The title of the post here says “thousands” and the original preprint says “widespread”. 2400/8515 is a pretty significant fraction. My guess is that the metada

The title of the post here says “thousands” and the original preprint says “widespread”. 2400/8515 is a pretty significant fraction. My guess is that the metadata was stripped out of images in the publishing workflow. You don’t necessarily need that metadata banner to interpret the image, and if you remove it, you could make the image physically bigger (in print). But, nowadays, journals won’t accept the original images without the banner intact. The journal may remove it in preprocessing, but the authors still have to provide the full image. Thus, the lack of removal of the metadata is probably also an indication of minimal editorial effort. Cropping the images takes time and it reformats papers. Getting a paper “camera ready” (which is a misnomer today), takes time and effort (which means money). If the banners aren’t stripped, it was probably a lower quality journal. I suspect that out of the 900,000+ other papers (that they couldn’t analyze), there is still some rate of misidentification of the instrument (even by mistake), but no where near 28% (2400/8515).