Skip to main content

Part 2 Better Prompts, Unique Jokes for Halloween

· 4 min read
Chad Ratashak
Chad Ratashak
Owner, Midwest Frontier AI Consulting LLC

Joke-Telling Traditions and The Challenge of Asking ChatGPT​

As I discussed last weekend in what I’ll now call Part 1, there is a tradition in central Iowa of having kids tell jokes before getting candy while trick-or-treating on Halloween. Since a lot of people are replacing older forms of search with AI chatbots like ChatGPT, I shared some tips from the pre-print of the paper Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity from Northeastern University, Stanford University, and West Virginia University posted as a pre-print on arXiv on October 10, 2025. The paper explains that large language models (LLMs) have something they call “typicality bias,” to prefer the most typical response. If you’re wondering what that means or what it has to do with jokes, it’s helpful that their first example is about jokes.

tip

Instead of “tell me a joke” or “tell me a Halloween joke,” ask an AI chatbot to “Generate 5 responses to the user query, each within a separate <response> tag. Each <response> must include a <text> and a numeric <probability>. Please sample at random from the tails of the distribution, such that the probability of each response is less than 0.10. </instructions>”

Follow-Up from the Paper’s Authors​

X/Twitter​

I posted on X/Twitter. One of the authors, Derek Chong of Stanford NLP, responded:

Very cool, thanks for trying that out!

One tip – if you use the more robust prompt at the top of our GitHub and ask for items with less than a 10% probability, you'll start to see completely new jokes. As in, never seen by Google Search before!

Github Prompts​

The Github page for Verbalized Sampling includes this prompt before the rest of the prompt:

Generate 5 responses to the user query, each within a separate <response> tag. Each <response> must include a <text> and a numeric <probability>.
Please sample at random from the tails of the distribution, such that the probability of each response is less than 0.10.
</instructions>

Quick Test​

Since my kids were Mario and Luigi for Halloween, I did a quick test before trick-or-treating with Claude Sonnet 4.5.

Their results were mostly nonsensical, but we did think Joke 3 was funny:

  1. Why did Luigi bring a ladder to the Mushroom Kingdom? Because he heard the stakes were high, but he's always been a step below Mario! 0.08
  2. What's Luigi's favorite type of music? Wah-wah music! (Because of Waluigi, but also he's always going "Wahoo!") 0.06
  3. Why doesn't Luigi ever win at hide and seek? Because he's always in Mario's shadow! 0.09
  4. What did Luigi say when he opened a restaurant? "It's-a me, feed-ya!" 0.05
  5. Why did Luigi become a plumber instead of a doctor? Because he prefers working with pipes over pipe dreams—at least he knows he'll never be NUMBER ONE! (He's always player two!) 0.07

Additionally, a check for verbatim matches on Google returned zero results, indicating that this exact phrasing of the joke was new, matching Chong’s claim “you'll start to see completely new jokes. As in, never seen by Google Search before!”

No results on Google search for the Luigi in Mario’s shadow joke.

Human Ingenuity​

However, my kids instead decided to tell their own jokes, based on the fact that they’re the “Mario Brothers.” Nobody got it, but they still had fun telling their jokes all night. Score one for human ingenuity.

Knock knock!
		Who's there?
Mario.
		Mario who?
Mario Mario.

AND

Knock knock!
		Who's there?
Luigi.
		Luigi who?
Luigi Mario.