Twitter says it will investigate after users find its preview of a photo, with a Black person and a white person, more frequently displayed the white person
MashableRachel Kraus
Context & Ripple Effects
Twitter's image previews have been machine-cropped since it adopted machine learning in 2018 to identify the most "salient" part of a photo. That automation ran unexamined until users noticed the crop favoring the white person in photos of a Black and a white person — prompting the investigation this article covers.
Users posting photos of multiple people can no longer assume neutral treatment from the preview crop, and Twitter must publicly examine a feature it had presented as a solved engineering problem.
Second-order effects
Twitter's move from automated saliency to user-controlled cropping shifts effort onto product changes rather than model fixes, and its later confirmation of bias against white people and women turns a user complaint into a documented defect other platforms' similar auto-cropping features now invite scrutiny over.
Third-order effects
The pattern — complaint, internal investigation, disclosed results, then paid external testing via bug bounty — points toward algorithmic bias being handled like security: audited, disclosed, and crowdsourced, rather than treated as an invisible property of deployed models.
The trend: Platform ML features are shifting from opaque automated decisions toward user control and externally incentivized audits as the standard response to demonstrated bias.
Trying a horrible experiment... Which will the Twitter algorithm pick: Mitch McConnell or Barack Obama? https://twitter.com/...
@lizkelley
Liz Kelley on x
thanks to everyone who raised this. we tested for bias before shipping the model and didn't find evidence of racial or gender bias in our testing, but it's clear that we've got more analysis to do. we'll open source our work so others can review and replicate. https://twitter.com…
@blackamazon
@blackamazon on x
How did you test before can we see that? What are human interventions? Was this out through an accessibility check ( because accessibility standards mean it should tell you what you're putting up before you do so so would have caught this..) https://twitter.com/...
@kumailn
Kumail Nanjiani on x
Wow. This would be shocking if it wasn't absolutely predictable. https://twitter.com/...
@alyssam_infosec
Alyssa Miller on x
Doesn't matter how much testing you do, when you don't have diversity in your teams working on these projects, when you don't correctly diversify your training sets, you get this trash. This needs to be addressed much sooner than testing before shipping. Major Fail Twitter! https…
@kimcrayton1
@kimcrayton1 on x
Here we go... Hey @lizkelley, your test did NOT find bias because your TESTS are biased...I'm not sure why this is so hard for y'all to understand Also, @Twitter is a multi-billion dollar org, this work should not rely on the FREE labor of the open source community to get right h…
@paraga
Parag Agrawal on x
This is a very important question. To address it, we did analysis on our model when we shipped it, but needs continuous improvement. Love this public, open, and rigorous test — and eager to learn from this. https://twitter.com/...
@neilcastle
@neilcastle on x
let's see if the twitter image algorithm is sexist as well as racist https://twitter.com/...