Twitter replaced its biased crop algorithm with full photos, handing the choice to users

In 2018 Twitter began cropping timeline photos with a saliency model: software trained on human eye-tracking estimated the region a viewer would look at first and centered the crop there, keeping timelines consistent and dense. In October 2020 users reported that the automatic crop did not serve people equitably — it appeared to favor white faces over Black ones and to crop women's bodies in ways men's were not. Twitter committed to analyzing the model again for bias.

Twitter's META and Content Understanding teams built a controlled experiment: random pairs of images of people of different races and genders were compared, and each image should have had a 50% chance of being chosen as the crop focus if the model were demographically neutral. The measured skews were systematic — a 4-point gap against Black individuals when paired with white individuals, an 8-point gap in favor of women in male-female pairs, and a 7-point gap in favor of white women versus Black women. Tests for male-gaze cropping found no significant objectification.

The researchers concluded that retraining toward parity would not fix the harm. Even a perfectly balanced model still overrides how people choose to represent themselves, and representational harm is not a statistical error; the paper noted that cropping to a single most-salient point could amplify disparities. Twitter's public conclusion was that not everything on the platform is a good candidate for an algorithm, and that how an image is cropped is a decision best made by people.

In March 2021 Twitter began showing standard aspect-ratio photos in full on iOS and Android instead of running them through the saliency crop, and after positive feedback launched the change to everyone by 19 May. The composer gained a true preview so authors could see the published image before posting, and the company published its analysis as a paper with open reproduction code, framing the whole exercise as public accountability.

Fairness metrics can't fix a crop that overrides self-representation — remove the model and let people choose

Statistical fairness could not solve representational harm: even a parity-neutral crop still decides how a person appears, so the model itself, not its parameters, was the problem.

A controlled paired-image protocol turned user anecdotes into measured, reproducible numbers, with race and gender gaps quantified and published alongside the code used to compute them.

Showing full images removed the whole class of cropping errors, including the disparity-amplifying single-point rule the paper named argmax bias, with no new model left to re-audit.

The composer preview gave authors control at the moment of publishing, so the fix restored agency instead of merely swapping one crop rule for another.

Releasing the paper and code made the response an auditable public record, which is why the announcement read as accountability rather than spin.

Before retraining a model that decides how people appear, ask if the output is a prediction; when it is self-representation, remove the model and hand the choice to the person.

By 19 May 2021 the full-image display had shipped to every iOS and Android user, and Twitter said the release reduced its dependency on ML for cropping. The analysis paper (arXiv 2105.08667, May 2021) was accepted to CSCW 2021 and appeared in PACM HCI in October 2021, with reproduction code in the twitter-research image-crop-analysis repository. The Hacker News discussion drew 197 points and 133 comments; most readers argued a manual cropper should have existed from the start, while some defended face-aware auto-cropping when an author sets no crop.