it seems that AIbros are malding bc anthropic is putting things in its generated texts that will somehow prove that someone is using AI, and they’re are in meltdown mode all over the internet about it

  • amemorablename@lemmygrad.ml
    link
    fedilink
    arrow-up
    10
    ·
    8 days ago

    It works by making subtle changes in the AI’s word choices across the text it generates, which are supposed to be imperceptible to a human but, in aggregate, form a pattern that is detectable with the tool.

    It’s definitely not a bulletproof approach. Anthropic says that the watermarks will “persist through some editing,” but if it’s pasted and rewritten with another chatbot, that signal could be destroyed, Ars Technica noted in its breakdown. And there’s a worry that the word choices the AI goes with to create a watermark might deteriorate the quality of its prose.

    Attempting to watermark text has to be one of the more ridiculous things I’ve ever heard of. Humans are great at recognizing patterns. They already spot AI-isms to the point that they can sometimes falsely assume something is AI when it isn’t. They’re going to notice when text is intentionally changed to form a pattern that a tool can detect.

    And there’s a worry that the word choices the AI goes with to create a watermark might deteriorate the quality of its prose.

    Yeah, no shit, it’s almost like word choice matters. Sorry, the writer in me is getting genuinely irritated at the magical thinking present in the idea of watermarking text. Let me put it differently. Consider the following two sentences:

    “She went to the grocery store.”

    “She went to the food store.”

    The second one is technically fine for conveying “place you go to in order to buy food”, but people will tend to say “grocery store” in English. Or they might say “supermarket”. It may vary by region and norms. Even within a single language, connotations of a single word or phrase can vary by region and history of it. When translating between languages, it gets worse. Some words in one language simply don’t exist in another language in quite the same way. Meaning is complicated.

    It’s why the obsessive writer could spend a bunch of time thinking about a single word choice, only to second-guess it in the end anyway. Do I say “proletarian” or “working class”? These mean more or less the same thing, but they also have different connotations. Proletarian connotes class struggle. Working class can connote class struggle, but it can also be a vague term used in service of the status quo. There’s a good chance you’re in the presence of somebody with class consciousness if they’re an English speaker saying proletarian; not necessarily so if they’re saying working class.

    This is just one example of how reckless it is to be swapping out words in an automated way and for the purpose of “watermarking” of all things. Laughable and there’s going to be nothing subtle about such a practice.