引導式測驗 · 中等 等級

偏見獎金與演算法漏洞獎金測驗

涵蓋 Twitter 的 2021 年有限作物模型挑戰、賞金設計、可重複性、激勵和限製作為自願審核方法。

問題 1 的 8

What model did Twitter’s 2021 algorithmic bias bounty challenge target?

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 model did Twitter’s 2021 algorithmic bias bounty challenge target?
    • Every post recommendation system on Twitter
    • Its image saliency and automatic-cropping model
    • A facial-recognition system
    • A bank credit model
  2. What did participants receive access to for the announced challenge?
    • A public demo API that returned example crops without exposing the model.
    • Generated image examples and evaluation labels without access to the generation code.
    • The saliency model and crop-generation code
    • A small set of saliency scores without the crop-generation code.
  3. Why is a bias bounty not equivalent to a security vulnerability bounty?
    • Security reports never need evidence
    • A bias bounty is always a software exploit test
    • A bias bounty automatically changes the law
    • Potential social harms need context, affected-population evidence and choices about fairness criteria
  4. If no researcher reports a bias during a bounty, what limitation remains?
    • It certifies legal compliance
    • It proves every group is treated equally
    • It shows no one used the system
    • The result does not prove the model is free of bias or harm
  5. Which detail makes a submitted bias finding more reproducible?
    • A conclusion without test data
    • A screenshot with no context
    • Model version, test population, metric, comparison and limitations
    • A list of protected groups only
  6. Why might a prize contest overrepresent easy-to-demonstrate harms?
    • Time, access and prize criteria can favor findings that are easy to test over less visible harms
    • Awards eliminate all conflicts of interest
    • Every harm is equally measurable
    • A model’s training data is always public
  7. Which practice helps protect affected people and good-faith researchers?
    • Encourage participants to scrape private accounts
    • Clear authorized scope, privacy safeguards, confidential disclosure and a response process
    • Publish raw sensitive data with each report
    • Make acceptance depend only on publicity
  8. How should an organization interpret a successful bias bounty?
    • As bounded evidence for risk management, not a fairness certificate
    • As a replacement for appeals and internal testing
    • As proof that all company models are unbiased
    • As legal approval from regulators