Fun experiment: if you see a greentext on reddit, try typing out one of the sentences word for word, and pasting it into the reddit search bar. It will show you one post, and that will be the post you copied that sentence from.
If reddit has the ability to mass OCR and index every image for searching, then almost every other platform has as well. So things like screenshotting text and posting it will still get indexed and linked to your account. Insta algos will probably derank your account if you have words like Palestine, Gaza, etc. even if you try to bypass that by screenshotting them instead of typing them.
Another aspect of this is sending images by WhatsApp, only doubly so, because those images end up in the gallery, and most phones nowadays automatically OCR them as well. This increases your attack vector, depending on your threat model, I guess.
My guess would be that the scanning is done at the time of creation and then the text and image description is store in a db. But yeah, one of the first tensorflow android examples they released was for local object detection and that was almost ten years ago, I think. I'm sure it's advanced quite a bit. We'll have narc LLMs installed soon enough.
Probably so. I'd assume that tech companies would want an opportunity to scan your saved pictures for ad-profile data, but also I'm surprised that a different OCR feature works locally and near-instantly
Specifically, my phone running Android 14 has a 'Select' feature that you can activate when switching between active apps, and so far it's been able to copy optically from any app that doesn't privacy-block its contents from appearing in the app switcher
Some googling reveals that, in my case, Samsung has it built in their OneUI and it doesn't have an option to be turned off. Good to know.