Well look at that, it works!!
I’ve found circle to search on android can read both variations depending on how you select the image
Already defeated…
In selection 2, you don’t seem to be selecting an image. You’re selecting text.
In method 1, it’s doing OCR. In method 2, its just reading the raw text data you selected. This is how you can copy text from thing A and Font A, and paste it into thing B with font B (depending on Thing B’s paste settings)
Nice ! I didn’t think the font had a serious purpose before I saw your excellent example…
THIN
BLU
ELINEB*
Here’s the text my phone extracted from that image:
FUCK THE POUXCE
What does this say?
It says “THIN BLUE LINE.”
The word BLUE is split between the second and third lines:
THIN
BLU
ELINE
•••
Edge detection moment
It works at full resolution, but it’s a cat and mouse game. If something like this becomes popular, then all that needs to be done is to start passing images through a filter first (blur, lower resolution, etc.), or just train a new vision model that can pick out both the blurred message as well as the decoy message.
The downloadable font is especially vulnerable since each letter is a straight map, so easily trainable
Person: does something impressive against AI Comments: it’s bad because it’s not perfect
Well, I’m glad that we now have a font that makes it look like I’m not wearing my glasses… when I’m wearing my glasses.
This is art
Most AI systems work by reading the pixels of an image up close
In the case of text aren’t LLM’s reading actual text ?Whats the usecase here ? For any text I dont want to be readable by ai scrapers and such, I need to use this font and render as image(s)?
This is not for the case of using text, it’s for protecting text in images. The problem with adding invisible text is that LLMs can defeat it by screenshotting and using OCR. This tries to plug that loophole.
Top points for effort, but if it’s making it hard to read even 2 words… I’d rather poison/prompt inject via invisible text and not make any humans cross-eyed trying to read that font
Invisible text doesn’t work for images though, which is what this is specific for. Invisible text is also defeated by taking a screenshot, so if you use this too then you have a chance of defeating both LLMs that read the text as well as those that screenshot and OCR.
Well, then glaze or Nightshade for ocr stuff. Point still stands, why make the experience worse for everyone just because it might get used in AI trainimg
This hurts my eyes, I see both at the same time 🤔
I think i might be a cylon. 🤖
We can’t be writing like this, fuck AI. We tailor our language to avoid sounding like it, analysis images to tell if it had been generated and now using fonts to avoid scraping. What a toxic, horrible invention. An invention that freely takes from workers, repackages it, and monetizes it all for a small class of capitalists while depriving the majority of its value created.
Doesn’t work on a dark background ? The decoy text is invisible…
