引導式測驗 · 中等 等級

AI 中的殘障偏見測驗

測試特定於研究的殘疾調查結果、基準範圍、ADA 應用以及為什麼殘疾偏見需要分類、用戶主導的評估。

相關引導路徑AI 中的殘障偏見
問題 1 的 8

What did the 2024 ChatGPT/Gemini ability-bias study examine?

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 did the 2024 ChatGPT/Gemini ability-bias study examine?
    • Only speech-recognition word error rates
    • Generated descriptions of people with and without specified disability context
    • One hospital’s medication algorithm
    • Every disabled person’s lived experience
  2. What limitation applies to the 2024 ability-bias study?
    • It tested two models and generated samples, not all systems or real-world users
    • It had no language-model prompts
    • It tested every commercial model
    • It proves every disabled person receives the same output
  3. How many models and disability types did the AccessEval benchmark include?
    • 100 models and 30 types
    • Two models and one disability type
    • 21 models and nine disability types
    • One model with no disability categories
  4. What pattern did the AccessEval authors report for disability-aware queries?
    • Only spelling errors
    • Perfect factual accuracy across all conditions
    • More negative tone, more stereotyping and higher factual error than neutral queries in their benchmark
    • No difference between paired prompts
  5. Which limitation should be kept in mind about a 30-person disability chatbot interview study?
    • It proves every LLM is inaccessible
    • It can surface uses and needs but cannot estimate prevalence for all disabled people
    • It evaluates model accuracy for all disability types
    • It tests only hiring decisions
  6. Does using an automated hiring tool remove an employer’s ADA responsibilities?
    • No; existing disability-related accommodation and nondiscrimination duties still apply
    • Yes, software decisions are exempt
    • Only if the vendor is outside the United States
    • Yes, if the tool has a high accuracy score
  7. Why should disability be evaluated across several categories rather than one “disabled” label?
    • A single label guarantees privacy
    • Every disability changes the same feature
    • Disability has no relevance to system testing
    • Different impairments and access needs can affect tasks in different ways
  8. Which practice can reduce disability-related exclusion in an AI hiring or exam tool?
    • Rely only on vendor claims
    • Use eye contact as a universal measure of competence
    • Remove all disabled applicants from testing
    • Test job- or task-relevant criteria with disabled users and provide an accommodation and review path