Dictionary Wiki
Open any tech article, product launch, or job listing and you will bump into a wall of AI jargon. Words that meant something narrow ten years ago now carry different baggage, and brand-new terms appear every quarter. If you want to read past the marketing copy, you need a working vocabulary that covers the math, the engineering, and the public debates all at once. This guide walks through the words that actually show up in papers, boardrooms, and news coverage, grouped by subfield so you can find what you need without scrolling past three definitions of the same thing.
Table of Contents
- 1. The Core Ideas Behind AI
- 2. How Machines Learn from Data
- 3. Neural Networks and Deep Learning
- 4. Language and NLP Terminology
- 5. Teaching Machines to See
- 6. Generative Models and LLMs
- 7. Robots and Autonomous Machines
- 8. Safety, Ethics, and Accountability
- 9. The Data and Hardware Stack
- 10. Where the Vocabulary Is Headed
1. The Core Ideas Behind AI
Every subfield of AI sits on top of a handful of older ideas about what it means for a machine to think. The terms below show up in almost every other section, so it pays to anchor them first. They also help separate legitimate capability claims from marketing — once you know what general AI would actually require, it becomes easier to spot when the word is being used loosely.
Artificial intelligence (AI) — The umbrella discipline concerned with building software and hardware that can reason, perceive, plan, or learn in ways usually associated with people. The label covers everything from spam filters to self-driving prototypes.
Narrow AI (weak AI) — A system built to be good at one job — sorting email, transcribing speech, beating humans at Go — without any ability to transfer that skill to unrelated problems. Every production AI you have ever used falls into this bucket.
General AI (strong AI) — A hypothetical system that could match a capable adult across any cognitive task, from writing a legal brief to learning a new musical instrument, using the same underlying mind. No one has built one, and there is no consensus on how close we are.
Turing test — Alan Turing's 1950 thought experiment in which a judge converses with a hidden human and a hidden machine and tries to tell them apart. Passing it has long served as a cultural shorthand for "real" intelligence, even if working AI researchers treat it as a philosophical marker more than a technical goal.
Algorithm — A finite, well-defined recipe a computer follows to turn inputs into outputs. Every model, heuristic, and training loop in this guide is ultimately one or more algorithms stacked together.
Keep these definitions handy as you read further. When someone claims a product is "true AI" or worries that "AI will soon be conscious," the argument almost always hinges on which of these words the speaker has in mind.
2. How Machines Learn from Data
Machine learning is where most working AI lives. Instead of writing explicit rules ("if the email contains these words, mark as spam"), engineers hand the computer a pile of examples and let it fit its own rules. That simple shift — from programming behavior to programming a learner — explains why ML tools power everything from Netflix suggestions to credit scoring.
Machine learning (ML) — A family of techniques where a system improves at a task by processing examples, rather than being told step by step what to do. The more representative data it sees, the better its predictions tend to get.
Supervised learning — Training that relies on labeled pairs: a chest X-ray tagged "pneumonia" or "clear," a loan application tagged "repaid" or "defaulted." The model learns the mapping and then applies it to fresh inputs.
Unsupervised learning — Techniques that hunt for structure in unlabeled data, such as grouping customers into behavioral clusters or compressing images into a smaller set of features that still capture what matters.
Reinforcement learning — A setup where an agent tries actions, sees what happens, and adjusts its strategy based on rewards and penalties. It is how systems learn to play Atari games, control data-center cooling, or fine-tune chatbots from user feedback.
Training data — The examples a model studies during learning. Its quality, diversity, and size largely determine whether the final system is reliable or brittle.
Overfitting — When a model memorizes quirks of the training set instead of learning the underlying pattern, it will ace the practice exam and fail the real one. Spotting and preventing overfitting is a daily concern for ML practitioners.
Feature extraction — Turning raw inputs (pixels, log files, transaction records) into a smaller, more useful set of signals the model can learn from. Good features can make the difference between a model that works and one that does not.
These terms describe the actual plumbing of most AI products. Once you can distinguish supervised from unsupervised work, or explain why overfitting matters, vendor pitches and research abstracts become much easier to parse.
3. Neural Networks and Deep Learning
Deep learning is the branch of ML that has driven the loudest headlines of the last decade. By stacking many layers of simple computational units, these models learn progressively more abstract representations of data — edges in early layers, shapes in the middle, whole objects or concepts near the top. The vocabulary below shows up constantly in research papers and product blog posts alike.
Neural network — A model made of connected nodes arranged in layers, loosely inspired by how biological neurons signal one another. Each connection has a weight that gets tuned during training.
Deep learning — ML carried out with neural networks that are many layers deep, which lets the model build up layered abstractions from raw input. Most state-of-the-art vision, speech, and language systems today are deep.
Convolutional neural network (CNN) — A network tuned for spatial data like photographs or satellite images. Convolutional filters slide across the input to pick up local features — a stroke, a corner, a pattern — and deeper layers combine them into objects.
Recurrent neural network (RNN) — A network built for sequences such as audio streams or sentences, where outputs loop back into the network so earlier inputs can influence later decisions.
Transformer — An architecture introduced in 2017 that uses a technique called self-attention to weigh every token in a sequence against every other token at once. Transformers are the reason today's language and multimodal models are as capable as they are.
Backpropagation — The learning algorithm that tells a neural network which weights to nudge, and in which direction, after each wrong answer. Without it, training deep models at scale would not be feasible.
If you can fluently name what a transformer does and why a CNN is a natural fit for images, you can read most product-level AI coverage without getting lost in the acronyms.
4. Language and NLP Terminology
Natural language processing is the art of getting computers to deal with text and speech in ways that feel useful. It blends linguistics, statistics, and engineering, and its vocabulary sits behind chat assistants, search, translation, legal e-discovery, and almost every customer-support tool on the market.
Natural language processing (NLP) — The slice of AI focused on reading, writing, and understanding human language, in both its written and spoken forms.
Tokenization — Chopping raw text into pieces — words, word fragments, or individual characters — that a model can feed into its math. How you tokenize has a surprisingly large effect on final quality.
Sentiment analysis — Estimating the mood of a piece of writing: a glowing product review, a hostile tweet, a neutral news summary. Brands use it to track reputation; researchers use it to study public opinion.
Named entity recognition (NER) — Scanning text for meaningful nouns — people, companies, cities, dates, dollar amounts — and tagging each one. It is a building block for search, compliance, and knowledge-graph work.
Word embedding — A way of representing each word as a vector of numbers so that words with related meanings land near each other in that space. It is what lets a model know "king" and "monarch" are cousins.
As voice interfaces and chat tools spread, NLP terms are creeping into everyday work. Knowing what tokenization is, for example, takes a lot of the mystery out of why a chatbot's answers sometimes break at odd places.
5. Teaching Machines to See
Computer vision gives machines a way to interpret pixels — still images, video frames, depth maps — as meaningful content. It sits behind the camera in your phone, the quality-inspection rigs in factories, medical imaging triage, and the perception stack of every self-driving prototype.
Computer vision — The field that teaches computers to extract useful information from visual data, whether that means counting cars on a freeway or flagging a suspicious mole in a dermatology photo.
Object detection — Finding specific items inside a picture and drawing a box around each one while labeling it. Think of the boxes that pop up around faces in a smartphone camera app.
Image classification — Assigning a single label to a whole image — "golden retriever," "sunset," "defective part" — based on what it contains overall.
Semantic segmentation — Labeling every individual pixel, so the model outputs a color-coded map of road, sidewalk, pedestrian, sky, and so on. It is how autonomous vehicles tell drivable surface from everything else.
Optical character recognition (OCR) — Turning pictures of text — a receipt, a scanned contract, handwritten notes — into editable characters a computer can search, index, or translate.
Vision vocabulary is becoming hard to avoid as cameras and AI chips show up in doorbells, cars, and warehouses. These five terms alone cover most of what you will encounter in product descriptions.
6. Generative Models and LLMs
Generative AI is the category responsible for the last few years of public fascination. Rather than classifying or scoring existing inputs, these systems produce new ones — a paragraph, an illustration, a song snippet, a block of working code. The vocabulary here is changing fast, but a few terms have already settled into common use.
Generative AI — Models that output fresh content — text, images, audio, video, software — by learning the statistical shape of a big training set and then sampling from it.
Large language model (LLM) — A very large transformer trained on enormous text corpora that can draft essays, summarize documents, answer questions, and write code with surprising fluency. Think of GPT-4, Claude, or Llama.
Prompt engineering — The craft of writing instructions that coax a generative model into producing the output you actually want, using tricks like explicit role descriptions, worked examples, or formatting constraints.
Hallucination — When a model states something false with total confidence — a made-up citation, a nonexistent API, a fake quote. Reducing hallucinations is one of the hardest open problems in LLM work.
Fine-tuning — Taking a general-purpose pretrained model and continuing training on a narrow, task-specific dataset so it behaves the way a particular company or use case needs.
Diffusion model — An image-generation approach that starts from pure noise and refines it step by step into a coherent picture, having learned during training how to reverse a gradual noising process.
Generative terms turn up in ordinary conversations now — hiring managers ask about prompt engineering, newspapers write about hallucinations. Being precise about what each word means is quietly becoming a workplace skill.
7. Robots and Autonomous Machines
Robotics is where AI meets hardware. A robot has to sense the world, decide what to do, and then actually move a motor. Autonomous systems — cars, drones, warehouse vehicles — extend the same pattern to machines operating without a steady human hand on the controls.
Basic Robot Parts and Terms
A robot is a programmable machine that carries out physical tasks, often guided by AI software. Its degrees of freedom count how many independent directions its joints can move, which sets the ceiling on how dexterous it can be. Actuators are the motors, pistons, and similar components that turn electrical signals into motion — the hardware equivalent of muscles. End effectors are whatever sits at the working tip: a two-fingered gripper on a pick-and-place arm, a welding torch on an assembly line, or a scalpel attachment on a surgical system.
Finding a Path Through the World
LIDAR fires short laser pulses and times their return to build a precise 3D model of nearby surfaces, which is why it shows up on top of so many self-driving test cars. SLAM — Simultaneous Localization and Mapping — lets a robot figure out both "where am I" and "what does this room look like" at the same time, which is essential when no one has pre-mapped the space. Path planning algorithms then take that map and compute a route that reaches the goal while steering around obstacles and respecting safety rules.
Working Alongside People
Cobots are collaborative robots engineered to share a workspace with humans, typically with padded housings, force limits, and sensors that halt motion when a person gets too close. Teleoperation lets an operator drive a robot from a distance — useful in bomb disposal, deep-sea work, or remote surgery. Haptic feedback sends touch-like signals back to that operator's controls, so a surgeon can sense the resistance of tissue or a technician can feel when a bolt has seated correctly.
8. Safety, Ethics, and Accountability
As models get bigger and show up in more decisions — hiring, lending, policing, medical triage — the vocabulary around responsibility, fairness, and risk has grown up alongside the technical terms. These words increasingly appear in legislation and contract language, not just research papers.
AI bias — Patterned unfairness in a model's outputs, usually inherited from skewed training data or from design choices that encode an implicit worldview. Bias can show up as lower accuracy for some demographic groups or as recommendations that quietly disadvantage them.
Explainability (XAI) — Techniques and design choices that make a model's decisions auditable, so a human can answer the question "why did it say no to this loan applicant?" rather than shrugging at a black box.
Alignment — The research problem of getting an AI system to reliably pursue what its designers and users actually want, including goals that are hard to specify precisely, like "be helpful but don't help me harm anyone."
Responsible AI — An organizational approach that bundles fairness testing, documentation, privacy protection, and human oversight into the way AI products are built and shipped, rather than bolted on at the end.
Deepfake — Synthetic audio or video, produced with generative models, that makes a real person appear to say or do something they never did. Deepfakes are now a standing concern for elections, impersonation scams, and non-consensual imagery.
For a non-specialist audience, this is probably the most consequential group of terms. The policy debates over AI — who is liable, what can be regulated, which uses should be banned — happen entirely in this vocabulary.
9. The Data and Hardware Stack
A model is only as good as the data feeding it and the machines training it. The terms below describe the infrastructure layer: where the numbers come from, where the heavy computation happens, and how models end up close to the user.
Big data — Datasets too large, fast-moving, or varied for a single database to handle comfortably. The classic shorthand describes them with four Vs: volume, velocity, variety, and veracity.
GPU (Graphics Processing Unit) — A chip originally built for 3D graphics that happens to be ideal for the massive parallel arithmetic neural networks require. Modern AI training clusters are stuffed with them.
Data pipeline — The chain of automated steps that pulls raw data from its sources, cleans it, transforms it, and delivers it to wherever analysts or models will use it. Broken pipelines are a leading cause of broken AI.
Cloud computing — Renting storage and compute from a provider's data centers over the network, which lets even small teams train or serve sizeable models without buying their own hardware.
Edge AI — Running models directly on phones, cameras, cars, or factory sensors instead of sending data back to a central server. The payoff is lower latency, better privacy, and operation even when the network is patchy.
These terms explain the less glamorous half of AI — the reason training runs cost millions, the reason data cleanup eats most of a project's timeline, the reason your voice assistant can respond quickly even with bad Wi-Fi.
10. Where the Vocabulary Is Headed
AI's word list is not a fixed syllabus. New concepts arrive almost monthly as research groups name techniques, companies coin product categories, and regulators invent compliance terms. Phrases like retrieval-augmented generation, mixture-of-experts, agentic workflow, model card, and AI Safety Institute barely existed a few years ago and are now routine in technical and policy conversations.
The good news is that the vocabulary tends to layer. Once you are comfortable with the core ideas in this guide — what training means, what a transformer is, what alignment is trying to solve — new terms usually slot into place as refinements or combinations of things you already understand. You do not need a PhD to keep up; you need patience and a habit of looking words up when they appear.
Whatever role you play — student picking a major, developer retooling your stack, manager evaluating a vendor, citizen reading a policy proposal — this shared glossary is what makes serious conversations about AI possible. Treat the terms here as a starting set, revisit them as the field shifts, and you will find the next wave of buzzwords much less intimidating than the last one.
Look Up Any Word Instantly on Dictionary Wiki
Get definitions, pronunciation, etymology, synonyms & examples for 1,200,000+ words.
Search the Dictionary