Here's Vinay running tests on that Twitter bias. While I believe the first dataset isn't necessarily representative of real-world examples (always plain white backgrounds in this data) it's good that more tests are being run and that it's open source. x.com/vinayprabhu/st…
Twitter are also going to open source their work: x.com/lizkelley/stat…
One big issue for me is also the lack of user control in this.