#58 | AI Glossary – 80 AI terms you pretend to know

TL;DR: An AI glossary of 80 AI terms you’ve probably heard, maybe used, and possibly can’t quite explain. No fluff. No PhD required.

AI glossary lovers ☀️,

A few weeks back, someone mentioned “model collapse” at a meetup. I nodded like I got it—I’d seen it all over LinkedIn, AI newsletters, that viral thread about the web filling up with AI slop.

But when someone else asked me to actually explain it, I froze. Something about AI eating its own output? Training data getting weird? I mumbled something vague and changed the subject.

Yikes. I thought I knew how to explain better.

We’re all swimming in AI terminology now. Tokens. Agents. Guardrails. Hallucinations.

These words show up in product updates, strategy decks, and Slack threads. We nod along. We use them confidently. But if someone pressed us for a quick, clear explanation?

Cue the awkward pause.

So I created my own AI glossary.

Not an exhaustive technical AI glossary and dictionary—I skipped the purely academic concepts, the AI terms only researchers whisper to each other, and the jargon that sounds impressive but means nothing to your actual work.

What’s here are AI terminologies that show up in real conversations, real tools, and real decisions you’re making this week.

Each entry is short. Some are serious. Some are a little cheeky. All are true.

Ready? Let’s go.

AI Glossary

1. AI Core Concepts & Foundations

The basics that shape how AI works—and what it can’t do. Start here if you’re still nodding along in meetings while secretly Googling under the table.

Generative AI — AI that creates new stuff (text, images, code) instead of just analyzing what already exists. The “generative” part is why your CEO suddenly cares about AI.

Large Language Model (LLM) — AI trained on massive amounts of text to generate human-like responses. ChatGPT, Claude, Gemini—all LLMs. Basically, very expensive autocomplete that somehow learned to write poetry.

Multimodal Models — AI handling multiple inputs at once: text, images, audio. GPT-4 analyzing your messy whiteboard photo and making sense of it? That’s multimodal doing its thing.

SOTA (State-of-the-Art) — The best-performing model available right now. Changes constantly. Often overhyped. Probably outdated by the time you finish reading this sentence.

AGI (Artificial General Intelligence) — Theoretical AI with human-level intelligence across all domains. Doesn’t exist yet—despite what the headlines, Twitter threads, and that one guy at parties keep insisting.

Narrow AI — AI is excellent at one specific task but completely useless at everything else. Your spam filter is brilliant at catching spam. Ask it to write a poem and it just stares at you.

Tokens / Context Window — Tokens are word chunks that AI processes. The context window is how much it remembers at once. Think: your entire novel versus a single tweet. That’s why long documents sometimes confuse it.

Parameters / Weights — Internal settings that let a model make decisions. More parameters usually mean smarter—but also slower, hungrier for compute, and way more expensive to run.

Training vs. Inference — Training builds the model. Inference runs it. Baking the cake versus eating it. You mostly deal with inference. Someone else did the baking.

Pretraining — When AI learns from massive data before you ever interact with it. ChatGPT knew the internet before you typed “hello.” You’re just the latest conversation.

Bias (AI context) — When AI systematically favors certain outcomes because of its training data. Not a moral failing—a technical one. But the effects can be very real.

2. Prompting & AI Workflows Terminologies

How you talk to AI—and actually get it to do what you need. Because “just ask it” is rarely that simple.

Prompt — The instruction you give AI to get a response. Your opening move. Some people overthink this. Others don’t think about it at all. Both approaches have consequences.

Prompt Engineering — Systematic, tested approach to crafting prompts for reliable results. The opposite of “let me just try random stuff and hope it works.”

System Prompt — High-level instructions setting AI’s overall behavior before you ask anything specific. The invisible rulebook running in the background.

Zero-Shot Learning — Asking AI to do something with zero examples. Impressive when it works. Frustrating when it doesn’t. You’re hoping AI just… figures it out.

Few-Shot / One-Shot Learning — Showing AI examples so it mimics the pattern you want. Three to five examples for a few-shot. One for one-shot. More examples, more clarity. Usually.

In-Context Learning — AI picking up patterns from examples you provide in the same conversation—no retraining needed. It’s reading the room. Sometimes quite well.

Chain of Thought (CoT) — Making AI show its work. Step-by-step reasoning instead of jumping straight to conclusions. Helps catch mistakes. Also helps you spot when it’s confidently wrong.

Breaking Down Complex Tasks — Splitting big goals into smaller steps so AI doesn’t lose the plot halfway through. Surprisingly effective. Surprisingly underused.

Copilot — an AI assistant helping with specific tasks like coding, writing, or design. Think co-pilot, not autopilot. You’re still flying this thing.

Tool Calling / Function Calling — Letting AI trigger real actions—sending emails, checking calendars, querying databases—instead of just generating text. Where things get genuinely useful. And occasionally risky.

3. AI Training & Model Development

How AI gets built and customized for specific jobs. Mostly happens before you ever touch it.

Fine-tuning — Teaching a pretrained model your specific style or domain without starting over from scratch. Efficiency meets customization.

Knowledge Distillation — Training a smaller, faster model by learning from a bigger one. That’s why DistilBERT exists. Same vibes, smaller footprint.

Synthetic Data — Artificially generated data used for training when real data is limited, private, or ethically complicated. Fake data solving real problems.

Structured vs. Unstructured Data — Spreadsheets are structured. PDFs, emails, Slack messages? Unstructured chaos. AI handles both—with varying degrees of grace.

Embeddings — How AI represents meaning as numbers so it can understand similarity and relationships. “Dog” and “puppy” end up close together. “Dog” and “refrigerator” don’t.

Grounding — Tying AI answers to actual facts so it stops inventing confident nonsense. Critical for anything important.

4. Agentic Workflows & AI Agents Terms

AI that actually does things, not just answers questions. The part everyone’s excited—and slightly nervous—about.

AI Agent — AI taking autonomous actions, not just generating responses. More than a chatbot. Less than Skynet. For now, anyway.

Agentic Workflow — Systems where AI plans and executes multi-step tasks with minimal human hand-holding. The future everyone keeps talking about. Arriving slower than the hype suggests.

Human-in-the-Loop (HITL) — Checkpoints where humans review AI decisions before execution. Not micromanagement—safety. There’s a difference.

5. Data, Retrieval & RAG AI Terms

How AI stays current and factual using your actual data. Because the internet circa 2023 isn’t always helpful.

RAG (Retrieval-Augmented Generation) — AI looks up relevant documents before answering, staying grounded in your knowledge base. The reason your company’s AI assistant can actually answer questions about your policies.

Vector Database — A specialized database storing data as numbers so AI can find similar information fast. The quiet hero behind every RAG system.

Grounding — Anchoring responses to real sources so AI stops making things up. Same concept as above, but important enough to mention twice.

6. AI Glossary for Tools, Settings & Performance

The controls you actually touch when using AI tools. Yes, those sliders in the settings menu do something.

Temperature — Lower means predictable, consistent answers. Higher means creative, varied responses. It’s the randomness dial. Turn it up for surprises. Turn it down if surprises scare you.

Top-P / Nucleus Sampling — Controls diversity of word choices by limiting AI’s options. Similar to temperature, it works differently under the hood. Most people never touch this.

Top-K — Restricts AI to choosing from the K most likely next words. Fewer choices, more focused output. Another lever most people ignore.

Max Tokens — Sets maximum response length. Hit the limit and the answer just… stops. Mid-sentence sometimes. Annoying if you didn’t plan for it.

Output vs. Input Tokens — Tokens you give AI versus tokens it generates back. Sometimes priced differently. Always counted. Matters for budgeting.

Latency — How long AI takes to respond. Lower is faster. Higher means you’re waiting. And waiting. And wondering if it crashed.

Inference Cost / Cost per Token — The price you pay for running the model, measured per token. Scales fast when you’re generating lots of content—or just running lots of queries.

GPTs / Custom GPTs — Personalized ChatGPT versions with specific instructions, knowledge, or actions built in. Your own AI specialist without building anything from scratch.

7. AI Terms for Hallucinations & Reliability

Making sure AI outputs are trustworthy. Because confident doesn’t mean being correct.

Hallucination — AI confidently generating false information. Not lying—it genuinely doesn’t know it’s wrong. Dangerous in high-stakes contexts. Amusing in low-stakes ones.

Guardrails — Boundaries built into AI to prevent certain outputs or behaviors. Safety nets for production systems. Less restrictive than they sound. Usually.

Evals / Evaluations — Tests measuring how well AI performs on specific tasks. Quality control before you deploy something to actual humans.

Watermarking — Invisible markers in AI-generated content so it can be traced and verified later. Useful when you need to prove something wasn’t written by a person.

Model Card — Documentation showing what data trained the model and its known limitations. Informed consent for AI. Often skipped. Shouldn’t be.

8. AI Terms for Governance, Risk & Ethics

How organizations deploy and manage AI responsibly. The stuff that keeps legal teams up at night.

AI Governance Framework — Rules and processes for safe, compliant AI use. Think NIST, ISO 42001, or whatever your company cobbled together last quarter.

Shadow AI — Unauthorized AI tools employees use without IT or compliance approval. It’s everywhere. You probably don’t know about half of it. Neither does your IT team.

AI Literacy — Workforce capability to understand and responsibly use AI. Increasingly non-negotiable. Decreasingly common.

Model Risk Management — Treating AI models like financial risks that need monitoring and controls. Borrowed from banking. Applied to algorithms. Makes more sense than it sounds.

Explainability (XAI) — Making clear why AI made a particular decision. Critical for regulated industries. Helpful for trust-building everywhere else.

Algorithmic Bias — Systematic unfairness baked into AI from training data or design choices. Not malice—still damaging. Still your problem to address.

Red Teaming — Ethically trying to break AI to find vulnerabilities before bad actors do. Offense as defense. Strangely satisfying when done well.

Prompt Injection / Jailbreaking — Hacking AI’s instructions to bypass guardrails. The reason security teams lose sleep. The reason guardrails exist in the first place.

Copyright / IP Liability — Legal questions about who owns AI-generated content. Still being fought in courts worldwide. No clear answers yet. Probably won’t be for a while.

Deepfake — Synthetic media (video, audio) made by AI that can impersonate real people convincingly. Impressive technology. Terrifying implications.

Alignment — Ensuring AI goals and values match human values. Sounds simple. Harder than it sounds. Becomes critical as systems get more powerful.

9. AI Terms for Healthcare

AI in clinical work and patient care. Where the stakes are highest and the regulations strictest.

Clinical Decision Support (CDS) — AI assisting doctors with diagnoses and treatment recommendations. Not replacing doctors—helping them process more information faster.

Ambient Scribing — AI listening to doctor-patient conversations and automatically writing clinical notes. Saves hours of admin work. Raises obvious privacy questions.

Precision Medicine — AI-tailored treatment recommendations based on individual genetics, history, and lifestyle. Personalized healthcare, not one-size-fits-all.

PHI (Protected Health Information) — Sensitive patient data protected under HIPAA. Requires careful handling. Requires even more careful AI handling.

10. AI Glossary for AI in Finance

AI in money, markets, and compliance. Fast, data-heavy, heavily regulated.

Robo-Advisors — AI systems automatically managing investment portfolios based on risk tolerance and goals. Financial planning without a human advisor. For better or worse.

Fraud Detection — AI spotting unusual transaction patterns in real-time. Your credit card company’s favorite use case. Works surprisingly well.

Algorithmic Trading — AI making buy/sell decisions in financial markets automatically. Speed and data processing humans simply can’t match. Also, occasionally, flash crashes.

KYC/AML Automation — AI automating “Know Your Customer” and anti-money laundering compliance checks. Faster, cheaper, more consistent. Still imperfect.

11. AI Glossary for Creative & AI Marketing

AI as a creative partner and marketing accelerator. Where things get fun—and ethically complicated.

Generative Search — Search results AI generates with synthesis and summaries, not just links. Google’s future. Your SEO team’s nightmare.

Text-to-Image — AI creating images from text descriptions. DALL-E, Midjourney, Stable Diffusion. The reason your feed is suddenly full of weird art.

Text-to-Video — AI generating video from text prompts. Sora made headlines. Still early days. Still genuinely impressive when it works.

In-painting / Out-painting — AI editing images by filling in or extending areas. Remove that photobomber. Extend the landscape. Magic, basically.

Style Transfer — AI copying an artist’s visual style onto new content. Technically impressive. Ethically debatable. Legally murky.

Augmented Creativity — AI as co-pilot in creative work, not replacement. Amplifying imagination rather than automating it away. At least, that’s the optimistic framing.

12. AI Glossary for AI in Education

AI in learning and teaching. Lots of potential. Lots of policy debates.

Adaptive Learning Paths — AI personalizing learning experiences to each student’s pace, style, and knowledge gaps. Education that actually adapts. Finally.

AI Ethics in Education — Teaching students responsible and ethical AI use. Critical literacy for the generation that will build with—and on—this technology.

13. AI Glossary for Emerging & Future-Facing Terms

Where AI is heading. Some of this is real. Some is hype. Figuring out which is which is part of the fun.

AI-Native / AI-First — Products or companies built entirely around AI capabilities. Couldn’t exist without it—not just enhanced by it.

Small Language Models (SLMs) — Efficient, cheaper AI models running on phones or local computers. Speed and privacy without cloud dependence. Getting better fast.

Edge AI — AI running on devices instead of the cloud. Faster, private, and works offline. Your phone is smarter than it looks.

AI Wearables — AI embedded in smart glasses, pins, and earbuds. Ambient intelligence you wear. Still finding its footing. Probably won’t stay awkward forever.

Model Collapse — The risk that AI trained on AI-generated content degrades over time. AI is eating its own tail. A real concern, not sci-fi.

Autonomous Agents — Self-directed AI that plans and executes without human prompting. Potential and peril in equal measure. The exciting-scary part of the conversation.

Ambient Intelligence — AI woven invisibly into environments, anticipating needs and acting quietly. Smart homes, cities, and workplaces of the future. Subtle until it isn’t.

Ok. That’s the AI glossary. Bookmark it. Reference it. Use it the next time someone drops ‘SOTA,’ ‘model collapse,’ ‘shadow AI,’ or even “SOTA RAG evals” in Slack, and you want to know what they’re actually talking about.

AI moves fast. These AI terms (may) stick around. Understanding them means you can participate in the conversation—not just nod along and hope nobody asks follow-up questions.

See you next week,

Mark
The AI Learning Guy
👋⚡😎

Other AI Glossary Sources

Not enough? Fair. Here are some of the best AI glossaries out there—for when you need a second opinion or want to go full rabbit hole.

The heavy hitters:

• ​Andreessen Horowitz AI Glossary​ — VC perspective, solid technical definitions, regularly updated
• ​TechCrunch: Simple Guide to Common AI Terms​ — Plain language, journalist-friendly, keeps up with the news cycle
• ​Wikipedia: Glossary of Artificial Intelligence​ — Exhaustive, academic, updated two days ago (seriously)

For business and strategy folks:

• ​Zendesk Generative AI Glossary​ — CX-focused, practical for anyone building customer-facing AI
• ​Coursera AI Terms & Definitions​ — Clean, educational, good for teams ramping up on AI literacy

For the technically curious:

• ​Moveworks AI Terms Glossary​ — Enterprise-focused, especially strong on agents and RAG
• ​Pryon RAG & LLM Glossary​ — Deep on retrieval systems and governance

Quick reference:

• ​The Drum: Essential AI Glossary for Marketers​— Skimmable, marketing-friendly

Note: No single website has all the answers. This list serves as a starting point for those who want to explore or satisfy their curiosity about AI.

Links: Links with * are affiliate links. See disclosure below.