AI Terms Explained: A Plain-English Glossary for 2026

So you nodded along in a meeting while someone said "we'll just RAG it against the embeddings and keep the temperature low". Everyone else nodded too. Statistically, most of them were bluffing as well.

AI jargon moves faster than anyone's ability to learn it, and almost every explainer out there is either a maths lecture or marketing soup. This is the plain-English version: the terms you'll actually meet, what they mean, and why you should care. Every term here also lives in the Tickd AI glossary, which is searchable, filterable and free.

Start with the four that explain everything else

Token. Models don't read words, they read tokens — chunks of roughly four characters. "Unbelievable" might be three tokens. This matters because pricing, speed and limits are all counted in tokens, so a chatty prompt genuinely costs more than a tight one.

Context window. How much the model can hold in mind at once: your prompt, the documents you pasted, and its own reply. Go past the edge and the oldest material silently falls off. Most "the AI forgot what I told it" complaints are a context window problem, not a memory problem.

Inference. Training builds the brain; inference is asking it a question. When someone says "inference costs", they mean the cost of every answer, forever — not the one-off cost of building the model.

Hallucination. A confidently wrong answer. The model isn't lying, because it has no concept of truth — it predicts plausible text. There is no fact-checker inside unless you build one.

Prompting terms worth knowing

System prompt — the standing instructions the model reads before your message: persona, rules, boundaries. Most of the difference between a good AI product and a bad one lives here.

Few-shot prompting — showing two or three worked examples so the model copies your pattern instead of inventing its own. It is the single highest-return trick in prompting.

Chain of thought — asking the model to reason step by step before answering. Slower, noticeably better on maths, logic and anything multi-step.

Temperature — the randomness dial. Low means predictable and repetitive; high means creative and occasionally unhinged. For structured output, keep it low.

Negative prompt — a list of what you don't want, common in image generation for killing artefacts and clichés.

If you'd rather skip the theory and get a prompt that works, the Tickd prompt generator builds them for coding, image, video and productivity tasks.

The model-architecture words

RAG (Retrieval-Augmented Generation) — fetch the relevant documents first, then ask the model to answer using only those. This is how you ground answers in your own data without retraining anything.

Embedding — a list of numbers that represents meaning. Similar ideas land near each other, which is what makes semantic search and RAG possible.

Fine-tuning — further training a base model on your examples so it adopts a style or task by default. Usually the wrong first answer: try prompting and RAG before you reach for it.

Distillation — training a smaller, cheaper model to imitate a bigger one. Most of the quality, a fraction of the bill.

Multimodal — handles more than text: images, audio, video in, and sometimes out.

Latent space — the compressed internal map of concepts a model has learned. Generation is really navigation through that map.

Builder vocabulary

Agent — a model given tools and a goal, allowed to loop: plan, act, observe, repeat. The word is wildly overused; ask whether it can actually loop before believing the label.

Tool calling — letting the model invoke your functions with structured arguments instead of describing what should happen.

MCP (Model Context Protocol) — a shared standard for plugging models into external tools and data.

Structured output — forcing answers into a fixed shape, usually JSON matching a schema, so your code can parse it safely.

Streaming — sending the reply token by token so the interface feels alive rather than frozen.

Guardrails — input filtering, output validation and refusal rules that stop bad responses reaching users.

Rate limit — the cap on requests or tokens per minute. Hit it, get a 429, back off and retry. Don't hammer it.

Vibe coding — describing what you want in natural language and iterating on what the AI produces, rather than typing every line. Ticks along nicely until you need to debug something you never read.

Creative-tool vocabulary

Diffusion model — starts from noise and repeatedly denoises it into an image, guided by your prompt.

Seed — the random starting number. Same seed plus same prompt equals a reproducible image, which is how you iterate deliberately instead of rerolling forever.

Inpainting — regenerating only a masked region while leaving the rest untouched. The fix for hands.

Upscaling — raising resolution after generation, inventing plausible detail rather than stretching pixels.

Style reference — an image supplied alongside the prompt so the model borrows its palette, texture and lighting.

Aspect ratio — the frame shape. Models compose genuinely differently at 16:9 versus 1:1, so set it on purpose.

To see how these play out per platform, the deep dives on Midjourney, Higgsfield and Figma Weave cover where each one actually shines, and each platform's own gallery shows live examples better than any screenshot we could paste here.

How to actually learn these

Don't memorise the list. Pick the five terms that touch what you're building this month — probably token, context window, system prompt, RAG and structured output — and use them out loud until they stop feeling borrowed. The rest arrive on demand.

When one shows up mid-task and you need the definition in five seconds, that's what the glossary is for: search it, filter by category, get one sentence, get back to work. And if the thing you're stuck on is a tool misbehaving rather than a word you don't know, the platform deep dives at Tickd point to the right support site — OpenAI, Claude or Gemini — instead of leaving you guessing.

Jargon's only gatekeeping until someone explains it. Consider it ticked off.