This issue is on OSX Monterey using Safari 16.1 When doing accessibility testing, we found a very unusual bug. Voiceover reads out AI generated text in images rendered by Safari, even though the image is marked as decorative and should not have alt text. How to reproduce: 1) Go to https://codepen.io/zoltan-dulac/pen/NWOBEQK using Safari/Voiceover 2) Tab into the "dummy link" 3) Use Voiceover's reading mode to read the content (i.e. press CAPSLOCK+A). Instead of ignoring the text like it should, it reads the Shadow DOM content that has text that looks like it was generated by an OCR function that attempts to read the text inside the image (i.e. the text on the van's door). There is another related issue that involves tabbing into a link that contains the image: 1) Go to same page https://codepen.io/zoltan-dulac/pen/NWOBEQK using Safari/Voiceover 2) Tab into the "dummy link" 3) Tab into the link. Sometimes the screenreader reads the URL of the link. Sometimes it reads the OCR text. If you tab out of it and tab backwards into it again, it may give different results. It is very inconsistent. I tagged this as a major issue since decorative images reading incorrect information can be very confusing to screen reader users and result in misunderstanding the content of a webpage.