AI Resources
AI Glossary
Plain-language definitions of the AI terms you’ll hear in vendor pitches, the news, and our workshops.
Jump to: Kinds of AI · Words for using AI · How a machine learns · Microsoft Copilot · Responsible use
Kinds of AI
- Artificial intelligence (AI): Software doing thinking-type work, like reading, sorting, predicting, and writing. It’s the umbrella for every term below.
- Machine learning: Software that learns patterns from examples instead of following rules someone typed in. Example: bookkeeping software that starts filing an expense in the right category after you’ve filed it that way a few times.
- Deep learning: Machine learning with many stacked layers, used for harder problems. Example: your phone unlocking when it recognizes your face.
- Generative AI: AI that creates something new, such as text, images, or slides, from a prompt. Example: a first draft of a customer email.
- Large language model (LLM): The engine inside ChatGPT, Microsoft Copilot, Google Gemini, and Claude. It predicts language one small piece at a time, which is why it writes fluently and can still be wrong.
- Agentic AI (AI agents): AI that takes actions across your tools instead of only writing text. Example: reading an email, updating a schedule, and drafting a reply. Because it acts, decide in advance what it may do without a person’s approval.
- Automation: Repeatable steps a system runs without you. It may include AI, but often it’s just rules. Example: an automatic appointment reminder.
- Multimodal AI: AI that works with more than text, such as images, audio, documents, or video. Example: asking a chat tool to read a photo of a whiteboard.
Words for using AI
- Prompt: What you type: your instructions, your context, and any material you paste in. It’s the part you fully control. See the Prompt Library.
- RACE: Role, Action, Context, Expectation. Our four-part framework for writing a prompt: who the AI should think like, the task, what else it should know, and what the output should look like.
- Training data: Everything a model learned from before you used it. It doesn’t include your organization’s files, people, or policies unless you provide them.
- Knowledge cutoff: The date a model’s training data ends. It knows nothing after that date unless the tool looks it up.
- Hallucination: A confident, fluent, wrong answer, like a citation, code section, or number that doesn’t exist. It comes from how these tools work, so check facts before you use them.
- Token: The unit a model reads and bills in, roughly three-quarters of a word. A one-page report is about 500 tokens.
- Context window: How much a model can hold in view at once, measured in tokens. If something isn’t in the window, the model doesn’t know it.
- RAG (retrieval-augmented generation): The tool looks up relevant documents first and puts them in the context window before it answers, so it works from real sources instead of memory.
- Grounding: Tying an answer to specific sources, either ones you paste in or ones the tool retrieves. Grounded answers make fewer things up, but grounded isn’t the same as verified.
- Sycophancy: A model’s tendency to agree with you, because these tools are tuned toward answers people rate highly. When a decision matters, ask it to argue the other side.
- Iteration: Improving a result with follow-up prompts, like “make it shorter” or “check this against my notes.” The second prompt is often where the work gets good.
How a machine learns
Three ways, with the nicknames we use in our workshops.
The Teacher
Supervised learning. Learns from examples with the right answers attached, like flashcards.
Example: a spam filter learning from the messages you mark as spam.
The Explorer
Unsupervised learning. Finds groupings in data when no answers are given.
Example: a streaming service grouping viewers with similar tastes.
The Gamer
Reinforcement learning. Learns by trial and error, with rewards for good moves and penalties for bad ones.
Example: game-playing AI that got good at chess by playing itself millions of times.
Chat tools are built with a mix of these, then trained on huge amounts of text. They learned what an expert sounds like, not how to be one. That’s why they can sound sure and still be wrong. Try the Teacher yourself with Teachable Machine.
Microsoft Copilot, sorted out
Copilot Chat
Microsoft’s AI chat assistant, at copilot.microsoft.com or in the Copilot app. Sign in with your work or school account to use it under your organization’s agreement with Microsoft, then paste in the material you want it to work from.
Microsoft 365 Copilot
Microsoft’s paid business license. It adds Copilot inside Word, Excel, PowerPoint, Outlook, and Teams, and it can search your organization’s email, files, and meetings. That search step is RAG. What you see inside the apps depends on your organization’s license.
Tip: in Excel, Copilot writes a formula and the spreadsheet does the math. In a chat window, the model predicts the math, so ask it to show its work and check one row by hand.
Faulkner students, faculty, and staff: sign in to Copilot with your Faulkner account. Faulkner doesn’t license the paid Microsoft 365 Copilot add-on, so use Copilot Chat in your browser or the Copilot app.
Responsible use
- The fit test: Three questions before you hand a task to AI. What goes in? Who sees the output? Can a mistake be undone? Your answers put the task in green (draft and go), amber (a second set of eyes), or red (a qualified person decides).
- Human review: A person checks facts, numbers, tone, and fit before AI output is used or sent. Ask: would you put your name on it exactly as written?
- Personal information (PII): Information that identifies a person, such as a name with an ID number, account details, grades, or health and financial records. Keep it out of AI tools your organization hasn’t approved.
- NIST AI Risk Management Framework (AI RMF): A free, voluntary framework from the U.S. National Institute of Standards and Technology for managing AI risk. It’s organized into four functions: Govern, Map, Measure, and Manage. Read it at nist.gov.