AI Glossary: 50 Terms Every Beginner Should Know

This glossary defines 50 essential AI and machine learning terms in plain language. Each term has a brief, beginner-friendly explanation.

A

Algorithm

A step-by-step procedure or formula for solving a problem or completing a task. In AI, algorithms are instructions that tell computers how to process data and make decisions.

API (Application Programming Interface)

A set of rules allowing different software programs to communicate. APIs let applications request data or services from each other. For example, a weather app uses an API to get weather data from a weather service.

Artificial General Intelligence (AGI)

A hypothetical AI system with human-level intelligence across all domains. Unlike today's AI (which specializes), AGI would reason, learn, and adapt as well as humans in any field. We don't have AGI yet.

Bias

Systematic error in AI systems, often reflecting biases in training data. If a model trained on biased data, it will make biased predictions. Reducing bias is critical for fair AI.

B

Bias (Continued above)

Chatbot

A program designed to simulate conversation. Chatbots use AI to understand questions and generate responses. They range from simple rule-based bots to advanced language model-based systems like ChatGPT.

C

Classification

A machine learning task where the goal is to assign items to predefined categories. Example: Email spam filter classifies messages as spam or not spam. The system learned from examples to make predictions.

Cloud Computing

Using servers accessed over the internet instead of local hardware. Cloud services host AI models and data. Companies like AWS, Google Cloud, and Azure provide cloud infrastructure for AI applications.

Computer Vision

AI systems that interpret images and videos. Computer vision can identify objects, recognize faces, read text, and detect anomalies. Powers autonomous vehicles, medical imaging, and photo search.

D

Data Mining

Extracting patterns and insights from large datasets. Data mining discovers relationships, anomalies, and trends that might not be obvious. Foundational to machine learning and business intelligence.

Dataset

A collection of data used to train, test, or evaluate AI models. Datasets are fundamental: their quality determines model quality. Bad data leads to bad models, even with perfect algorithms.

Deep Learning

Machine learning using deep neural networks (many layers). Deep learning powers modern AI breakthroughs: image recognition, language models, game-playing AI. Named for the depth of network layers.

Deepfake

Synthetic media (video, audio) created using AI to manipulate or create false content. Deepfakes use deep learning to replace faces or generate realistic fake videos. Major concern for misinformation and fraud.

Diffusion Model

A type of generative AI that creates images by gradually removing noise. Systems like Stable Diffusion and DALL-E use diffusion models to generate images from text descriptions. Powerful but computationally intensive.

E

Embedding

Representing words, images, or concepts as numerical vectors. Embeddings capture semantic meaning: similar words have similar embeddings. Critical for how modern AI systems understand and compare information.

Epoch

One complete pass through an entire training dataset. During training, models see data many epochs (usually 10-100). Each epoch helps the model improve, but too many can cause overfitting.

F

Fine-tuning

Adjusting a pre-trained model on new, specific data. Rather than training from scratch, you start with a model trained on general data and adjust it for your specific task. Much faster and requires less data.

Foundation Model

A large, general-purpose AI model trained on vast amounts of data. Foundation models (like GPT-4 or LLaMA) can be fine-tuned for specific tasks. The base upon which many AI applications build.

G

GAN (Generative Adversarial Network)

Two neural networks competing: one generates fake data, one tries to detect fakes. This competition improves both. GANs create realistic images, deepfakes, and synthetic data.

GPT (Generative Pre-trained Transformer)

A type of large language model. GPT systems (like ChatGPT, GPT-4) are trained on massive text corpora to predict the next word. Pre-trained means they're already trained; you can use them directly or fine-tune them.

GPU (Graphics Processing Unit)

A processor designed for graphics but excellent for AI training. GPUs parallelize computation, making them 10-100x faster than CPUs for neural network training. Why AI research requires expensive hardware.

H

Hallucination

When AI systems generate false information confidently. Large language models sometimes "hallucinate" facts that aren't true but sound plausible. A major limitation to watch for.

Hyperparameter

Settings you choose before training (learning rate, batch size, number of layers). Unlike parameters learned during training, hyperparameters are manually set. Choosing good hyperparameters matters for model performance.

I

Image Recognition

AI systems identifying objects, people, or concepts in images. Image recognition powers facial recognition, medical diagnosis, autonomous vehicles. Deep learning made this practical.

Inference

Using a trained model to make predictions on new data. Unlike training (where models learn), inference is where trained models are used in production. Inference is usually faster and cheaper than training.

L

Large Language Model (LLM)

A neural network trained on massive amounts of text to understand and generate language. LLMs like ChatGPT, Claude, and Gemini can answer questions, write text, and hold conversations. Changed how people access AI.

M

Machine Learning

AI systems that improve through experience and data rather than explicit programming. Instead of coding every rule, you provide examples and let algorithms learn patterns. Foundational to modern AI.

Model

A trained mathematical system that makes predictions. Models are trained on data, then used to make predictions on new data. A model might predict house prices, recognize faces, or generate text.

N

Natural Language Processing (NLP)

AI systems understanding and generating human language. NLP powers chatbots, translation, sentiment analysis, and text search. One of the most successful AI domains.

Neural Network

AI systems inspired by biological brains, with interconnected nodes (neurons). Neural networks learn by adjusting connection weights. Deeper networks (more layers) can learn complex patterns.

O

Open Source

Software with publicly available code that anyone can view, modify, and distribute. Open source AI (like Llama, Stable Diffusion) democratizes AI. Community-driven development.

Overfitting

When a model memorizes training data instead of learning generalizable patterns. An overfitted model performs great on training data but poorly on new data. Avoiding overfitting is key to good models.

P

Parameter

A variable a model learns during training (like weights and biases). A neural network with millions of parameters learns complex patterns. More parameters don't always mean better models, can cause overfitting.

Prompt

Instructions given to an AI system (usually text) to elicit desired output. "Write a poem about cats" is a prompt. Quality prompts produce quality outputs. Prompt engineering (crafting good prompts) is a skill.

Prompt Engineering

Crafting effective prompts to get the best responses from AI systems. Techniques include being specific, providing examples, and asking for step-by-step reasoning. Small prompt changes can dramatically improve results.

R

RAG (Retrieval-Augmented Generation)

Combining retrieval (finding relevant documents) with generation (producing text). RAG lets AI systems reference external knowledge instead of relying purely on training data. Reduces hallucinations.

Reinforcement Learning

Training AI systems through rewards and penalties. The system learns to maximize rewards. Powers game-playing AI (AlphaGo), robotics, and autonomous driving. Different from supervised learning.

S

Robotics

Physical systems with AI for autonomous control. Robots combine computer vision, motor control, and decision-making. Applications range from manufacturing to exploration to home assistance.

Sentiment Analysis

Determining emotional tone in text (positive, negative, neutral). AI systems analyze reviews, social media, or customer feedback to understand sentiment. Useful for brand monitoring and user research.

Stable Diffusion

An open-source image generation model. Creates images from text descriptions using diffusion techniques. More accessible than proprietary models because it can run locally.

Supervised Learning

Training AI on labeled examples (input-output pairs). You provide examples with correct answers; the model learns to predict answers for new inputs. Most common machine learning approach.

T

Token

A unit of text (usually a word or subword) that AI models process. Large language models work with tokens, not characters. Understanding token limits helps with prompt design.

Training Data

Data used to teach AI models. The quantity and quality of training data determines model quality. Biased training data creates biased models. "Garbage in, garbage out" applies strongly to AI.

Transfer Learning

Using a model trained on one task as a starting point for another. Transfer learning lets you use knowledge from large datasets for new tasks with limited data. Dramatically speeds up development.

Transformer

An AI architecture using attention mechanisms to process sequences. Transformers (developed by Google in 2017) power modern large language models. Revolutionary for NLP.

Turing Test

A test of artificial intelligence proposed by Alan Turing: Can a human judge distinguish between AI and human conversation? Controversial as a measure of intelligence but historically important.

U

Unsupervised Learning

Training AI on unlabeled data to discover hidden patterns. No correct answers provided; the model finds structure. Used for clustering, dimensionality reduction, and anomaly detection.

V

Vector Database

A database optimized for storing and searching embeddings (numerical vectors). Vector databases enable semantic search: finding similar items by meaning, not keywords. Growing importance for AI applications.

Weight

Parameters in neural networks connecting neurons. During training, weights are adjusted. Millions of tiny weight adjustments create learning. Similar to how human brain synapses strengthen with use.

Z

Zero-shot Learning

An AI system solving tasks without prior training examples. Large language models do zero-shot learning: they can answer questions about topics they weren't explicitly trained on. Powerful but not perfect.

Want to dive deeper? Explore our other articles: Will AI Replace My Job?, Quantum Computing, or APIs for practical context on how these terms apply.