Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Twitter’s image-cropping system produced measurable differences in which people and parts of images remained visible. The company’s 2021 tests found disparities by gender and race; later research and a public bias-bounty challenge identified additional patterns involving age cues, skin tone, disability-related composition and other forms of representation. Those findings show unequal outcomes, not proof that engineers deliberately encoded those categories. Twitter ultimately stopped automatically cropping standard-aspect-ratio photos on mobile, giving users more control over how their pictures appeared.
How Twitter’s image-cropping algorithm worked
Twitter introduced saliency-based cropping in 2018 to make photos fit the timeline consistently and allow more Tweets to appear at a glance. The system estimated which region of an image a viewer might look at first, assigned saliency scores to image areas, and placed the highest-scoring point at the center of the crop.
This approach could be useful when a wide or tall image had to fit a smaller preview: instead of cropping from a fixed edge, the system tried to preserve what it judged most important. But it also meant the algorithm—not the person who posted the image—made a consequential choice about what viewers would see in the timeline preview.
What Twitter’s 2021 tests found
In a May 19, 2021 engineering post, Twitter reported differences from demographic parity in its saliency-cropping test. The company said the results favored women over men by 8%, and white people over Black people by 4%. The reported difference was 7% for white women compared with Black women, and 2% for white men compared with Black men.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitches#1 Best Overall
These figures describe results in Twitter’s reported experiment; they are not a universal score for every image, user or context. They do, however, show why a seemingly neutral image-processing feature drew scrutiny: the predicted focal point could affect whose face or body survived a crop, and the reported outcomes were not evenly distributed across the tested groups.
The separate check for objectification
Twitter also described a limited check of 100 male-presenting and 100 female-presenting images. In each group, about three images were cropped away from the head. The company said it did not find significant evidence of objectification bias in that check; some crops away from the head centered on details such as sports-jersey numbers.
That result is narrower than a general claim that the algorithm could not objectify people. It concerns one company-run check with a particular set of images and does not negate the separate race and gender disparities Twitter reported.
What the bias-bounty challenge added
In August 2021, Twitter published findings from a bias-bounty challenge that invited participants to identify harms in the cropping system. The winning submission used counterfactual comparisons and found that the model tended to favor slimmer, younger, feminine and lighter-skinned faces. A second-place submission found that the model rarely selected white-haired people as the salient person in images containing multiple faces, and examined spatial gaze bias in group photos that included people with disabilities.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Other submissions identified a preference for lighter-skin-tone emoji and a bias favoring English over Arabic script in memes. Twitter also said participants surfaced potential harms affecting veterans, religious groups, disabled people, older people and people communicating in non-Western languages. These are different kinds of findings: some concern which person or feature a crop selects, while others concern how language or identity is represented in an image.
Why the phrase “racist, ableist and ageist” needs context
The labels describe observed disparities and representational risks. They do not establish that Twitter’s engineers intentionally designed the model to discriminate against Black people, disabled people or older people. Twitter’s bounty report said the patterns appeared embedded in the saliency model and could have been learned from human eye-tracking data. A system can reproduce patterns in its training signals or evaluation environment without anyone explicitly instructing it to prefer one demographic group.
That distinction does not make the effects trivial. If a model repeatedly crops out a person’s face, privileges one appearance over another, or makes a disability-related subject less visible, the result can shape how people are represented and encountered—even when the system is optimizing for a different goal, such as a tidy preview.
How “argmax bias” can magnify small differences
Researchers Yee, Tantipongpipat and Mishra described a mechanism they call “argmax bias.” Twitter’s method selected the single region with the highest saliency score. When multiple people or image regions have similar scores, a small difference between them can determine which one wins. Repeating that all-or-nothing choice across many images can turn a modest underlying preference into a systematic difference in what remains visible.
Free tools Windows power users keep installed
One-click scans. No signup required.
The issue is not simply whether a model has a high or low score on one demographic-parity measure. A system might meet a numerical target and still stereotype people, omit them from important parts of a composition, or deny the person who posted an image control over its presentation. The researchers therefore argued that quantitative parity measures should be considered alongside qualitative and human-centered evaluation.
Rank #4
What changed after the controversy
Twitter’s mitigation was to display standard-aspect-ratio photos uncropped on mobile, rather than automatically applying the saliency crop to those images. This gave the person posting the photo more agency over the image’s composition and avoided the model’s single-point selection for that case.
The change was not a claim that every image in every Twitter surface would always appear in its original dimensions. Twitter described the mobile behavior for standard-aspect-ratio photos; the available account does not establish that all formats, clients or display contexts were covered identically.
What independent research says
A broader audit presented at WACV 2022 examined saliency-cropping systems, including Twitter’s. It reported that male-gaze-like cropping can occur in real-world full-body images and assessed race-and-gender disparities in whether faces survive a crop. This provides independent context for the concerns, but it is a broader audit rather than the same experiment as Twitter’s 2021 engineering tests.
Best Value
What a less harmful design would prioritize
The central design lesson is to avoid treating the algorithm’s preferred focal point as the only valid way to present an image. The researchers recommended removing saliency-based cropping in favor of approaches that preserve user agency. Options they discussed include showing the original image where possible and letting users choose among candidate focal points.
- User control: Let the person posting the image decide what the preview emphasizes, or preserve the original composition when practical.
- More than one candidate: If a crop is necessary, offer plausible focal points rather than letting one top-scoring region silently determine the result.
- Broader evaluation: Test which people and features survive the crop across demographic groups, while also evaluating stereotyping and representational harm that a single parity number may miss.
- Reproducible methods: Publish enough about test construction and evaluation for others to understand and reproduce the findings. In October 2020, after users raised concerns, Twitter acknowledged that its earlier test method should have been published for reproducibility.
These principles apply beyond a single social network: any automated preview system makes editorial choices when it decides what a viewer sees and what is cut off.
Quick Recap
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.
