Google rolled out AI overviews across the United States this month, exposing its flagship product to the hallucinations of large language models.

Google rolled out AI overviews across the United States this month, exposing its flagship product to the hallucinations of large language models.

  • barsoap@lemm.eeEnglish
    5·
    6 months ago

    That’s not hallucinations (in particular), that’s concept bleed. Try the following:

    1. Acquire a human experimental subject. Ask them:
    2. What colour is snow?
    3. What colour is the fridge (point to a white fridge)?
    4. What do cows drink?

    …and hear them answer “milk”. “White, cold, drink, cow” are all wired to “milk” in our heads logic comes later. It’s quite a bit harder to trick humans with this than AIs because we do have the capacity to double-check but if you simply want to bend an answer, not have it be completely nonsensical, it’s quite easy.

    Also your 40k or Iron Maiden result might very well still be Battletech. E.g. when it comes to image composition. Another explanation would be low resolution in the prompt encoding, that’d be similar to boomers calling your PS5 a Nintendo. Most likely though it has only seen two or three Battletech images (face it, it’s not that popular in comparison) and thought “eh looks like a Nintendo that’s where I’ll store it”, Humans and current-gen AI are different in principle in that regard as we can come up with encoding strategies, they can’t. Something something T3 systems and need for exponential amounts of data.