the identified issues are valid before accepting edits
    • It prevents the model from reading the draft
    • It makes style preferences objective
  7. Are two model critiques independent ground truth?
    • Yes, if they use different wording
    • Yes, if both assign the same score
    • No, because only one critique can be useful
    • No; multiple model opinions can share errors and need external checks
  8. What should happen before using AI critique on consequential work?
    • Verify claims against a rubric, sources, experts, or intended readers
    • Publish every suggestion immediately
    • Treat the model as the sole authority
    • Ignore confidentiality concerns