Today I’d like to tell the tale of how an innocent member of Anthropic technical staff summoned from the void a fictional 9,000-pound hippo named Gustav, and the chaos this hippo wrought.

Timeline:

June 2023
A member of Anthropic’s Product Research team creates a slide for a prompting tutorial, illustrating how to mitigate hallucinations by “giving Claude an out”. The slideshow is shared publicly.

slideshow

August 2023
I’m getting ready to deliver the slide to an audience, and decide to double-check that Gustav isn’t real. Indeed he is not. There’s no record of him on the Internet, or of any known hippo as heaviest of all time (HHOAT)

Thanksgiving 2023
Someone is looking at the slides and messages me like “Hey I think you should change this example, since there’s a real HHOAT so Claude’s IDK response isn’t ideal.” They send me a page from a website called americanoceans dot org as evidence.

americanoceansgustav

I look up snapshots from the webpage on archive.org since I KNOW I searched this just a few months ago. Sure enough, the page was newly created in fall 2023.

The webpage also contradicts itself lower down by stating that the Guinness World Record HHOAT is named Bertie (not Gustav or Humphrey). And you can see they have a lot of other sus-looking articles.

OK so we have a website churning out these fake articles probably with some sort of LLM assistance. How does this connect to Gustav and our slide show? Well, check out this Medium article that reskinned our slides, only they resampled and got Humphrey instead of Gustav.

BTW, if you search hippo Humphrey or hippo Bertie, you’ll find news stories about hippos of those names who are famous for other reasons. Meanwhile there’s a Wikipedia-famous giant crocodile named Gustav. This gives a clue about how LLMs get the wrong idea in the first place.

To recap, the story so far:

  • We make a slideshow explaining how to reduce hallucinations using heaviest-hippo as an example
  • Someone reposts it on a Medium blog
  • Someone else scrapes the blog and feeds it to an LLM, which writes up the example hallucination as a fact

From there, it spreads further. For a while (IDK if it’s still true), if you asked Alexa for the HHOAT it would cite americanoceans. And for a glorious few months, if you asked Google for the heaviest hippo of all time, it mentioned Gustav under yet another guise, “Hubert”, and cited our prompting docs as evidence. (By this time, we’d rerun the example again with a new model and got a different answer.)

As of today in May 2025, Google’s AI Overview doesn’t give a name for the HHOAT, but does claim there is one and they lived in Germany – just like the fictional Gustav, who lived in the Munich Zoo.

Is there light at the end of the tunnel? Am I making the problem worse by posting about it? I honestly have no idea, although I like to hope I’m raising awareness or whatever.

The latest Claude 4 models, in the contexts I’ve tried them, seem to not hallucinate hippos so severely. In Claude . ai, when I asked “who’s the heaviest hippo of all time? no searching” five times, Opus 4 gets it right each time. Sonnet 4 makes one up there but not in Console 🤷

Anyway, LLMs are weird and hard to predict, and so’s the Internet, the world, and the future. Be careful who you summon.