AI at Work Jeopardy

How AI tools work, good prompting, hallucinations, and what never to paste into a chatbot. A free board your team can copy and play with phone buzzers.

6 categories · 30 clues · Final Jeopardy · everyone buzzes in from their phone

For work

Copies all 30 clues into an editor of your own. Free, no account needed.

Every clue on this board

This page is public, so anyone can read the answers — copy the board and swap a few clues before game day.

How Language Models Work

$200
Language models read and write text in these small chunks, which can be whole words or pieces of words.
What are tokens?
$400
Rather than looking facts up, a language model writes by repeatedly predicting this.
What is the next word?
$600
The maximum amount of text a model can take into account at once is called this.
What is the context window?
$800
This setting controls how random a model's output is, with lower values giving more predictable text.
What is temperature?
$1,000
The 2017 paper Attention Is All You Need introduced this neural network architecture.
What is the transformer?

Writing Good Prompts

$200
Telling the model to act as an experienced editor is assigning it one of these.
What is a role?
$400
Including a few examples of what you want in the prompt is this kind of prompting.
What is few-shot prompting?
$600
Telling the model exactly how you want the answer laid out, such as a table or bullet list, specifies this.
What is the output format?
$800
Asking a model to think step by step is associated with this prompting technique.
What is chain-of-thought?
$1,000
Splitting a complex task into a sequence of smaller prompts, each feeding the next, is called this.
What is prompt chaining?

Hallucinations

$200
Before relying on an AI-generated statistic, you should do this against an original source.
What is fact-checking?
$400
In a 2023 New York case, lawyers were sanctioned for filing a brief with made-up examples of these produced by a chatbot.
What are case citations?
$600
A model may invent details about events after this date, where its training data ends.
What is the knowledge cutoff?
$800
This technique reduces made-up answers by having the model fetch relevant documents and answer from them.
What is retrieval-augmented generation?
$1,000
Keeping a person responsible for reviewing AI output before it is used is known by this phrase.
What is human in the loop?

Protecting Company Data

$200
Some public AI chatbots may use what you type to improve future models, so never paste in this kind of information.
What is confidential information?
$400
Before using an AI tool for work, check this company document setting out approved and banned uses.
What is an acceptable use policy?
$600
This term describes employees using AI tools that IT has not approved or does not know about.
What is shadow AI?
$800
Names, home addresses and Social Security numbers are examples of this kind of information.
What is personally identifiable information?
$1,000
This attack hides malicious instructions in content a model reads, such as a webpage or email.
What is prompt injection?

AI Vocabulary

$200
AI that creates new text, images or audio is called this.
What is generative AI?
$400
This term describes an AI system that can take actions and complete multi-step tasks on its own, not just chat.
What is an AI agent?
$600
Models that can handle text, images and audio together are described with this word.
What is multimodal?
$800
Further training a pretrained model on a smaller, specialized dataset is called this.
What is fine-tuning?
$1,000
These lists of numbers represent the meaning of words or documents and power semantic search.
What are embeddings?

AI History

$200
This British mathematician proposed the imitation game in a 1950 paper.
Who was Alan Turing?
$400
This IBM computer defeated world chess champion Garry Kasparov in a 1997 match.
What is Deep Blue?
$600
This New Hampshire college hosted the 1956 summer workshop often called the founding event of AI as a field.
What is Dartmouth College?
$800
Joseph Weizenbaum's 1960s MIT chatbot, which imitated a psychotherapist, was named this.
What is ELIZA?
$1,000
This DeepMind program defeated Go champion Lee Sedol in 2016.
What is AlphaGo?

Final Jeopardy — AI Vocabulary

This term describes a generative AI model confidently producing false or fabricated information, such as a citation to a source that does not exist.

What is a hallucination?

How to run this board

Generative AI tools have arrived in most workplaces faster than the training has. This board covers the basics of how large language models work, writing a useful prompt, spotting and checking hallucinations, keeping confidential data out of the wrong tools, AI vocabulary, and a short history of the field.

Treat it as an introduction or refresher that sits alongside your organization's AI use policy, not a substitute for it. Edit the data-handling clues to name the tools your company has approved and what may and may not go into them.

A useful house rule for the hallucination column: any team that answers has to say how they would verify the output. That habit is the most valuable thing the game can leave behind.

More from the library

Spotted a clue that's wrong or out of date? Tell us and we'll fix the board. Or start from a blank board and write your own.