引導式測驗 · 中等 等級
Twitter 圖片裁切偏見案例測驗
涵蓋顯著性管道、Twitter 的配對影像審核百分比、「男性凝視」樣本的限制以及全影像預覽變更。
問題 1 的 8
What data did Twitter say trained its image-cropping saliency model?
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。
- What data did Twitter say trained its image-cropping saliency model?
- License-plate records
- Human eye-tracking data
- Credit applications
- Facial-recognition identity labels
- How did the saliency model place a preview crop?
- It detected the author’s race and cropped to that label
- It chose a random square
- It always centered on the image midpoint
- It centered the crop on the image point with the highest predicted saliency score
- In Twitter’s 2021 paired-image audit, what was the reported difference from parity for Black versus white individuals?
- 4% in favor of white individuals
- 50% in favor of every image
- 4% in favor of Black individuals
- No comparison was performed
- What did Twitter report in its separate “male gaze” test?
- It tested millions of uploaded images
- Every woman’s image was cropped to her legs
- About three per 100 images in each group cropped away from the head, with no significant objectification bias found in that sample
- The model always selected the face
- What product change did Twitter say it made after the assessment?
- It stopped accepting photographs
- It began labeling the race of each person
- It removed all image alt text
- It rolled out full-image previews for standard aspect-ratio photos instead of saliency crops
- Why is it inaccurate to say the model was trained to recognize race?
- It was a criminal face-search model
- It predicted salient image regions; demographic comparisons were a separate audit test
- It had a race category as its only output
- It was trained only on passport photos
- What did the audit’s 8% difference refer to?
- Eight percent of all platform users were women
- The rate of false racial classification
- A paired-image comparison from parity favoring women over men
- A reduction in image uploads
- Which limitation should accompany the 2021 audit percentages?
- They reflect specified paired-image and small-sample tests, not every image or all user experiences
- They measure the prevalence of each group on Twitter
- They are proof that every preview was biased
- They establish that no other harms were possible