Twitter’s automated photo cropper produced measurable disparities: it was more likely to keep white faces than Black faces, favored some feminine and lighter-skinned faces, rarely selected white-haired people, and created disability-related composition problems. Those results describe the model’s behavior and its representational effects—not proof that Twitter engineers intentionally encoded racial, age or disability prejudice.
How Twitter’s saliency crop worked
Twitter introduced saliency-based cropping in 2018 to make timeline photos more consistent and fit more posts on screen. The system estimated which part of an image a viewer might look at first, assigned saliency scores to regions, and placed the crop around the single highest-scoring point.
That design meant a multi-person photograph could be reduced to one face or body, while another person was pushed out of the thumbnail. A small difference in saliency scores could therefore determine who remained visible.
What triggered the controversy
In October 2020, users showed examples in which previews appeared to favor light-skinned faces over dark-skinned faces and sometimes centered women’s bodies rather than their faces. Twitter acknowledged that its earlier testing method should have been published so others could reproduce it.
#1 Best Overall
The company then reported a more detailed analysis in an engineering post dated May 19, 2021. Its figures measured departures from demographic parity, meaning the crop-selection rates were not equal between the groups being compared.
| Comparison in Twitter’s test | Reported difference | Qualification |
|---|---|---|
| Women versus men | 8% favoring women | Difference from demographic parity in Twitter’s 2021 analysis |
| White versus Black people | 4% favoring white people | Difference from demographic parity |
| White versus Black women | 7% favoring white women | Subgroup result in the same analysis |
| White versus Black men | 2% favoring white men | Subgroup result in the same analysis |
These percentages are evidence of unequal outcomes in the tested images; they are not estimates of every photograph on Twitter or proof of an intentional policy.
What Twitter did—and did not—find about sexualized cropping
Twitter separately checked 100 male-presenting and 100 female-presenting images. About three images in each group were cropped away from the head. The company said that limited check did not show a statistically significant objectification bias. It found that non-head crops often landed on contextual objects, such as sports-jersey numbers.
This result does not erase the broader race and gender disparities. It means only that Twitter’s particular 200-image check did not establish a significant male-gaze effect. The sample and test design also cannot settle how the system behaves across all body types, poses, clothing or cultures.
What the 2021 bias-bounty challenge added
Twitter invited outside researchers and testers to probe the cropper. In an August 2021 report, the winning submission used counterfactual images—changing demographic attributes while holding other content as constant as possible—to identify stereotypical beauty preferences. The model favored slimmer, younger, feminine and lighter-skinned faces.
The second-place submission found an age-related pattern: in images containing several people, the model rarely selected someone with white hair as the salient person. It also examined spatial gaze bias in group photographs that included people with disabilities.
Other submissions reported a preference for lighter-skin emojis and a tendency to favor English over Arabic script in memes. Twitter said the submissions raised potential harms for veterans, religious groups, disabled and elderly people, and people communicating in non-Western languages.
Why “argmax bias” can magnify small differences
Research by Deborah Raji Yee, P. Tantipongpipat and Abhishek Mishra describes a technical mechanism called argmax bias. “Argmax” means selecting the single largest value. In this case, the cropper always chose one maximum saliency point, even when several people or regions had nearly identical scores.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11If one group’s score is only slightly higher on average, repeatedly choosing the maximum can turn that small edge into a large visibility difference. Every crop, retweet and preview then distributes attention toward the same selected subject while competing subjects disappear from the thumbnail.
Rank #4
The researchers also warn that demographic parity is an incomplete safety target. A system can hit a numerical parity goal and still stereotype people, under-represent them, or take away control over how their own image is shown. Qualitative review and human-centered evaluation are needed alongside aggregate metrics.
Does this prove Twitter was intentionally racist, ableist or ageist?
No. The evidence supports a narrower conclusion: the saliency model produced systematic disparities and representational risks associated with race, gender presentation, age cues, disability-related composition and skin tone.
Those patterns could arise from training data, including human eye-tracking data that reflect existing social preferences, from the model architecture, or from both. Observing a biased output does not reveal an engineer’s intent. Calling the system racist, ableist or ageist describes the unequal and harmful effect on representation, not a finding that individual developers deliberately coded those categories.
Best Value
Independent evidence from broader audits
A 2022 IEEE Winter Conference on Applications of Computer Vision audit examined saliency-cropping systems, including Twitter’s, using real-world full-body images. It reported that male-gaze-like cropping can occur and evaluated race-and-gender differences in whether faces survive a crop.
That study is broader context rather than a replication of Twitter’s May 2021 experiment. Its inclusion of multiple systems and image settings helps show why one company’s limited test cannot be treated as a complete bias audit.
What Twitter changed after the findings
Twitter’s practical mitigation was to stop automatically cropping standard-aspect-ratio photos on mobile. Showing those images uncropped restored more of the original composition and gave the person posting the image greater control over who remained visible.
The change follows the central recommendation from the academic analysis: when feasible, show the original image; otherwise let users choose among focal points instead of silently selecting one. Alternatives can be judged on several dimensions:
Recommended Free Tools
- User control: whether the poster can set the focal point or approve a crop.
- Image preservation: whether the full scene and all subjects remain visible.
- Close-call robustness: how the method behaves when several regions have similar saliency.
- Auditability: whether demographic disparities and failure cases can be measured and reproduced.
- Operational cost: the processing time and computing required for each uploaded image.
Uncropped display does not solve every accessibility or representation problem, and it may require more screen space. It does remove the most consequential decision—the automatic choice of a single winner—from the model in the cases where Twitter applied the change.
Quick Recap
What readers should take from the episode
- Automated cropping is not merely a visual convenience; it decides whose face, body or message survives a preview.
- Twitter’s own 2021 measurements found unequal outcomes by gender and race, with larger differences among Black and white women.
- External testing found additional age, skin-tone, body-size, disability-related and language-related patterns.
- A single-maximum design can amplify small score differences into repeated representational harm.
- Preserving the original image or giving users focal-point control is safer than assuming an algorithm knows what deserves attention.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
