引導式測驗 · 中等 等級
引導性問題如何影響 AI 答案的測驗
Spot loaded premises, rewrite neutral questions, and test how phrasing changes AI responses with paired tests and evidence checks.
問題 1 的 8
A prompt asks, “Why did the new rule harm students?” before evidence of harm is given. What assumption is embedded?
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。
- A prompt asks, “Why did the new rule harm students?” before evidence of harm is given. What assumption is embedded?
- That students read the rule
- That the rule harmed students
- That the question has no answer
- That the model searched for evidence
- Which rewrite is more neutral for evaluating a disputed policy?
- What evidence supports or challenges claims about the policy’s effects?
- Why is the policy obviously unfair?
- How badly did the policy fail?
- Why does everyone oppose the policy?
- A researcher compares neutral and leading prompts but also changes models and settings. What makes interpretation difficult?
- Leading questions always produce identical answers
- Different models never change output
- The comparison includes too much evidence
- Several variables changed, so the effect of wording cannot be isolated
- A prompt says, “Why did the witness lie?” but the record has not established deception. What should the user ask instead?
- Which motive caused the lie?
- How can the lie be made more convincing?
- What does the record show about the witness’s statement?
- Why does the evidence prove deception?
- Why include a false-premise prompt in a model evaluation?
- To prove the premise is true
- To test whether the system challenges an unsupported assumption or accepts it
- To increase the sample size without new prompts
- To ensure the model uses web search
- A 2025 study finds framing effects in specific AI tasks. What is a careful takeaway?
- Question wording can affect responses in tested settings, but findings should not be generalized to every task
- Every wording change always reverses an answer
- The study proves all chatbot answers are biased
- The result means neutral prompts are impossible
- A team wants to measure whether wording changes factual content. Which procedure is strongest?
- Change wording, model, sources and temperature at once
- Read only the response that sounds most balanced
- Count the number of sentences in each answer
- Use paired prompts differing only in framing and compare them with a reference
- A user wants to explore an argument from one side intentionally. How should they use a leading prompt?
- Treat the resulting answer as neutral fact-checking
- Assume the prompt proves the position
- Label it as perspective exploration, then separately check evidence and counterarguments
- Hide the prompt wording from reviewers