Reminds me of an anti-crawl mechanism I encountered some time ago in a financial data provider's website: for all numbers in the table, a special font is used where 0~9 are randomly rendered as different chars (e.g. '0' is rendered as 5, '1' is rendered as 8, etc.). The backend server returns the "encoded" chars, and is then correctly "decoded" by the font. The font changes after each reload. So humans always see the correct numbers, but when some crawler uses the HTML source, the numbers are incorrect.
Everyone rushing to make their content AI-friendly made me want to figure out how to make content AI-unfriendly. Basically human-written words meant for human eyes only.
So I built "SoulsOnly.ttf": a font for humans not AI, and keyboard firmware to type in it.
The implementation of a font can be "hacked" to make what looks like gobbledegook to a computer, render as legible to humans. Copying and pasting text written in the font into AI to summarize is almost impossible. And to avoid AI image analysis, a version of the font can be loaded with the glyphs scattered and require a simple "focus" interaction by the reader to begin reading. Note that a sufficiently prompted AI agent can definitely read this, so it's not meant to be cryptographically sound, more just unfriendly to the common AI reader!
It's actually not really easy for AI, without the agent doing some actual coding itself to reverse engineer the font file, or to take screenshots at different variable font intervals to zone in on the "focused" version of the variable font. All of that being said, the intention (beyond just having fun creating it) was to make it AI "unfriendly" so AI bots doing broad quick reads of it are going to be left with gobbledegook encoded characters.
Most llms can equally engage with text in picture form as text in token form. In fact my initial research on this (later corroborated by actual published papers) indicate that this is a cheap way to save on tokens.
Oh interesting and good to know on the token savings with this technique. My test with claude had it use vision and then programmatically test different variable font input variables (mimicking the user scrub interaction) until it was able to OCR it.
As they said in the comment you replied to: "Note that a sufficiently prompted AI agent can definitely read this, so it's not meant to be cryptographically sound, more just unfriendly to the common AI reader!"
Reminds me of an anti-crawl mechanism I encountered some time ago in a financial data provider's website: for all numbers in the table, a special font is used where 0~9 are randomly rendered as different chars (e.g. '0' is rendered as 5, '1' is rendered as 8, etc.). The backend server returns the "encoded" chars, and is then correctly "decoded" by the font. The font changes after each reload. So humans always see the correct numbers, but when some crawler uses the HTML source, the numbers are incorrect.
How would a human copying/pasting a number work?
In my experience (PDF contract sent by a house seller), copy paste was broken.
That said, after 15 minutes of gently massaging the PDF with claude, it was pretty easy to drop the substitutions and restore the original text.
Everyone rushing to make their content AI-friendly made me want to figure out how to make content AI-unfriendly. Basically human-written words meant for human eyes only.
So I built "SoulsOnly.ttf": a font for humans not AI, and keyboard firmware to type in it.
The implementation of a font can be "hacked" to make what looks like gobbledegook to a computer, render as legible to humans. Copying and pasting text written in the font into AI to summarize is almost impossible. And to avoid AI image analysis, a version of the font can be loaded with the glyphs scattered and require a simple "focus" interaction by the reader to begin reading. Note that a sufficiently prompted AI agent can definitely read this, so it's not meant to be cryptographically sound, more just unfriendly to the common AI reader!
Break search and screw over your disabled readers with this one weird trick! Legal in multiple countries
I love this! But won't the machine easily pick up on this?
It's actually not really easy for AI, without the agent doing some actual coding itself to reverse engineer the font file, or to take screenshots at different variable font intervals to zone in on the "focused" version of the variable font. All of that being said, the intention (beyond just having fun creating it) was to make it AI "unfriendly" so AI bots doing broad quick reads of it are going to be left with gobbledegook encoded characters.
Most llms can equally engage with text in picture form as text in token form. In fact my initial research on this (later corroborated by actual published papers) indicate that this is a cheap way to save on tokens.
Oh interesting and good to know on the token savings with this technique. My test with claude had it use vision and then programmatically test different variable font input variables (mimicking the user scrub interaction) until it was able to OCR it.
As they said in the comment you replied to: "Note that a sufficiently prompted AI agent can definitely read this, so it's not meant to be cryptographically sound, more just unfriendly to the common AI reader!"
Hilarious. Nice work
Thanks! Font against the machine! lol
lol
Fuck blind people I guess?
I don't know why you'd feel so hostile towards the blind, but you do you...
Hilarious that Claude was used to make it.
I truly lovely this as a conceptual exercise. However, I worry it will be easy for an agent to decompose. That said, well done.
neat idea! it is slightly amusing to find "btarbell and claude committed 2 weeks ago" in an anti-ai project.
turns out the master's tools will dismantle the master's house
That remains to be seen.