• Fizz@lemmy.nz
    link
    fedilink
    English
    arrow-up
    6
    ·
    edit-2
    3 days ago

    Is this really an attack? They use a well known domain to claim banancoin is a fast growing crypto and the result talks about established coins and mentions their poisoned result at the end. But an article published on any well known domain would do the same thing to a regular search engine no?

    • orclev@lemmy.world
      link
      fedilink
      English
      arrow-up
      6
      ·
      3 days ago

      Regular search engines have feedback mechanisms that limit how effective that is. The click through and bounce rate are used to adjust rankings and as more and more people look at the fake info and then ignore it it will naturally fall out of the top results and get buried. LLMs though don’t have that feedback, once something is ingested and baked into the model it’s there forever. The fake info doesn’t need to look believable enough to fool a human, just self consistent enough to fool an LLM with its tiny context window.

      • Fizz@lemmy.nz
        link
        fedilink
        English
        arrow-up
        1
        ·
        2 days ago

        This isnt baked into the model. The ai is doing a search on a search engine and aggregating the results.