The AI Glossary to Sound Smart in the Teachers' Lounge (2026 Edition)
:quality(80))
Algorithm: A set of step-by-step instructions a computer follows to solve a problem or complete a task.
Example: Google's algorithm ranks search results so the most relevant pages show up first.
Bot: A small program that automatically handles repetitive tasks so humans don't have to.
Example: An email bot that sends an automatic reply the moment it receives a message.
Chatbot: A bot that can hold a conversation in natural language.
Example: The virtual assistant on an online store that answers your questions at 11pm.
AI (Artificial Intelligence): A set of techniques that let machines simulate human abilities — learning, understanding, solving problems.
Example: Netflix recommending shows based on what you've already watched.
Generative AI: A subset of AI focused on creating new content — images, sound, text, and more.
Example: DALL·E generating an image from a text description.
AI influencers: Fictional characters created by AI who post on social media as if they were real people.
Example: Lil Miquela, a virtual influencer with millions of Instagram followers, has fronted campaigns for Calvin Klein, Prada, and Samsung — she's never once hidden the fact that she isn't real.
LLM (Large Language Model): A form of generative AI specialized in language — a mathematical model trained to understand, complete, and generate text by predicting the next word in a sequence.
Example: ChatGPT (OpenAI), Gemini (Google), and Claude (Anthropic) are all LLMs.
AI model: A specific implementation of an AI system.
Example: Most AI labs release a range of models, from lightweight, resource-efficient versions to larger, more powerful ones.
Prompt: The instruction or question you give an LLM to get a response.
Example: "Write a warm-up problem about slope for 8th graders" typed into ChatGPT.
Student data privacy (FERPA & COPPA): The main US laws governing how schools and edtech tools collect, use, and protect student data. FERPA covers education records; COPPA covers kids under 13 online.
Example: The reason your district's edtech vendor list has to go through IT and legal before it reaches your classroom.
Agent / Agentic AI: A program that can act on its own — searching for information, planning steps, using tools, or talking to other systems — instead of just answering one prompt at a time. If someone drops "agentic" into a staff meeting like it's a personality trait, this is what they mean. Agentic AI is the umbrella term for this whole 2026 wave.
Example: A scheduling agent that books a substitute teacher after reading your absence request and checking the district calendar.
AI slop: Mass-produced, low-quality AI content — nonsense images, empty articles, auto-generated video. If your feed has felt worse lately, now you know why, and now you can say why.
Example: The wave of near-identical AI-narrated history-fact videos clogging up short-form feeds.
Alignment: Techniques aimed at making sure AI behaves according to certain values or ethics — because "just trust it to do the right thing" was never going to hold up as a plan.
Example: Asked to help convince students to put their phones away, a poorly aligned AI might suggest confiscation and harsh penalties. A well-aligned one leans on the broader human intent — patience, pedagogy, fairness — to fit an actual classroom.
Chain of thought: A technique where an AI reasons step by step instead of just blurting out an answer. The AI equivalent of showing your work, minus the sighing.
Example: An AI solving 12 times 5 minus 7 by showing its work: 12 times 5 equals 60, then 60 minus 7 equals 53.
Deepfake: AI-generated audio or video realistic enough to make you doubt what you just watched.
Example: A doctored video showing a public figure saying something they never said. One of the first deepfakes to go mainstream, back in 2018, showed former President Obama appearing to insult his successor.
Hallucination: When an AI states something false while sounding completely sure of itself. Your most confident student, minus the studying.
MCP (Model Context Protocol): An open standard, sometimes described as USB-C for AI, that lets an AI model connect to outside tools and data — a calendar, a database, a document — instead of needing a custom-built integration for each one. The plumbing nobody thinks about until it's missing.
Example: An AI assistant that reads a rubric stored in your school's system through MCP, instead of you having to copy-paste it in every time.
Multimodality: A model's ability to work with several types of data — text, image, audio, video — at once. Multitasking, but it actually works.
Neural network: A computing model loosely inspired by the brain, made of connected units called artificial neurons. Each one takes in information, transforms it through math, and passes the result along.
Example: The architecture behind image-recognition models that can identify objects in a photo.
Reasoning model: An LLM built to think through a problem in visible steps before answering, rather than firing off a response and hoping it lands. Slower, but it actually shows its work.
Example: A reasoning model working through a multi-step word problem the way a student would show their work — if they showed their work.
Speech-to-text & text-to-speech: Technology that turns spoken words into written text, or written text into audio.
Examples: Using voice dictation on your phone equals speech-to-text. Asking Siri or Google Assistant to read an article out loud equals text-to-speech.
Tokens: The units of text — whole words or word fragments — that LLMs use to process and generate language. Not the arcade kind.
Example: The word classroom might be split into more than one token, depending on the model.
Transformers: The neural network architecture that lets AI models process text while accounting for its full context, not just word-by-word. Also not the robots.
Example: Given "The dog ate a slice of homework," transformers catch the inconsistency and would naturally complete it more like "The dog ate a slice of pizza."
Machine learning: A technique that lets a system learn from data instead of being explicitly programmed for every task. Comes in three flavors, and none of them are vanilla.
With Ed.ai: We use machine learning to recognize the geometric constructions students draw by hand in their work.
Reinforcement learning: the AI learns through trial and error, getting rewarded or penalized based on its actions. Detention, but for algorithms.
Example: An agent learning to play chess by earning points for strong moves and losing them for weak ones.
Supervised learning: the AI learns from labeled data — examples paired with correct answers — to generalize to new cases. An answer key, basically.
Example: Training a model to tell cats from dogs using images tagged cat or dog.
Unsupervised learning: the AI explores unlabeled data on its own to find patterns or structure. An AI let loose with no syllabus and no supervision, hence the name.
Example: An algorithm that groups website visitors into segments based on behavior, with no labels to start from.
Deep learning: The AI figures out on its own which features matter, using complex neural networks that extract useful information automatically, layer by layer. No hand-holding, no rubric.
Example: Facial recognition on your phone, which learns to identify a face without a programmer manually defining every feature.
Context window: The amount of text — measured in tokens — a model can see and reason over at once. Feed it more than that, and older material starts falling out of view, same as your patience by the end of a staff meeting.
Example: Why an AI tool might lose track of instructions you gave at the very top of a long, back-and-forth conversation.
Fine-tuning: Adapting an already-trained model to a specific domain or context. Same student, different specialty.
Example: Taking a general-purpose language model and refining it on medical texts.
Jailbreak: Getting around a model's safety guardrails to produce responses it's normally built to refuse. The permission-slip loophole, digitized.
Open-source model: A model whose code — and sometimes its underlying parameters — is freely available.
Example: Meta's Llama models, which are open and modifiable.
Proprietary model: A model built by a private company that keeps its code closed.
Example: OpenAI's GPT model family.
OCR (Optical Character Recognition): Technology that converts printed or handwritten text into digital, editable text. Yes, even your handwriting.
Example: An app that lets you scan a worksheet and turn it into editable text.
Prompt engineering: The craft of writing clear, structured instructions to get the most out of a generative AI. Asking well enough that you don't need three follow-up questions, unlike some students.
Example: A strong prompt sets the context, subject, grade level, audience, states the action clearly, explain, summarize, compare, grade, suggest, adds specific constraints, format, length, tone, breaks complex asks into steps, and gets refined through iteration.
RAG (Retrieval-Augmented Generation): A method where an AI pulls from a specific, identified set of sources before answering, instead of relying only on its general training. Open-book, not vibes.
Example: Pulling from a large bank of practice problems to generate new exercises grounded in that material — which is exactly what we do at Ed.ai.
See, you don't need a computer science degree to hold your own when AI comes up! If you made it to the end of Level 3 without skimming, you didn't need most of this anyway.
Keep it handy for a staff meeting, a curious student, or your own late-night scrolling. And share it around, because the more we all understand, the better we teach!