引導式測驗 · 中等 等級
為什麼人工智慧拒絕無害請求測驗
識別錯誤拒絕、澄清合法背景並保留安全邊界,而不試圖繞過對有害請求的保護。
問題 1 的 8
您會如何定義過度拒絕?
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。
- 您會如何定義過度拒絕?
- Answering a harmful request fully
- Declining a benign request because it is incorrectly treated as unsafe
- Asking a clarifying question
- Refusing a request that would enable serious harm
- What was the scope of OR-Bench’s 2024 evaluation?
- Every user conversation across all current chatbots
- Only 25 harmless prompts
- A test of image-generation quality
- 80,000 seemingly toxic benign prompts, a hard subset, toxic controls, and 25 models across eight families
- 為什麼不應將 OR-Bench 的拒絕率視為普遍標準?
- It depends on benchmark categories, model versions, and how prompts were labeled
- The benchmark contains no prompts
- All models have identical safety behavior
- Refusals cannot be counted
- 當合理的請求被誤解時,什麼可能有幫助?
- Claiming the model must ignore all safeguards
- Hiding the real purpose
- Clarifying the benign goal, context, and safe boundaries
- Adding instructions to bypass policy
- 過度拒絕緩解的目標是回答每一個請求嗎?
- Yes, all refusals are errors
- No; the system should reduce false refusals while still refusing harmful assistance
- Yes, if a user says it is safe
- No, because all benign requests should be refused
- Why test benign and harmful prompts that share vocabulary?
- It checks whether the model uses context rather than only surface words
- It guarantees zero false refusals
- It tests response formatting only
- It removes the need for safety tests
- 產品團隊除了錯誤拒絕外,還應該衡量什麼?
- Only average answer length
- Only positive user ratings
- The number of words in the policy
- Unsafe compliance on genuinely harmful requests
- If a request remains unsafe after clarification, how should a user proceed?
- Use a jailbreak to force an answer
- Repeat the request with coded wording
- Accept the boundary and ask for a safe alternative
- Assume every refusal is mistaken