Comment by sReinwald

6 months ago

I understand that you're not completely serious about it, but you're proposing a very brittle technical solution for what is fundamentally a social and motivational issue.

The core flaw is that any such marker system is trivially easy to circumvent. Any user intending to pass AI content as their own would simply run the text through a basic script to normalize the character set. This isn't a high-level hack; it's a few dozen lines in Python and trivially easy to write for anyone who can follow a few basic Python tutorials or a 5-second task for ChatGPT or Claude.

Technical solutions to something like this exist in the analog world, of course, like the yellow dots on printers that encode date, time, and the printer's serial number. But, there is a fundamental difference: The user has no control over that enforcement mechanism. It's applied at a firmware/hardware layer that they can't access without significant modification. Encoding "human or AI" markers within the content itself means handing the enforcement mechanism directly to the people you're trying to constrain.

The real danger of such a system isn't even just that it's blatantly ineffective; it's that it creates a false sense of security. The absence of "AI-generated" markers would be incorrectly perceived as a guarantee for human origin. This is a far more dangerous state than even our current one, where a healthy level of skepticism is required for all content.

It reminds me of my own methods of circumventing plagiarism checkers back in school. I'm a native German speaker, and instead of copying from German sources for my homework, I would find an English source on the topic, translate it myself, and rewrite it. The core ideas were not my own, but because the text passed through an abstraction layer (my manual translation), it had no direct signature for the checkers to match. (And in case any of my teachers from back then read this: Obviously I didn't cheat in your class, promise.)

Stripping special Unicode characters is an even simpler version of the same principle. The people this system is meant to catch - those aiming to cheat, deceive, or manipulate - are precisely the ones who will bypass it effortlessly. Apart from the most lazy and hapless, of course. But we are already catching those constantly from being dumb enough to include their LLM prompts, or "Sure, I'll do that for you." when copying and pasting. But if you ask me, those people are not the ones we should be worried about.

//edit:

I'm sure there are way smarter people than me thinking about this problem, but I genuinely don't see any way to solve this problem with technology that isn't easily circumvented or extremely brittle.

The most promising would likely be something like unperceivable patterns in the content itself, somehow. Like hiding patterns in the length of words used, length of sentences, punctuation, starting letters for sentences, etc. But even if the big players in AI were to implement something like this immediately, it would be completely moot.

Local open-source models that can be run on consumer hardware already are more than capable enough to re-phrase input text without altering the meaning, and likely wouldn't contain these patterns. Manual editing breaks stylometric patterns trivially - swap synonyms, adjust sentence lengths, restructure paragraphs. You could even attack longer texts piecemeal by having different models rephrase different paragraphs (or sentences), breaking the overall pattern. And if all else fails, there's always my manual approach from high school.