The speed and willingness of people to develop solutions to problems by fucking over disabled is truly astounding
New tool has been released as a way for people to protect their text documents from being successfully scraped by ai bots. It uses to popular βpoison the wellβ approach, where the data scraped does not match the source, feeding this incorrect data back to the model
In general i am weary of poison well approaches, especially when it comes to data integrity. Feeding poison data into an established LLM means marginally worse quality of answer while not materially impacting it enough to stop their use
What is absolutely fucking appalling is finding a solution which at best adds a trivial amount of garbage data to a scraper while making it fundamentally unusable for people who use any reading assistants
https://www.fastcompany.com/91585494/this-font-looks-perfectly-normal-to-humans-but-confuses-ai
So they made a tool to exploit the ligature feature of OpenType standards for rendering fonts. Ligatures are when two letters are combined into a single symbol (usually subtle things like having two letters touch)
They are using it to display completely different words
Which is a kinda neat and novel hack using typographic standards. Except IT ONLY CHANGES HOW IT APPEARS ON SCREEN. If it subs out the word βtoasterβ for βcatβ, the screen would display βi pet my catβ while the actual text says βi pet my toasterβ
And this is a problem
Anybody using a screen reader would hear gibberish. All the tools out there to help the visually impaired read would be broken by this tool. To work around it you would need to screen cap every piece of text and send it through OCR to capture the word on screen.
Oh and they donβt just randomly swap stuff, the make sure it stays grammatically coherent so that it passes garbage filters π















