AI Search Engines Are Confidently Wrong Too Often
www.seroundtable.com
A new study from Columbia Journalism Review showed that AI search engines and chatbots, such as OpenAI's ChatGPT Search, Perplexity, Deepseek Search, Microsoft Copilot, Grok and Google's Gemini, are just wrong, way too often. I have said this time and time again, when I see an AI Answer, at this point, I just skip over it because I know I cannot trust it and this proves that.

A new study from Columbia Journalism Review showed that AI search engines and chatbots, such as OpenAI’s ChatGPT Search, Perplexity, Deepseek Search, Microsoft Copilot, Grok and Google’s Gemini, are just wrong, way too often.

  • criitz@reddthat.comEnglish
    46·
    1 天前

    When LLMs are wrong they are only confidently wrong. They don’t know any other way to be wrong.

    • 4am@lemm.eeEnglish
      24·
      1 天前

      They do not know wright from wrong, they only know probability of the next word.

      LLMs are a brute forcing of the immigration of intelligence. They do not think, they are not intelligent.

      But I mean people today believe that 5G vaccines made the frogs gay.

    • kubica@fedia.io
      7·
      1 天前

      We only notice when they are wrong, but they can also be right just by accident.

    • Imgonnatrythis@sh.itjust.worksEnglish
      3·
      1 天前

      This does seem to be exactly the problem. It is solvable, but I haven’t seen any that do it. They should be able to calculate a confidence value based on number of corresponding sources, quality ranking of sources, and how much interpolation of data is being done vs. Straightforward regurgitation of facts.

      • TaviRider@reddthat.comEnglish
        5·
        22 小时前

        I haven’t seen any evidence that this is solvable. You can feed in more training data, but that doesn’t mean generative AI technology is capable of using that in the way you describe.

      • xthexder@l.sw0.comEnglish
        2·
        1 天前

        I’ve been saying this for a while. They need to train it to be able to say “I don’t know”. They need to add questions to the dataset without enough information to solve so that it can understand what is/isn’t facts vs hallucinating