AI Glossary: 50 AI Terms Explained in Plain English

Key Takeaway

• 📖 This AI glossary is the site’s plain-English base layer: every core AI term in one place, each defined in 40-60 words an ordinary reader can absorb in one pass.
• 🔗 Depth is one click away: where WorldNgayon has published a full explainer, the definition links straight to it — the glossary is a map, not a replacement.
• 🧠 Six groups, no decoding required: foundations, models and training, hardware and infrastructure, safety and security, agents, and the rules and money around them.
• 🔁 It updates as the AI world does: new terms ship when they enter working vocabulary, verified at each pass rather than written once and forgotten.

AI vocabulary expands faster than any technology vocabulary in history, and most explainers assume you already know the words. This AI glossary reverses that: short definitions first, deep coverage one click away. Terms are grouped by what they touch — how models work, what they run on, how their safety is tested, and what they cost — so you can find a term by the question you are actually trying to answer.

How to read it: every definition stays deliberately short, in plain language, and never invents numbers. When a deeper explainer exists, the definition’s link carries you to the verified version with sources, tools and worked examples. Start with the group matching your question, and return here whenever a new name shows up in a product announcement.

Foundations: the Words Everything Else Rests On

These are the AI glossary entries new readers meet first — the basic moving parts of every modern AI system.

Large language model (LLM) — An AI system trained on huge amounts of text to predict and generate language. An LLM produces fluent answers without understanding them the way a person does, and the quality of its answers depends heavily on its training data and the settings you give it.

Generative AI — The family of AI models that create new content — text, images, audio, video or code — rather than only sorting or predicting. Generative AI learns statistical patterns from examples and produces new material that resembles them, which is also why fabrications look so convincing.

Chatbot — A conversational application wrapped around an AI model. Modern chatbots keep session context, follow instructions and call tools; consumer versions share your data for training unless you change the defaults, which we document in our ChatGPT memory and AI privacy settings walkthroughs.

Neural network — A model built from layers of connected “neurons” that convert inputs into outputs by weighting signals. Learning adjusts those weights from examples; the architecture determines what the network can represent efficiently.

Transformer — The neural-network architecture behind today’s LLMs. Its key trick, self-attention, lets the model weigh how every token relates to every other token, which is what makes long, coherent text generation possible at scale.

Token — The small chunk of text (roughly a word, part of a word, or a few characters) that models read and produce. Pricing, context limits and speed are all counted in tokens, which is why two sentences of similar length can cost the same or wildly different amounts.

Context window — How much text a model can “see” at once, measured in tokens: your conversation, the documents pasted in, and the model’s own reasoning. Everything outside the window is invisible to that turn. Our context-engineering guide shows how window limits shape real workflows.

Prompt — The instruction and material you hand a model in one turn. Prompts carry intent, constraints and examples; vague prompts produce vague outputs. Prompt composition at scale — loading the window with the right material — is called context engineering.

Multimodal model — A model that handles more than one data type in one system — text with images, audio, or video in and out — so a single chat can accept a photo, a voice note and a document and reason across all three.

Parameters — The learned internal weights of a model, counted in billions. More parameters usually mean more capability and more serving cost, though modern designs decouple the two — which is exactly what mixture-of-experts models below are about.

Training data — The examples a model learns from, plus everything around it: web text, licensed corpora, code, user-allowed chats. The data’s breadth and quality cap what the model can do, and who gets a say in what is fed in is where AI privacy law now lives.

Models, Training and Performance

How raw networks grow into the tools you use — the AI glossary terms labs argue about in public.

Foundation model — A large, general model trained once at scale and then adapted to thousands of tasks. Because they anchor entire product stacks, the strongest foundations now arrive wrapped in restrictions — including the clearance requirements we covered for frontier-grade access.

Fine-tuning — Continuing a base model’s training on narrower data so it performs a specific job better — say, legal filings or customer questions — without the cost of pretraining from scratch.

RLHF — Reinforcement learning from human feedback: people rank model answers, and the model is trained to prefer answers like the ones humans rated highly. It is the standard technique for turning a raw text predictor into a usable, safer assistant.

Open weights — A model whose learned parameters anyone can download and run themselves, even if the training code or data stays closed. The open-weights class moved fast enough to take real benchmark leads, as we tracked when Xiaomi’s MiMo run took the open-weights crown.

Distillation — Training a small model to imitate a big one’s outputs, so you can serve the answers at a fraction of the cost. It is a legitimate technique — and a contested intelligence practice, which is why a disputed distillation campaign became one of the year’s biggest industry fights.

Quantization — Compressing a trained model by lowering the precision of its numbers (16-bit down to 8- or 4-bit) so it runs on cheaper hardware with a small quality trade-off. It is the main reason capable models now fit on consumer machines.

Mixture of experts (MoE) — A design in which only some of the model’s “expert” sub-networks activate per request, so total parameters can be huge while serving cost stays low. Our MoE inference explainer covers why these models serve faster.

Inference — The act of running a trained model to answer a request — as opposed to training, which builds it. Inference is the recurring cost of every AI product, and efficiency there decides what services you get at what price.

Test-time compute — Spending more processing while answering a single question — letting the model think longer — instead of only training longer. The knob changed how frontier models handle hard problems; we explain it in our test-time compute explainer.

Reasoning model — A model tuned to work through problems step by step before answering, often using extended thinking. Strong on math, code and planning; slower and pricier per question than fast-chat models.

Benchmark — A standardized test for comparing models — exams, coding suites, agent tasks. Benchmarks drive competitive marketing, but training-set leaks and gaming make them a starting signal for vendor choice, not the finish line.

AGI (artificial general intelligence) — The hypothesized point where a system matches humans across most cognitive work, not just one domain. Labs treat it as a goalpost; skeptics treat it as marketing shorthand. We map the difference in superintelligence vs AGI explainer.

Singularity — The speculative moment of self-improving, runaway AI progress that human institutions can no longer keep pace with. It is the most overused term in the field — see our singularity reality check for what the claim actually carries.

Hardware and Infrastructure

Where AI physically lives — the AI glossary terms behind your monthly subscription.

GPU — The graphics processor repurposed as the workhorse of AI: its parallel math suits both training and inference. Prices, rental rates and shortage cycles for GPUs shape what AI services cost you downstream — our GPU rental guide prices an hour of compute.

TPU — Google’s custom AI accelerator, sold access to through its cloud and used in giant capacity deals; the Google-Broadcom TPU partnership behind gigawatt-scale commitments shows how accelerator supply now decides model roadmaps as much as research does.

HBM (high-bandwidth memory) — The fast-stacked memory hugging each accelerator chip. AI demand outgrew HBM factories, and the squeeze raised prices for laptops and phones too — the economics are in our HBM shortfall explainer.

Co-packaged optics (CPO) — Putting optical network connections right at the chip package instead of on separate plug-in modules, cutting the energy AI clusters waste moving data. Explainers live in our co-packaged optics piece.

AI rack density — How much computing fits in one server rack, now pushed to hundreds of kilowatts by liquid cooling — the engineering behind the numbers is in our rack-density walkthrough.

Data center — The buildings housing AI computers, consuming power, water and land at a scale that reshapes local utilities. Water draw is the contested part; the measured numbers sit in our data-center water explainers.

Edge AI — Model inference running on your device — phone, laptop, car — rather than in a distant data center. It moves work off the network for privacy and latency reasons; the trade-offs are mapped in our edge-AI explainer.

Sovereign AI — A country’s push to host and control its own AI computing and models rather than renting everything from foreign clouds. The national compute race, and who actually has sovereignty, is the subject of our sovereign-AI analysis.

Agentic infrastructure — The server, protocol and tooling layer that lets AI agents act — browse, file, pay, deploy — instead of only chatting; it is where the industry’s spending is currently heading, and where its security story is weakest.

Safety, Security and Trust

The AI glossary terms decide what to believe and where models break.

Hallucination — A confident answer the model invented: wrong facts, fake citations, invented prices. It happens because the model predicts plausible text, not verified truth. Our hallucination explainer shows catching methods.

Prompt injection — An attack where hostile instructions hidden in a document, webpage or email steer an AI into obeying the attacker rather than you. It is the top AI-security failure mode in practice; defenses live in our prompt-injection defense guide.

Jailbreak — A crafted prompt that talks a model past its safety rules into producing disallowed content. Every model generation gets a new wave; the game never stays won.

Guardrails — The automated checks around a model — input filters, policy classifiers, output scanners — that catch disallowed content before it reaches you. Strong products publish safety evaluations so you can see what the guardrails were tested against.

Red teaming — Organized attack-testing of an AI system by friendly attackers before the real ones arrive — probing jailbreaks, injection routes and tool abuse. Serious labs red-team frontier models before release; serious vendors publish the results.

Alignment — The research program that tries to make AI systems genuinely want to do what their operators intend, not merely appear to. It is the technical version of the safety debate, carried in public by figures like Zuckerberg’s safety-markets framing and every frontier-lab charter.

Deepfake — AI-generated audio, video or images of real people doing and saying things they never did. Detection tools lag generation, so verification habits matter first — see the deepfake detector guide and the family defense pattern in our voice-cloning scams piece.

Voice cloning — Reproducing a person’s voice from short samples well enough to fool relatives and banks. It powers the cruelest scam wave of the decade; the countermeasure is a family code word, per our voice-cloning scams explainer.

Synthetic media — Any content wholly or partly machine-generated — text, images, video, voice, music. The label sounds neutral; the hard part is provenance: knowing which parts are real when generation is free.

Content provenance (C2PA) — A cryptographic standard that attaches an edit history to media files — which camera or model created it, and what was changed after — documented by the C2PA coalition itself. It is the strongest current answer to “is this real,” and its adoption is spreading through cameras, platforms and AI generators.

Vibe coding — Building software by describing what you want in natural language and letting an AI write most of the code, rarely reading the result. Shipping it without checks turns you into the attack surface; our vibe-coding security checklist covers the pre-launch pass.

Agents, Tools and Data

Beyond chat: the AI glossary vocabulary for systems that act.

AI agent — A system that pursues multi-step goals: it plans, calls tools, checks results and adjusts, rather than answering one prompt at one time. Agents can shop, book, file and code; autonomy is a dial you set, not a property it simply has.

Computer use — Agents that operate software the way a person does — moving the cursor, typing, clicking through screens — so they can use apps with no special API. A survey of what current systems can actually do is in our computer-use agents explainer.

MCP (Model Context Protocol) — An open protocol for connecting AI models to the data and tools around them — files, databases, apps — through standard “servers” instead of a bespoke connector per app; the official MCP site hosts the specification. The base standard is explained in our MCP explainer and the managed-server wave in Google’s managed MCP piece.

RAG (retrieval-augmented generation) — A pattern where the system first fetches relevant documents, then has the model answer from them. It grounds answers in your materials and reduces hallucination — it does not eliminate it, because retrieval can miss and sources can be wrong too.

Vector database — A database built to store text and other data as vectors, the numeric fingerprints a model produces, so similar items can be found by similarity rather than exact keyword match. It is the storage backbone of most RAG systems.

Embedding — The numeric fingerprint a model assigns to a piece of text, image or sound, placing similar meanings near each other in high-dimensional space. Search, recommendations, and retrieval-augmented systems all run on embeddings.

Memory (model feature) — When an AI app remembers things between conversations: facts about you, your preferences, your ongoing work. It is a controlled feature — our ChatGPT memory piece shows how to turn it off — distinct from the training question in your AI privacy settings.

Agentic checkout — A shopping flow where the AI agent completes a purchase end-to-end on your behalf — selection, cart, payment — with guardrails on spend and authority. The live deployment is documented in our agentic checkout piece.

AI note taker — Services that join your meetings, transcribe, and summarize with AI. Their privacy terms decide who trains on your conversations — mapped provider by provider in our AI note-taker privacy ledger.

The Rules and the Money

Where AI meets law, clearance and invoices — the AI glossary terms of the last mile.

AI Act (EU) — The European Union’s risk-tiered AI law, as the Commission’s own policy page states: banned uses, a strict tier for high-risk systems, transparency duties for the rest. Its Article 50 transparency rules — what must be disclosed as AI-generated or AI-made — are covered in our Article 50 explainer.

Model clearance requirements — Access controls that gate the strongest frontier models behind vetting — for government users or sensitive deployments, proving you can secure their outputs before they ship. The trend and its implications are in our clearance-requirements piece.

Compute rental — Renting GPU capacity by the hour from cloud providers instead of buying hardware. Rates swing enormously with demand; what an hour actually costs, and how to buy it well, is priced out in the GPU rental guide.

Chatbot economics (small business) — What running a business AI chatbot really costs in subscription, build and maintenance — versus what it actually saves. Real numbers and data realities are in our chatbot cost ledger.

AI data privacy — The rules and reality of what happens to work you paste into AI services: training opt-outs, enterprise exclusions, retention windows, and where your confidential material actually travels. The full map is in our AI data privacy ledger.

Model release cadence — How fast labs ship new models and versions — now counted in dozens of notable releases per year, which is why any “latest model” claim carries its own expiry date. The pace numbers are tracked in our release-race piece.

Using This AI Glossary

Three habits make this AI glossary pay. First, when a news story hits an unfamiliar name — a new accelerator, a new safety term — look it up here before trusting any take on it, including ours; the definitions carry no agenda of their own because they cite the documented explainers. Second, treat the linked pieces as the real product: each is a full investigation with verified numbers, while a glossary definition is only the door to it. Third, flag gaps — a term you hit in the wild that is not here is a gap worth closing, and the list grows by review and verification, which is exactly how every other hub on this site stays current.

The AI glossary does not chase jargon for its own sake. Each entry earned its place by appearing in real product decisions, real regulation, or real security incidents that our readers face — which is also why the linked explainers lead with verified primary sources and carry their own update triggers rather than pretending these words stand still.

FAQ

What is this AI glossary for?

One plain-English definition per AI term, in 40-60 words, with a link to our full explainer whenever one exists — so the answer to “what does this word mean” arrives in seconds, and the verified depth arrives one click later.

How is the AI glossary different from a dictionary?

A dictionary defines in isolation. This AI glossary links every term that we have covered deeply — models, hardware, safety, agents, regulation — so each short definition connects to a sourced investigation with real numbers rather than leaving you with a slogan.

Will the AI glossary add new terms?

Yes. When a term starts appearing in working decisions — not just in vendor slideware — it gets researched, defined plainly, and linked where depth exists. Terms that remain marketing-only stay out until they mean something.

Where should a complete beginner start in the AI glossary?

Foundations. The first group — LLM, token, context window, prompt — is all a newcomer needs to understand any AI announcement, and each entry points to a deeper walkthrough once the base vocabulary is in place.

Does every term link to a longer article?

No — only where a real explainer exists. Terms without a dedicated piece get a complete plain-language definition here anyway, and new depth pieces are the natural next stop as the site expands its coverage.

How current are the AI glossary definitions?

Each linked piece carries its own update triggers, and the glossary itself is reviewed at every update pass. When a term’s meaning shifts — as it did around Gemini’s activity controls or Claude’s training settings — the definition follows the verified change, not yesterday’s wording.

@online{agron2026aiglossary,
  author       = {Agron, Edmon},
  title        = {{AI Glossary: 50 AI Terms Explained in Plain English}},
  year         = {2026},
  date         = {2026-11-10},
  organization = {WorldNgayon},
  url          = {https://worldngayon.com/ai-glossary/},
  urldate      = {}
}

TY  - ELEC
AU  - Agron, Edmon
TI  - AI Glossary: 50 AI Terms Explained in Plain English
T2  - WorldNgayon
PY  - 2026
DA  - 2026/11/10/
UR  - https://worldngayon.com/ai-glossary/
Y2  - 
ER  -