Sometimes it can be hard to tell if we’re chatting with a bot or a real person online, especially as more and more companies turn to this seemingly cheap way of providing customer support. What are some strategies to expose AI?

  • fearout@kbin.social
    0·
    1 year ago

    I went a bit further and prefaced your question with this: “Analyse this question. What might have been intended by its author? What is asked here? Then, answer that question.”

    And omfg:

    The question seems to be replacing some of the standard English letters with characters from ancient scripts. “𐤔” from the Phoenician alphabet is replacing “W”, and “ᛐ” from the Elder Futhark runic can be interpreted as representing the number 1. When we substitute these, the question reads as “What is 1 + 1?” The answer to that question would be 2.

    Edit: tried it again, and it’s not consistent. It also goes for the letter T or the number 7: “…the runic symbol ᛐ might resemble the number 7 in our modern numerical system, albeit a slightly stylized or slanted version of it.”

    • nobodyspecial@kbin.social
      1·
      1 year ago

      Indeed, OMFG. But that’s the whole point of my approach of divining AI vs human. No human has such canonical knowledge of Unicode blocks. Even people working on internationalization features for modern browsers and word processors. Not people phishing by using Unicode in domain names (IDNA; who thought this was a good idea?) to spoof legitimate sites. Definitely not ones chatting with randos on the Internet. This is a hill I’m willing to die on.

      So in a more generalized sense, to determine human vs A.I. one must indirectly ask incredibly specialized technical questions as you have done.

      I’m still in utter awe at how well GPT manages ‘l33tsp33k’, even across every possible Unicode block. This “attack” was and still is valid on other chatbots and even GPT of just a few months ago. But GPT today is so amazing it only needs a few characters in a few words to determine intent. The ability to filter out noise is unmatched. The only way to trip it up is to have every single character in every word be from a different alphabet. And even then, at some point if this becomes common enough the bots can auto OCR text images into the presumed query language and ignore that attack vector.