Alpha Set

Every card in the Alpha Set, the first set of Tokens cards: AI concepts from the five AI4K12 big ideas, each with a plain-language definition and a question to talk about. Tap a card to read it up close.

Showing 70 of 70 cards

Not Alive. System, Humans and AI, Common, cost 1. Capability 1, Trust 3, 1 Token. Definition: AI systems can perform human-like tasks but aren't alive or conscious. Discuss: What's one thing a human can do that no card in this deck can?

The Human Trainer. System, Humans and AI, Common, cost 2. Capability 1, Trust 3, 2 Tokens. Ability: Whenever this System Runs a Task, you may put the top card of your deck into your Training Set. Definition: Humans label the examples an AI learns from. Discuss: If a labeler makes mistakes, what does the model learn?

Human-in-the-Loop. Tool, Humans and AI, Common, cost 2. Ability: {E} — Remove up to 2 Flags from target System. Definition: A design where a human reviews or overrides an automated decision before it takes effect. Discuss: Name a task where you'd always want a human to double-check the AI.

Chatbot. System, Humans and AI, Common, cost 1. Capability 1, Trust 2, 1 Token. Definition: A conversational agent that talks with people via text or voice. Discuss: How can you tell whether you're talking to a chatbot or a real person?

Agent. System, Humans and AI, Uncommon, cost 3. Capability 2, Trust 2, 2 Tokens. Ability: Fast-Tracked. Definition: An AI program that senses its environment, decides, and acts — physical (robot) or software. Discuss: What’s the difference between an "agent" and a simple chatbot?

AI Career Path. Action, Humans and AI, Uncommon, cost 1. Effect: Draw a card. Definition: People work on AI as data labelers, model trainers, ethicists, and more — not just "programmers.". Discuss: Which AI-related job sounds most interesting to you, and why?

Manual Override. Action, Humans and AI, Common, cost 2. Effect: Target opponent's System can't Run a Task next turn. Definition: A person steps in to pause an automated process before it can act again. Discuss: When should a person be allowed to override an AI's decision?

AI Isn't Magic. Action, Any Data Type, Chase, cost 2. Effect: Choose target System. It can't use abilities until its controller's next turn. Definition: AI isn't sentient or magical — it's a tool built by people. Discuss: What's something people assume AI can do that it actually can't?

Python. Tool, Humans and AI, Common, cost 1. Ability: {E} — Draw a card. Definition: A programming language popular for building AI models because it's readable and has huge community support. Discuss: Why might a beginner-friendly language end up being what experts use too?

Fast-Moving Lab. System, Humans and AI, Rare, cost 3. Capability 3, Trust 2, 2 Tokens. Ability: Fast-Tracked. Definition: An AI research company that races to release powerful new tools to the public quickly. Discuss: What's a tradeoff of a company releasing powerful AI tools quickly?

Safety Lab. System, Humans and AI, Rare, cost 3. Capability 2, Trust 4, 2 Tokens. Ability: Highly Trusted. Definition: An AI company that makes safety research a core part of how it builds AI systems. Discuss: Should a company that builds AI also be the one checking it's safe?

AI Assistant. System, Humans and AI, Uncommon, cost 3. Capability 2, Trust 3, 2 Tokens. Ability: Highly Trusted. Definition: A conversational AI built to help people with writing, questions, and analysis. Discuss: What would you want to double-check before trusting a chatbot's answer?

Viral App. System, Humans and AI, Common, cost 3. Capability 3, Trust 2, 2 Tokens. Ability: Fast-Tracked. Definition: An AI chatbot app that spread to millions of people faster than almost any app before it. Discuss: Why might a chatbot become popular faster than almost any app before it?

Built-in AI. System, Humans and AI, Common, cost 2. Capability 2, Trust 2, 2 Tokens. Definition: An AI assistant built right into search, email, and other apps people already use every day. Discuss: What's different about an AI built into apps you already use daily?

Defunct. Action, Humans and AI, Uncommon, cost 2. Effect: Discard target Tool. If it's a Legacy Tool, draw a card. Definition: A company, product, or service that has shut down and no longer exists. Discuss: What happens to your data when an app you use shuts down?

Abstraction. System, Rep. & Reasoning, Common, cost 1. Capability 1, Trust 2, 1 Token. Definition: A representation that keeps some features and leaves others out. Discuss: If you drew a map of your school, what would you leave out?

Feature Vector. Tool, Rep. & Reasoning, Common, cost 2. Legacy Tool. Ability: {E} — Look at the top 2 cards of your deck. Put one into your hand, the other on the bottom. Definition: A list of numbers encoding an object’s key traits so an algorithm can use them. Discuss: What 3 "features" would you use to describe your favorite animal to a computer?

Classifier. System, Rep. & Reasoning, Common, cost 2. Capability 2, Trust 2, 2 Tokens. Definition: A model that sorts input into categories, like spam vs. not-spam. Discuss: What's something in your life that gets automatically "classified" for you?

Predictor. System, Rep. & Reasoning, Common, cost 2. Capability 3, Trust 1, 2 Tokens. Definition: A model that outputs a continuous value, like estimating a house’s price. Discuss: What would you want an AI to predict for you, if it could?

Recommender. System, Rep. & Reasoning, Common, cost 2. Capability 3, Trust 1, 2 Tokens. Definition: A model that suggests items based on what similar users liked. Discuss: Has a recommendation ever felt "too accurate"? Why might that happen?

Decision Tree. System, Rep. & Reasoning, Uncommon, cost 3. Capability 2, Trust 4, 2 Tokens. Ability: Highly Trusted. Definition: A series of yes/no tests that leads to a decision — explainable step by step. Discuss: Sketch a 3-question decision tree for what to eat for lunch.

Search Tree. System, Rep. & Reasoning, Rare, cost 3. Capability 3, Trust 3, 3 Tokens. Ability: Whenever this System Runs a Task, look at the top 3 cards of your deck; put one into your hand. Definition: A branching map of possible moves an AI explores to find the best one, like in tic-tac-toe. Discuss: Why might exploring every possible move become impossible for a more complex game?

Ambiguous Input. Action, Rep. & Reasoning, Common, cost 1. Effect: Target opponent's System can't Run a Task next turn. Definition: Poorly represented or unclear data makes it hard for an AI system to act reliably. Discuss: Why might unclear instructions make an AI "freeze up" or give a bad answer?

Discontinued. Action, Rep. & Reasoning, Uncommon, cost 2. Effect: Discard target Tool. If it's a Legacy Tool, draw a card. Definition: A product or service that is no longer made or supported, even if some people still use it. Discuss: Why might a company warn people before it stops supporting a feature?

Sensor. System, Machine Learning, Common, cost 1. Capability 1, Trust 2, 1 Token. Definition: Hardware that turns a physical signal (light, sound) into data a computer can use. Discuss: What sensors does your phone have, and what does each one "sense"?

Training Data. Action, Machine Learning, Common, cost 2. Effect: Big Data 2 (or pay by turning a Cost 2+ System sideways). Put the top 2 cards of your deck into your Training Set face-down. Definition: The examples a model learns from before it's used. Discuss: Why would a model trained only on cat photos fail to recognize dogs?

Bias in Data. Action, Machine Learning, Common, cost 1. Effect: Target opponent's System gets +1 Flag. Definition: A systematic skew in a dataset that leads to unfair or inaccurate outputs. Discuss: Can you think of a dataset that might accidentally leave a group out?

Supervised Learning. System, Machine Learning, Common, cost 2. Capability 3, Trust 3, 2 Tokens. Definition: Training a model on labeled examples so it learns to match inputs to known answers. Discuss: Why does supervised learning need humans to label the answers first?

Neural Network. System, Machine Learning, Uncommon, cost 2. Capability 4, Trust 2, 2 Tokens. Definition: Many simple processing units organized to jointly solve a complex problem. Discuss: Neural networks are inspired by brains — how are they different from real ones?

Large Language Model. System, Machine Learning, Rare, cost 5. Capability 5, Trust 2, 4 Tokens. Ability: Fast-Tracked. Definition: A neural network trained on massive text data, used for chat, writing, and more. Discuss: Why might an LLM sound confident even when it's wrong?

Reinforcement Learning. System, Machine Learning, Uncommon, cost 4. Capability 3, Trust 2, 3 Tokens. Ability: Whenever this System is Deprecated, gain 1 Token. Definition: A model learns by trial and error, guided by a reward signal instead of labeled answers. Discuss: What's a game where trial-and-error is exactly how you'd learn to play well?

Retraining Pause. Action, Machine Learning, Common, cost 2. Effect: Target opponent's System can't Run a Task next turn. Definition: A model taken offline for retraining can’t be used again until the update finishes. Discuss: Why do AI companies periodically retrain or update their models?

Neuron. System, Machine Learning, Common, cost 1. Capability 1, Trust 2, 1 Token. Definition: The smallest processing unit in a neural network — many of them together form the whole model. Discuss: How is one neuron different from the whole Neural Network card?

Tensor. Tool, Machine Learning, Common, cost 1. Ability: {E} — Look at the top card of your deck. You may put it on the bottom. Definition: A multi-dimensional grid of numbers — the basic data structure nearly every AI model computes with. Discuss: Why might a model need numbers arranged in a grid instead of a list?

CUDA. Tool, Machine Learning, Common, cost 2. Ability: {E} — Target System gets +1 Capability until end of turn. Definition: A platform that lets AI models run on graphics chips (GPUs), training far faster than on a regular computer. Discuss: Why would training a huge model need special, faster hardware?

ImageNet. Action, Machine Learning, Uncommon, cost 2. Effect: Big Data 3 (or pay by turning a Cost 3+ System sideways). Put the top 3 cards of your deck into your Training Set face-down. Definition: A massive labeled image dataset that helped launch the modern boom in computer vision. Discuss: Why did one giant, carefully labeled dataset change what AI could do?

GAN. System, Machine Learning, Uncommon, cost 2. Capability 3, Trust 1, 2 Tokens. Ability: Adversarial. Definition: Two networks compete: one makes fakes, the other tries to catch them. Discuss: Why might two competing AIs both get better?

Diffusion Model. System, Machine Learning, Uncommon, cost 2. Capability 3, Trust 2, 2 Tokens. Ability: Whenever this System Runs a Task, remove 1 Flag from it. Definition: A model that generates images by starting with noise and gradually refining it into a picture. Discuss: How is starting from random noise different from drawing from scratch?

Unsupervised Learning. System, Machine Learning, Uncommon, cost 2. Capability 3, Trust 2, 2 Tokens. Ability: Whenever this System Runs a Task, look at the top 2 cards of your deck; you may rearrange them. Definition: Training a model to find patterns in data with no labels. Discuss: How could a computer sort things without being told the categories?

Transformer. System, Machine Learning, Rare, cost 4. Capability 4, Trust 2, 3 Tokens. Ability: Attention. Whenever this System Runs a Task, look at the top 2 cards of your deck; put one into your hand. Definition: A neural network design that weighs which parts of the input matter most — the architecture behind most modern LLMs. Discuss: Why might "paying attention" to the right words matter for a good answer?

GPT. System, Machine Learning, Uncommon, cost 3. Capability 3, Trust 2, 2 Tokens. Ability: Whenever this System Runs a Task, draw a card, then discard a card. Definition: An LLM pre-trained on huge amounts of text. Discuss: GPT means Generative Pre-trained Transformer. Why pre-train?

GPU. Tool, Machine Learning, Uncommon, cost 3. Ability: {E} — A System you deployed this turn can Run a Task or Run an Audit this turn. Definition: A graphics processing unit: a chip that does thousands of calculations at once, which makes training AI much faster. Discuss: Why would a chip built for video games end up powering AI?

CPU. Tool, Machine Learning, Common, cost 1. Legacy Tool. Ability: {E} — Draw a card, then discard a card. Definition: A computer's central processing unit: the general-purpose chip that runs a program's steps one after another. Discuss: Why might a general-purpose chip be slower for AI than a specialized one?

RAM. Tool, Machine Learning, Common, cost 2. Legacy Tool. Ability: {E} — Put an Action card from your discard pile on top of your deck. Definition: A computer's short-term memory, holding the data a program is working with right now. Discuss: What happens to what's in a computer's memory when the power goes off?

Embedding. System, Machine Learning, Common, cost 2. Capability 2, Trust 3, 2 Tokens. Definition: Turning words or images into numbers so similar things end up close together. Discuss: Which words would an AI place close to "dog"?

Fine-Tuning. System, Machine Learning, Uncommon, cost 3. Capability 2, Trust 3, 3 Tokens. Definition: Training an existing model a little more on a smaller set of examples to make it a specialist. Discuss: What would you fine-tune an AI to be really good at?

Outdated. Action, Machine Learning, Uncommon, cost 2. Effect: Discard target Tool. If it's a Legacy Tool, draw a card. Definition: A model trained on old data that no longer reflects how things are now. Discuss: Why might an AI trained years ago give wrong answers about today?

Fairness Check. Tool, Ethical AI Design, Common, cost 2. Ability: {E} — Put 1 Flag on target System whose Capability is higher than its Trust. Definition: Evaluating whether an AI system's decisions are fair to everyone affected. Discuss: Whose idea of "fair" should count when people disagree?

Bias Audit. System, Ethical AI Design, Common, cost 2. Capability 1, Trust 4, 1 Token. Ability: Highly Trusted. Definition: A deliberate review of a dataset or model to find and reduce unfair skew. Discuss: If you found bias in a model, what's the first thing you'd check?

Model Card. Tool, Ethical AI Design, Common, cost 1. Ability: Once per game, you may remove all Flags from a System you control instead of it being Deprecated. Definition: A document describing an AI model's training, performance, and limitations — like a nutrition label. Discuss: Why might a company be reluctant to publish a model card?

Explainability. System, Ethical AI Design, Common, cost 2. Capability 1, Trust 3, 2 Tokens. Definition: An AI's ability to justify its decision by pointing to the features that led to it. Discuss: Should self-driving cars be required to explain every decision? Why?

Privacy Shield. Tool, Ethical AI Design, Uncommon, cost 2. Ability: Systems you control can't be targeted by opponents' Action cards. Definition: Design practices that protect personal data used to train or run an AI system. Discuss: What personal data would you not want used to train an AI, even anonymized?

Accountability. System, Ethical AI Design, Uncommon, cost 2. Capability 2, Trust 4, 2 Tokens. Definition: Who is responsible when an AI system causes harm — the designer, the user, or the company. Discuss: If a self-driving car crashes, who should be held responsible?

Ethical Framework. System, Ethical AI Design, Rare, cost 4. Capability 2, Trust 5, 3 Tokens. Ability: Highly Trusted. Definition: A structured set of criteria (fairness, safety, transparency..) used to evaluate an AI system. Discuss: Pick one value (fairness, privacy, safety) — which matters most to you, and why?

Compliance Review. Action, Ethical AI Design, Common, cost 2. Effect: Target opponent's System can't Run a Task next turn. Definition: A formal check to confirm an AI system meets required standards before it keeps operating. Discuss: Should a company have to pause an AI system while it's being investigated for bias?

Mandatory Recall. Action, Ethical AI Design, Rare, cost 4. Effect: None of your opponent's Systems can Run a Task next turn (including ones they deploy that turn). Definition: When an AI system is found unsafe, regulators can require it be pulled from use until it’s fixed. Discuss: Should the government be able to shut down an unsafe AI product?

Informed Consent. System, Ethical AI Design, Uncommon, cost 3. Capability 2, Trust 3, 3 Tokens. Definition: Asking people for clear permission before their data is used to train or run an AI. Discuss: Would you let an app train AI on your photos? What would you need to know?

Red Teaming. System, Ethical AI Design, Uncommon, cost 3. Capability 2, Trust 4, 2 Tokens. Ability: Adversarial. Definition: Testers who try on purpose to make an AI fail, so problems get fixed before release. Discuss: If you were testing a chatbot, how would you try to trick it?

Expired. Action, Ethical AI Design, Uncommon, cost 2. Effect: Discard target Tool. If it's a Legacy Tool, draw a card. Definition: Permission or approval that has run out and must be renewed before it counts again. Discuss: Should permission to use your data run out after a while? Why?

AI in Daily Life. System, Societal Impacts, Common, cost 1. Capability 1, Trust 2, 2 Tokens. Definition: AI already shapes recommendations, searches, and filters in everyday products. Discuss: List 3 places AI touched your day today, before this game.

Deepfake. System, Societal Impacts, Common, cost 2. Capability 3, Trust 1, 1 Token. Definition: Synthetic media generated to convincingly show something that didn't happen. Discuss: How could you check whether a video is real before sharing it?

Job Disruption. Action, Societal Impacts, Common, cost 2. Effect: Target opponent's System can't Run a Task next turn. Definition: AI can automate some tasks, changing which jobs exist or what skills they require. Discuss: Name a job you think AI can't replace, and explain why.

Environmental Footprint. Action, Societal Impacts, Common, cost 1. Effect: Each player discards the top card of their Training Set. Definition: Training and running large AI models consumes significant energy and water. Discuss: Would knowing an AI's energy cost change how often you'd use it?

Data Privacy Trade-off. System, Societal Impacts, Common, cost 2. Capability 2, Trust 1, 2 Tokens. Definition: Using a "free" AI tool often means trading personal data for convenience. Discuss: What’s a "free" app you use that’s probably paid for with your data?

Regulation Debate. Action, Societal Impacts, Uncommon, cost 2. Effect: Choose one: each player may Train an extra card this turn; or target System can’t be Audited this turn. Definition: Ongoing public debate over what laws should govern how AI is built and used. Discuss: Should there be an age limit on using generative AI tools? Why or why not?

Digital Divide. System, Societal Impacts, Rare, cost 3. Capability 1, Trust 3, 3 Tokens. Ability: This System's Tokens can't be reduced by opponents' abilities. Definition: Unequal access to AI tools and AI education across different communities. Discuss: If AI tools cost money, who might get left behind?

Cognitive Offload. Action, Societal Impacts, Common, cost 1. Effect: Draw a card, then discard a card. Definition: Relying on AI to think, remember, or decide for you — convenient, but a skill you may stop practicing. Discuss: What's something you'd lose if you let AI always do it for you?

Data Center. System, Societal Impacts, Uncommon, cost 3. Capability 2, Trust 3, 3 Tokens. Ability: Whenever this System Runs a Task, discard the top card of your Training Set. Definition: A building full of computers that run AI. Discuss: Should towns welcome data centers that use lots of water?

Obsolete. Action, Societal Impacts, Uncommon, cost 2. Effect: Discard target Tool. If it's a Legacy Tool, draw a card. Definition: Technology that's been replaced by something newer and is no longer supported. Discuss: What's a gadget your family used to use that's now obsolete?

Alignment. System, Any Data Type, Chase, cost 3. Capability 2, Trust 5, 3 Tokens. Ability: Whenever this System Runs a Task, remove 1 Flag from another System you control. Definition: Making sure an AI system’s goals match what people actually want it to do. Discuss: If you built an AI helper, what rules would you give it first?