I think Twitter is intentionally hiding tweets that have images attached containing lots of text.
I noticed this had no interaction and 3 impressions in 8 minutes, and apparently it's completely blocked for other accounts; second image from my alt.
This is extremely alarming. https://twitter.com/atomicthumbs/status/1329930847315660800
I noticed this had no interaction and 3 impressions in 8 minutes, and apparently it's completely blocked for other accounts; second image from my alt.
This is extremely alarming. https://twitter.com/atomicthumbs/status/1329930847315660800
This is the second time I've noticed this specifically, and the first time I've discovered that it's actually blocked, though I suspect I've been subject to this before.
Here's a link to the tweet, let's see what happens https://twitter.com/atomicthumbs/status/1331036152774631426
Here's a link to the tweet, let's see what happens https://twitter.com/atomicthumbs/status/1331036152774631426
The quote tweet looks normal on my end, but the screenshots in the replies show that it's blocked for everyone else. This is designed to be invisible to the people targeted by it.
I performed an experiment, with the blocked images and similarly composed lorem ipsum images, and notable keywords from the blocked tweets in the tweet text. It's OCRing the images; tweets with the blocked screenshots are unviewable, but lorem ipsum's fine https://twitter.com/atomicthumbs/status/1331045904594345984
other possibilities: this is limited to my account, or there's something about the specific composition of my screenshots that's bothering the OCR/ML systems that analyze tweet images. I'm undertaking another experiment https://twitter.com/0x0dada/status/1331046019275120642
I ran another series of tests. It blocks the original Wikipedia image and the original converted to greyscale, but allows reflowed text and the same image with added whitespace around it. It's not OCR, it's ML, and it's a very stupid system doing it https://twitter.com/atomicthumbs/status/1331049945508376576
It's not just my account, either: https://twitter.com/kingdomakrillic/status/1331051735306149890
After more tests, my conclusion is: something, possibly Twitter's new anti-disinfo/anti-Q system, is running a deep learning based image recognition model trained on images they'd like to ban, and because image classifiers aren't meant for that, it's throwing false positives.
Not necessarily malicious, just stupid as hell. However, unlike the 5G corona frequencies thing, this one is actively harmful, since it completely blocks the visibility of random tweets with images containing text.
It's meant to be invisible, but I have followers who like the sort of things I tweet, so I can immediately tell something's wrong. Any Q person with an obsessive Q-based following can tell this as well, and easily adjust the image to bypass it.
It's both harmful and ineffectual!
It's both harmful and ineffectual!
You can also bypass it by linking directly to the image in question; this URL is from the first blocked tweet https://pbs.twimg.com/media/EnjKAgKUcAAEiMc?format=png&name=small
@kingdomakrillic did more tests. apparently the portion of the image that sets off the badly made neural network is the line containing the phrase "McLean, Virginia"... the location of the CIA headquarters
(that is a coincidence, the system in question is almost certainly unable to recognize the content of text in an image, only that it *looks like* something it's been told to block)
proposal: i am going to make a system that blocks conspiracy disinfo
reality: a corporation spent tens of millions of dollars constructing a machine that blocks the encyclopedia, entirely by accident https://twitter.com/vvspire/status/1331069510091157504
reality: a corporation spent tens of millions of dollars constructing a machine that blocks the encyclopedia, entirely by accident https://twitter.com/vvspire/status/1331069510091157504