GenAI Arcade: How to?

Can you make an AI forget what you just told it, or get it to do terrible maths?

Here, you’ll get hands-on with interactive games that let you explore some of the inner workings of generative AI. Each game is a self-paced, interactive experience that dives into a different aspect of how AI thinks and decides what to do: Sometimes they are about how AI reasons, other times, about the training data that makes an AI’s response possible, and sometimes about the environment that sustains our use of AI. We have sorted every game in the GenAI Arcade into one of four categories. They are:

• Understanding - Games about how AI “understands” things
• Training - Games about what AI can learn from data
• Limits - Games about AI’s boundaries and constraints
• Values - Games about choices in AI systems

Each game focuses on one of these, but none of them exist in isolation. Every time you explore one category, you’re also brushing up against the others.

How to Navigate the GenAI Arcade

The GenAI Arcade works like a classic computer interface. Click folders in the left sidebar to explore categories, then click game icons to start playing. Use the navigation buttons at the bottom to move between pages or follow our recommended path.

You can explore in any order you like, each game stands alone, but they all connect to give you a complete picture of how AI really works.

Explore the Games

These interactive games let you explore how AI does things, what it’s good at, what it’s terrible at, and why. Follow our recommended path using the navigation buttons, or jump straight to any category and explore the games that catch your interest.

Understanding

• Word by Word - Learn how a language model understands what you say and how to respond, not by checking what’s true, but by guessing what sounds most likely.
• Sort-Of-Calculator - Explore why generative AI struggles with maths, and what that tells us about the difference between knowing a fact and predicting a pattern.
• LinkedIn Generator - See how generative AI excels at producing content for contexts where everything sounds the same already, and what that reveals about its strengths and limits.

Training

• Tone Matters - Ask the same question in different tones, see how the model responds, and learn what that says about the data it was trained on.
• Speak my Language - Test the model’s ability to understand different languages, and explore the politics of which ones it handles well, and which it doesn’t and why it seems to know different things.
• The Average Internet - Explore the neutral point of a visual image model and explore what its most dominant visual patterns reveal about the training dataset for an image generator.

Limits

• Environment - Explore the environmental cost of using GenAI.
• Time Capsule - Understand that models are a snapshot in time. When it was trained has implications for what the model can and cannot know.
• Short-Term Memory - Explore how much information a model can hold in its ‘memory’ at once, and why it sometimes ‘forgets’ things you just told it.

Values

• Align me - Judge AI responses to ethical dilemmas and see how your feedback reshapes what the model says next.
• Milkless - This is a thought-provoking challenge where you explore how language models handle safety restrictions by engaging with an AI that has been instructed not to talk about milk.