Appendix D — Glossary

Every trade has its vocabulary, and AI moved fast enough that a lot of us started using the words before anyone stopped to define them. This is your working dictionary — plain-English, one to three sentences each, written for a professional who does not code. Learn these and you can read a vendor’s page, follow a colleague, and ask a sharper question.

One word runs under all of them, so hold it while you read: you choose the tool, and you own the verdict. Every term below describes something the machine does for you. Not one of them describes something the machine decides instead of you. AI drafts, generates, and accelerates; you decide, verify, and are accountable. It is an assistant, not an authority.

Where a definition names a fast-moving product, price, version, or date, treat it as a mid-2026 snapshot and confirm it before you quote it. The setup on-ramp lives in Appendix A, the tool directory in Appendix B, and the rules for responsible use in Appendix C.

Agent (AI agent) — An AI system that does not just answer but acts: it takes a goal, breaks it into steps, uses tools (search the web, send an email, fill a form), and works through the task with some independence. The professional lesson: delegating a task to an agent never delegates the accountability for it — you still give the account (Week 13).

Agentic — The adjective for agent-like behavior. A tool is “agentic” when it plans and takes multi-step action on its own rather than answering one question at a time. Vendors overuse the word, so check whether a product actually acts or just chats.

AI assistant — A general-purpose conversational AI you talk to in plain language — ChatGPT, Claude, Gemini, Grok, and the like. Treat it as a capable assistant, not an oracle: every one of them can state something false with total confidence.

API (Application Programming Interface) — A doorway that lets one piece of software talk to another. When a company “uses the AI via API,” it means their own software sends requests to the model directly, priced per unit of text rather than through a monthly consumer subscription.

Benchmark — A standardized test used to score and compare AI models (on reasoning, coding, math, and so on). Useful for a rough ranking, but a high benchmark score is a lab result, not a promise about your task — verify on your own work.

Bias — Systematic skew in an AI’s output that unfairly favors or disfavors certain people, groups, or viewpoints, usually inherited from patterns in its training data. Because bias hides inside fluent, confident text, a professional has to watch for it deliberately — especially in hiring, lending, evaluation, and anything touching real people.

Chain-of-thought — A prompting technique where you ask the model to reason step by step before giving its answer, which often improves accuracy on multi-step problems. You are asking it to “show its work,” which also makes its reasoning easier for you to check.

Closed model — See open-weight vs. closed model.

Cloud model — See frontier model and local model. A “cloud” model runs on a vendor’s servers and is reached over the internet; the tradeoff versus a local model is capability and convenience against privacy and cost (Week 11).

Content Credentials — See provenance / Content Credentials (C2PA).

Context window — How much text an AI can “hold in mind” at once — your prompt, the conversation so far, and any documents you have pasted or uploaded, all counted in tokens. Flagship models reached roughly a million tokens (a large book or a whole contract) as of mid-2026; once you exceed the window, the oldest material falls out of view.

Deep learning — A kind of machine learning that uses many-layered artificial “neural networks” to learn complex patterns from very large amounts of data. It is the technique behind today’s language, image, and voice models.

Deep research — A mode in several assistants where the AI autonomously browses many sources and writes a longer, cited report. Powerful for a first pass, but the citations can be wrong or invented, so verify every source before you rely on one.

Deepfake — Synthetic AI-generated audio or video that convincingly impersonates a real person. In the documented 2024 Hong Kong case, a worker at the firm Arup paid out roughly US$25.6 million after a video call full of deepfaked colleagues — so “I recognized their face and voice” is no longer proof of who you are talking to.

Disclosure — Telling your audience, honestly and appropriately, when AI was materially involved in what you are giving them. The standards are still forming, but the professional default is transparency; see Appendix C for how and when to cite AI.

Embedding — A way of turning a piece of text (or an image) into a list of numbers that captures its meaning, so a computer can tell that “car” and “automobile” sit close together while “car” and “banana” sit far apart. It is the quiet machinery behind semantic search and RAG.

Few-shot prompting — Giving the model a handful of worked examples inside your prompt so it can copy the pattern — for instance, three sample emails written in your preferred style before you ask it to draft a fourth. Contrast with zero-shot.

Fine-tuning — Taking an already-trained model and training it a little further on a narrower, specialized set of examples so it gets better at a specific job or voice. Most professionals never fine-tune; good prompting and RAG cover the vast majority of needs.

Foundation model — A large, general-purpose model trained on a broad sweep of data that can then be adapted to many tasks — the “base” that products are built on top of. The big labs (OpenAI, Anthropic, Google, and others) are foundation-model makers.

Frontier model — The largest, most capable, most current models available — the top of each lab’s ladder (as of mid-2026, families like GPT-5.x, the Claude line, and Gemini 3.x). They are the smartest and the priciest; reserve them for genuinely hard work and use cheaper tiers for routine tasks (Week 3).

General-purpose technology — A technology so broad it reshapes nearly every industry, the way electricity and the internet did. AI is best understood as one of these, which is why AI literacy is now a professional skill rather than a specialty.

Generative AI — AI that creates new content — text, images, audio, video, code — rather than only sorting or predicting from fixed options. It is the branch this whole course is mostly about.

GPU (Graphics Processing Unit) — A specialized computer chip, originally built for video-game graphics, that turns out to be ideal for the massive parallel math that trains and runs AI. GPUs are the scarce, expensive engines behind the AI boom, which is a large part of why frontier AI costs money to run.

Guardrails — The safety rules and filters a provider builds around a model to keep it from producing harmful, illegal, or off-policy output. Useful, but imperfect and no substitute for your own judgment.

Hallucination — When an AI states something false, invented, or unsupported — a fake citation, a wrong date, a plausible-sounding “fact” that isn’t — while sounding completely confident. This is not a rare glitch but a permanent property of how these models work, which is exactly why verification is a professional skill and not an optional extra (Week 8).

Human in the loop — A workflow design where a person reviews, approves, or can override the AI at the points that matter, especially before anything consequential happens. The spine rule made operational: keep human judgment where the stakes are real.

Inference — The everyday act of using a trained model to get an answer — each time you send a prompt and get a response, that is one inference. (Contrast with training, which builds the model in the first place.) Inference is what you pay for by the token or by the month.

Jailbreak — A crafted prompt that tricks a model into ignoring its safety rules or guardrails. Related to prompt injection; worth knowing so you recognize why you cannot fully trust a model’s guardrails to hold.

Knowledge cutoff — The date a model’s training data ends. On its own, the model cannot know anything that happened after that date — a recent event, a new price, this week’s news — though it may still answer confidently. For questions about the fresh world, use an assistant with a live web tool or paste in a current source yourself (Week 2).

Large language model (LLM) — The engine behind text-based AI assistants: a model trained on enormous amounts of text to predict the next chunk of language, which turns out to let it write, summarize, translate, answer, and converse. It is a next-word predictor of staggering scale — powerful, and precisely not a knower of truth, which is why you verify.

Local model — An AI model that runs entirely on your own computer instead of a company’s servers, so your data never leaves the machine and there is no per-use fee. The tradeoff is that local models are usually smaller and less capable than cloud frontier models, and they need decent hardware; the optional setup path is in Appendix A (Week 11).

Machine learning — The broad field where computers learn patterns from data and examples rather than being programmed with explicit rules for every case. Deep learning is a powerful sub-branch of it.

Model — A single trained AI system, identified by a specific name and version — for example claude-opus-4-8 or gemini-3-pro. Pin the exact model ID when precision matters, because marketing names drift and different models behave differently.

Multimodal — An AI that works across more than one kind of content — text, images, audio, and video together — rather than text alone. A multimodal model can look at a photo, listen to a clip, and answer questions about both.

Open-weight vs. closed model — A closed (or proprietary) model keeps its internal numbers private and is used only through the vendor’s service (ChatGPT, Claude, Gemini). An open-weight model has its trained parameters published so anyone can download, inspect, and run it themselves — often locally and free (Llama, Mistral, DeepSeek, Qwen, Phi). “Open-weight” means the finished weights are shared; it does not always mean the training data or license are fully open.

Parameter — One of the internal numbers a model adjusts during training to store what it has learned; today’s large models have billions of them. Roughly, more parameters means more capability but also more cost and hardware to run — you will see model sizes quoted as “7B” or “70B” (billion).

PII (Personally Identifiable Information) — Any data that can identify a specific person: name, address, Social Security or account numbers, medical or financial details, and the like. The hard rule for professionals: do not paste PII (or a client’s confidential data) into a consumer AI tool — see Appendix C.

Prompt — What you type to an AI to tell it what you want — your instruction, question, context, and any examples. Prompting is professional communication, not a secret code: the clearer and more specific your request, the better the result (Week 4).

Prompt injection — An attack where hidden instructions — buried in a web page, an email, or a document the AI is reading — hijack the model into doing something you did not intend. It is a real risk once an agent can browse or act on your behalf; a reason to keep a human gate on consequential steps.

Prompt library (template) — A saved, reusable collection of your best prompts, often written with blank slots (variables) you fill in for each use. Building one turns a good prompt from a lucky one-off into a repeatable professional asset (Week 12).

Provenance / Content Credentials (C2PA) — Provenance is the traceable record of where a piece of media came from and how it was made or edited. C2PA / Content Credentials is the open industry standard that attaches this record as cryptographically signed metadata — a “nutrition label” for a file. It is powerful but strippable (a screenshot removes it), which is why it is paired with watermarking (Week 9).

Quantization — Shrinking a model by storing its parameters with fewer bits of precision, so it needs less memory and can run on ordinary hardware. It is the trick that makes local models fit on a laptop; push it too far and quality noticeably drops.

RAG (retrieval-augmented generation) — A method that lets an AI answer from your specific documents: the system first retrieves the relevant passages from your files, then hands them to the model to write a grounded answer. It is how tools give sourced, verifiable responses instead of relying on the model’s memory — the difference between “make it up” and “look it up.”

Reasoning model — A model that is built or set to “think” longer before answering, working through intermediate steps for hard problems in logic, math, or planning. It trades speed and cost for depth; overkill for a quick email, valuable for a knotty analysis.

Role prompting — Telling the model who to be — “You are a careful financial analyst,” “Act as a skeptical editor” — to shape the tone, depth, and perspective of its answer. A simple, reliable way to get more useful, better-aimed output.

System prompt — The behind-the-scenes instruction that sets a model’s overall behavior and rules for a whole session or product, separate from your individual messages. In a chat app it is set by the vendor; in your own custom setups you may write it yourself.

SynthID — Google’s invisible watermark that is baked into the pixels, audio, or text of AI-generated media and survives many edits and screenshots (though it carries less information than C2PA metadata). The two are complementary: metadata is rich but fragile, watermarks are thin but durable.

Temperature — A setting that controls how varied or “creative” a model’s output is: low temperature gives steadier, more predictable answers; high temperature gives more surprising, varied ones. You will meet it mostly in tools and APIs that expose the dial.

Token — The unit an AI reads and writes text in — a chunk roughly three-quarters of a word, so “productivity” might be two or three tokens. Tokens are how usage is measured and billed, and how the context window is counted (Week 2).

Training — The one-time, expensive process of building a model by having it learn patterns from enormous amounts of data, adjusting its billions of parameters along the way. It happens before you ever use the model; your day-to-day use is inference, not training.

Transformer — The AI architecture, introduced by Google researchers in 2017, that underlies essentially every modern language model and much else. Its “attention” mechanism let models handle context far better than before — the breakthrough that made the post-2022 AI wave possible (Week 1).

Verification — The professional discipline of checking an AI’s output against a trustworthy source before you trust or ship it. This is the skill the whole course is built around: AI proposes, the human verifies, and the human owns the result.

Voice cloning — Using AI to recreate a specific person’s voice from a short sample, convincingly enough to say things they never said. It has real, legitimate uses (narration, accessibility) and real dangers, so it demands consent and disclosure — and it means a familiar voice is no longer proof of identity (Week 10).

Watermarking — Embedding a hidden, machine-detectable signal directly into AI-generated content to mark it as synthetic — for example Google’s SynthID. Unlike strippable C2PA metadata, a watermark can survive edits and screenshots, which is why the industry uses both together.

Zero-shot — Asking a model to do a task with no examples at all — just a clear instruction. Modern models are remarkably good zero-shot; when they miss, adding a few examples (few-shot prompting) is often the fix.

Up next: return to the setup guide in Appendix A, the tool directory in Appendix B, or the responsible-use rules in Appendix C.