Artificial intelligence comes with a blizzard of jargon — parameters, tokens, hallucinations, agents, alignment — and half the time, even the people using the words seem unsure what they mean. This AI glossary for beginners cuts through the noise: more than sixty essential AI terms, each explained in plain English with no assumed background. Whether you are a student, a professional trying to keep up, or just AI-curious, this is your reference for understanding what everyone is talking about.
Terms are grouped by theme so you can read straight through or jump to what you need. Bookmark this page — you will come back to it.
Core Concepts: What AI Actually Is
Artificial Intelligence (AI)
The broad field of building machines that perform tasks requiring human-like intelligence — understanding language, recognizing images, making decisions, solving problems. "AI" is an umbrella term; nearly everything else in this glossary is a specific type of AI or a technique used to build it.
Machine Learning (ML)
A subfield of AI where systems learn from data rather than being explicitly programmed with rules. Instead of a programmer writing "if the email contains these words, mark it spam," an ML system studies thousands of labeled emails and figures out the patterns itself. Most modern AI breakthroughs are machine learning breakthroughs.
Deep Learning
A subfield of machine learning that uses artificial neural networks with many layers (hence "deep"). Deep learning powers today's biggest advances: image recognition, language models, voice assistants, and game-playing systems. It became practical in the 2010s as data and computing power grew.
Neural Network
A computing model loosely inspired by the brain: layers of interconnected "neurons" that transform inputs into outputs. Each connection has a weight, and training adjusts those weights until the network produces the right answers. A modern language model is a neural network with billions of such weights.
Algorithm
A precise set of step-by-step instructions for solving a problem — the recipe a computer follows. In AI discussions, "the algorithm" usually means the overall method a system uses, from a simple sorting procedure to a complex training routine.
Model
In AI, a model is the trained artifact — the neural network with its learned weights — that you actually use. Training produces the model; inference is running the model to get answers. When people say "a new model dropped," they mean a newly trained and released AI system.
Training
The process of teaching an AI model by feeding it data and adjusting its internal weights to reduce errors. Training a frontier language model can take months and enormous computing resources; using the finished model (inference) is comparatively cheap.
Inference
Running a trained model to produce outputs — asking a chatbot a question and getting an answer. Inference is what happens every time you use an AI product; training is the one-time (or periodic) process that created it.
Parameters
The adjustable internal values of a neural network — its "weights." Model size is often described by parameter count (millions, billions, or trillions). More parameters generally mean a more capable model, though architecture and training data matter enormously too.
Narrow AI (Weak AI)
AI designed for a specific task — spam filtering, chess, voice transcription. All AI in practical use today is narrow AI, even the impressive stuff. It contrasts with the hypothetical general AI below.
Artificial General Intelligence (AGI)
Hypothetical AI with human-level ability to understand, learn, and apply knowledge across any domain — the kind of flexible intelligence people have. AGI does not exist yet; whether and when it will is one of the most debated questions in technology.
Superintelligence
A hypothetical AI that surpasses the best human minds in essentially every domain. A staple of science fiction and long-term AI safety discussions, not something that exists today.
How Models Learn
Supervised Learning
Training where the model learns from labeled examples — inputs paired with correct answers. Show it ten thousand photos labeled "cat" or "dog," and it learns to classify new photos. The workhorse behind spam filters, medical image analysis, and recommendation systems.
Unsupervised Learning
Training on unlabeled data, where the model finds structure on its own — grouping similar customers together, for instance, without being told the groups in advance. Used for clustering, anomaly detection, and discovering hidden patterns.
Self-Supervised Learning
A clever middle path: the model creates its own labels from the data. A language model, for example, learns by predicting the next word in billions of sentences — the "label" is just the actual next word. This technique unlocked training on internet-scale data without human labeling, and it is the foundation of modern large language models.
Reinforcement Learning (RL)
Training where an AI agent learns by trial and error, receiving rewards for good actions and penalties for bad ones — like training a dog, but with math. Famous for game-playing AI (chess, Go, video games) and increasingly used to fine-tune language models based on human feedback.
Reinforcement Learning from Human Feedback (RLHF)
A technique where human reviewers rank AI outputs, and the model is trained to prefer highly-ranked responses. RLHF is a big reason modern chatbots are helpful, polite, and aligned with what users want — it was a key ingredient in making assistants like ChatGPT practical.
Fine-Tuning
Taking a pre-trained model and continuing to train it on a smaller, specialized dataset — adapting a general language model for medical Q&A, legal drafting, or your company's support tickets. Far cheaper than training from scratch.
Transfer Learning
The general principle behind fine-tuning: knowledge learned from one task transfers to related tasks. A model trained to recognize everyday objects can be adapted to recognize manufacturing defects with relatively little new data.
Overfitting
When a model memorizes its training data instead of learning general patterns — like a student who memorizes past exams but cannot handle new questions. Overfit models perform well on training data and poorly on new data. Preventing it is a central concern in machine learning practice.
Training Data
The data used to teach a model. Its quality, diversity, and biases shape everything the model does — which is why discussions of AI fairness and accuracy so often come back to training data.
Data Labeling / Annotation
The human work of tagging training data — marking tumors in scans, labeling sentiment in reviews, ranking chatbot responses. An enormous, often invisible labor force powers the AI boom through annotation.
Language Models and How They Work
Large Language Model (LLM)
An AI model trained on vast amounts of text to understand and generate human language. ChatGPT, Claude, and Gemini are LLMs. "Large" refers to their scale — billions of parameters trained on trillions of words.
Transformer
The neural network architecture behind modern LLMs, introduced in a 2017 research paper ("Attention Is All You Need"). Transformers process entire sequences of text at once and learn which words relate to which — the breakthrough that made today's language AI possible.
Attention Mechanism
The core trick of the transformer: when processing a word, the model "attends to" the other words most relevant to understanding it. In "the animal didn't cross the street because it was too tired," attention helps the model figure out that "it" refers to the animal. This context-sensitivity is what makes modern language AI so fluent.
Token
The unit of text a language model reads and writes — roughly a word fragment. "Unbelievable" might be two tokens ("un" + "believable"). Models have token limits (context windows) on how much text they can consider at once, and API pricing is usually per token.
Context Window
How much text a model can consider at one time — its working memory. Measured in tokens; modern models handle anywhere from thousands to millions. A larger context window lets a model work with whole books, long conversations, or large codebases without forgetting the beginning.
Prompt
The input you give an AI model — your question, instruction, or text. Everything about working effectively with AI starts with the prompt, which is why "prompt engineering" became a discipline.
Prompt Engineering
The practice of crafting effective inputs to get better AI outputs — being specific, providing context and examples, structuring requests. Less about trickery than about clear communication: the model can only work with what you tell it.
System Prompt / System Instructions
Hidden (or user-configurable) instructions that set a model's behavior for a whole conversation — its persona, rules, and constraints. Many AI products let you customize these to tailor the assistant to your needs.
Temperature
A setting controlling how random a model's outputs are. Low temperature gives focused, predictable answers (good for factual work); high temperature gives more creative, varied ones (good for brainstorming). If an AI product lets you adjust "creativity," it is usually adjusting temperature.
Hallucination
When a model generates false information with complete confidence — invented citations, wrong dates, fabricated facts. It happens because models predict plausible text, not verified truth. The single most important limitation for beginners to understand.
Zero-Shot / Few-Shot Learning
Using a model with no examples (zero-shot: just the instruction) or a handful of examples (few-shot: showing it 2–3 examples of the desired input-output pattern). Few-shot prompting is one of the simplest ways to improve results — show, don't just tell.
Chain-of-Thought Prompting
Asking a model to reason step by step ("think through this carefully before answering"). This measurably improves performance on math, logic, and multi-step problems — and letting you see the steps helps you catch errors.
Retrieval-Augmented Generation (RAG)
A technique where the model looks up relevant documents before answering, grounding its response in real sources. RAG powers many enterprise AI assistants and reduces hallucinations by giving the model actual facts to work with instead of relying solely on training memory.
Generative AI and Creativity
Generative AI
AI that creates new content — text, images, music, video, code — rather than just analyzing or classifying. LLMs, image generators, and voice synthesizers are all generative AI. The technology behind the current AI boom.
Diffusion Model
The technique behind modern AI image generators: the model learns to reverse a process of gradually adding noise to images, so starting from pure noise it can "denoise" its way to a coherent picture guided by your text prompt. Elegant, powerful, and the reason AI art exploded in quality around 2022.
Text-to-Image
AI that generates images from written descriptions — the category including tools like Midjourney, DALL-E, and Stable Diffusion. You describe; it depicts. Quality and controllability have improved dramatically year over year.
GAN (Generative Adversarial Network)
An older generative technique where two neural networks compete — one generating fakes, one detecting them — each improving until the fakes are convincing. GANs powered the first wave of deepfakes and AI art before diffusion models took over image generation.
Deepfake
AI-generated or AI-manipulated video, audio, or images depicting real people saying or doing things they never did. A dual-use technology: legitimate uses in film and accessibility, serious risks for fraud and misinformation.
Voice Cloning
AI that reproduces a specific person's voice from samples of their speech. Useful for audiobooks, dubbing, and assistive tech — and a growing vector for phone scams, which is why verification habits matter.
Style Transfer
Applying the visual style of one image to the content of another — your photo rendered as a watercolor, for instance. One of the earlier popular creative AI techniques.
AI Agents and Advanced Systems
AI Agent
An AI system that can take actions toward a goal — not just answering questions but using tools, browsing the web, running code, and completing multi-step tasks with some autonomy. The shift from chatbot to agent (from advisor to doer) is one of the biggest trends in applied AI.
Autonomous Agent
An agent designed to work toward goals with minimal human supervision, breaking tasks into subtasks and executing them. Powerful for research and automation; raises obvious questions about oversight and control.
Tool Use / Function Calling
The capability that lets a model invoke external tools — searching the web, querying a database, running calculations — rather than relying only on its training. This is what turns a language model into something that can actually do things in the world.
Multi-Agent System
Multiple AI agents working together — one researching, one writing, one reviewing, for example. An emerging pattern for complex tasks, modeled loosely on human teamwork.
Human-in-the-Loop
A design where humans review or approve AI actions at key points — the standard approach for keeping automated systems safe and accountable, from content moderation to medical AI.
Computer Vision and Perception
Computer Vision
AI that interprets visual information — recognizing objects in photos, reading license plates, detecting tumors in scans. The field behind facial recognition, self-driving perception, and quality inspection in factories.
Image Recognition / Classification
Identifying what is in an image ("this is a cat"). One of the earliest deep-learning triumphs and now a commodity capability.
Object Detection
Going beyond classification to locate objects within an image — drawing boxes around every car in a street scene. Essential for robotics, autonomous vehicles, and surveillance.
OCR (Optical Character Recognition)
Converting images of text — scanned documents, photos of signs — into machine-readable text. Often the first step before AI translation or document processing.
Facial Recognition
Identifying or verifying people from their faces. Technically mature, socially controversial — subject to bans and regulation in several jurisdictions over privacy and bias concerns.
Robotics and the Physical World
Robotics
The field of designing and building physical machines that sense and act in the world. Modern robotics increasingly pairs traditional engineering with AI for perception, navigation, and manipulation.
Embodied AI
AI that exists in a physical (or realistically simulated) body — robots that learn by interacting with environments rather than just processing text and images. Considered by many researchers a key path toward more capable, grounded intelligence.
AI Safety, Ethics, and Society
AI Alignment
The research problem of ensuring AI systems pursue goals aligned with human values and intentions — that they do what we actually want, not just what we literally asked for. One of the most important long-term challenges in the field.
Bias (in AI)
Systematic unfairness in AI outputs, usually inherited from training data — a hiring tool that disadvantages certain groups, a face recognizer that works worse on some demographics. Mitigating bias is an active area of research, policy, and (increasingly) law.
Explainability (XAI)
Techniques for understanding why an AI made a particular decision. Important for trust and accountability in high-stakes uses like lending, hiring, and medicine — though deep models remain stubbornly opaque.
AI Safety
The field concerned with preventing harm from AI systems — from near-term issues (misinformation, misuse) to long-term questions about advanced systems. Encompasses both technical research and governance.
Responsible AI
The practice (and corporate function) of developing and deploying AI with attention to fairness, transparency, privacy, and societal impact — ethics translated into engineering and policy checklists.
Red Teaming
Deliberately trying to break an AI system — eliciting harmful outputs, finding jailbreaks — before deployment, so vulnerabilities can be fixed. Borrowed from cybersecurity, now standard practice for major AI releases.
Jailbreak (AI)
A prompt or technique that bypasses an AI model's safety restrictions. A cat-and-mouse game between attackers and developers, and a reminder that no safeguard is perfect.
Model Collapse
A hypothesized degradation where models trained increasingly on AI-generated content get progressively worse — like a photocopy of a photocopy. A real concern as synthetic text floods the internet, and an active research topic.
Business and Practical Terms
API (Application Programming Interface)
The way developers connect to AI models programmatically — sending prompts and receiving answers from their own apps. Most AI products you use are built on model APIs behind the scenes.
Token Limit / Rate Limit
Practical constraints on API usage: how much text per request (token limit) and how many requests per minute (rate limit). The economics of building on AI revolve around these numbers.
Open-Source Model
A model whose weights are publicly released, letting anyone download, inspect, modify, and run it. The open ecosystem (models like Llama and Mistral) enables startups, researchers, and hobbyists to build without depending on big providers.
Closed / Proprietary Model
A model kept private and accessed only through a company's API or product. Typically the most capable frontier models, traded off against less transparency and vendor dependence.
Foundation Model
A large model trained on broad data that can be adapted to many downstream tasks — the "foundation" others build on. The term captures the platform-like role of today's biggest models.
Benchmark
A standardized test for measuring AI capabilities — question sets for math, reasoning, coding, and more. Benchmarks drive progress and marketing claims alike; treat vendor-reported scores as directional, not definitive.
Inference Cost
What it costs to run a model per query — the key economic number for AI products. Falling inference costs are quietly one of the most important trends in AI: capabilities that cost dollars last year cost cents today.
FAQ
What is the difference between AI and machine learning?
AI is the broad field of building intelligent machines; machine learning is the dominant approach within it — systems that learn patterns from data instead of following hand-written rules. Deep learning is then a subfield of machine learning using multi-layered neural networks. Think of them as nested circles: AI contains ML, which contains deep learning.
Do I need to know math to understand AI basics?
No. The conceptual ideas — learning from data, predicting the next word, adjusting toward fewer errors — are graspable without equations. Math matters for building models professionally, but for using AI well and understanding the news about it, plain-English concepts like the ones in this glossary are enough.
What does it mean when an AI "hallucinates"?
It means the model generated false information fluently and confidently — a fake citation, a wrong date, an invented fact. It happens because language models predict plausible-sounding text rather than checking facts. Always verify important claims from AI against independent sources.
What is an AI agent, in simple terms?
A chatbot answers your questions; an agent does things — searching the web, using software tools, completing multi-step tasks toward a goal you set. Agents combine a language model's reasoning with the ability to take actions, which makes them far more useful and also deserving of more oversight.
Are open-source AI models safe?
Open models offer transparency and independence — anyone can inspect and run them — which many consider a safety advantage. They also lower the barrier for misuse, since safeguards can be removed. Like most powerful technologies, the answer depends on the specific model, its safeguards, and who is using it for what.
What should I learn first if I want to work with AI?
Start with practical fluency: learn to prompt effectively, understand hallucinations and verification (see our How to Fact-Check AI Outputs: 2026 Guide), and get comfortable with one or two AI tools deeply — our What Are the Best AI Tools? 2026 Expert Roundup roundup is a good starting point. From there, branch by interest — data skills for analysis roles, API basics for builders, domain applications (marketing, education, healthcare) for specialists. Conceptual literacy like this AI glossary for beginners provides beats tool-specific tricks, because tools change constantly.
Conclusion: Your AI Glossary for Beginners Cheat Sheet
Sixty-plus terms, one core insight: this AI glossary for beginners proves AI jargon is far less intimidating once each word is anchored to a plain-English meaning. You now know the difference between training and inference, why hallucinations happen, what makes an agent more than a chatbot, and what all those debates about alignment and AGI are actually about. Keep this AI glossary for beginners bookmarked — the field invents new terms monthly, but nearly all of them build on the foundations above. And the next time someone drops "RAG pipeline with RLHF-tuned foundation models" in a meeting, you will know exactly what they mean — and whether they do.
Join the discussion
Have a question or a perspective to share?