• poke@sh.itjust.works
    link
    fedilink
    arrow-up
    1
    ·
    1 year ago

    If you have a known algorithm for generating those hard-to-read images, then it really wouldn’t be that difficult to generate a large enough set yourself to train a custom ML model to solve them. The same would apply to audio challenges.

    Only one person would need to do it then they could share the process, potentially automating others being able to bypass as well.

    I like the idea of captcha being open, but unlike encryption as far as I know we don’t have a starting point on something that is actually easier for humans when all information is available. Until something like that exists, open sourcing to implement and improve it doesn’t make sense if you want an effective product.

    • BOB_DROP_TABLES@lemmy.ml
      link
      fedilink
      arrow-up
      1
      ·
      1 year ago

      The text is not generated. It’s from photos of books that failed ocr. The photos are then distorted to make it even harder in order to become that captcha. 2 words are used 1 is a control (to know if the response is correct), the other is one they what to know what says (to add to the pool of words and finish digitizing the book).