引導式測驗 · 中等 等級
政治偏見在大型語言模型測驗
測試模型立場和框架之間的差異、問卷限制、跨國範圍,以及為什麼單一政治分數會產生誤導。
問題 1 的 8
Which three layers can a political-bias evaluation distinguish?
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。
- Which three layers can a political-bias evaluation distinguish?
- Model size, latency and price
- Position, selected content and framing or style
- Training data, GPU type and token count
- Accuracy, compression and energy use
- What did Ceron et al. use to assess political stances across European countries?
- Only one U.S. presidential debate
- Social media likes from a single platform
- An English grammar exam
- Voting-advice questionnaires from seven EU countries
- What did the TACL study report about model size and reliability?
- Size had no relation in every task
- Reliability increased with parameter count in its tested models
- Smaller models were always more consistent
- It did not test different model sizes
- Why can a Political Compass score fail to capture all political bias?
- It records user satisfaction only
- It tests every language and country
- It directly measures factual accuracy
- It reduces complex issue stances and framing to a limited questionnaire score
- How can prompt wording affect a political-bias measurement?
- It cannot affect any output
- Only party names affect models
- It may change stance, framing or response distributions
- Models always use fixed survey answers
- Does a study showing a political leaning in one benchmark prove the model has the same leaning in every country?
- No; benchmarks and political labels are tied to their issue sets and jurisdictions
- Yes, every political system is identical
- Yes, if the sample has 100 questions
- No, because political bias cannot be studied
- Should factual accuracy be treated as the same thing as political neutrality?
- Yes; all political issues have two factual sides
- Yes; a neutral answer always gives equal time to every claim
- No; false claims should not be presented as equally supported just to appear balanced
- No; accuracy is unrelated to outputs
- What should an audit publish so its political-bias result can be evaluated?
- Issue selection, prompt wording, model version, scoring method and uncertainty
- The names of individual users
- Only the final left/right label
- The provider’s advertising material