Accuracy — How often an AI system gets the answer right.

Active learning — A training method where the AI asks humans to label the examples it is most unsure about.

Adversarial attack — A trick designed to make an AI system fail, such as slightly changing an image so the AI misidentifies it.

Agent — An AI system that can take steps toward a goal, such as searching, planning, using tools, writing files, or completing tasks.

Agentic AI — AI that does more than answer questions; it can act, plan, use tools, and pursue multi-step goals.

AGI (Artificial General Intelligence) — A hypothetical AI that could perform most intellectual tasks humans can do. This does not currently exist in any settled, agreed-upon sense.

AI (Artificial Intelligence) — Computer systems designed to do tasks that normally require human intelligence, such as recognizing images, understanding language, making predictions, or solving problems.

AI alignment — The effort to make AI systems behave in ways that match human values, intentions, and safety needs.

AI assistant — A conversational AI system that helps users write, search, summarize, code, plan, or reason.

AI audit — A formal review of an AI system to check its safety, fairness, accuracy, security, legality, or reliability.

AI benchmark — A test used to compare how well different AI systems perform.

AI ethics — The study and practice of making AI fair, safe, accountable, transparent, and beneficial.

AI governance — Rules, policies, processes, and oversight structures for controlling how AI is built and used.

AI hallucination — When an AI confidently gives information that is false, made up, or unsupported.

AI model — The trained system that produces predictions, answers, classifications, images, text, or other outputs.

AI policy — Laws, rules, institutional guidelines, or public policies about AI.

AI safety — Work focused on preventing AI from causing harm.

AI system — The full setup around an AI model, including data, software, user interface, rules, tools, monitoring, and human oversight.

Algorithm — A set of steps or instructions a computer follows to solve a problem.

Algorithmic bias — Unfair or skewed results caused by data, design choices, assumptions, or social inequalities built into a system.

Annotation — Adding labels or notes to data so an AI can learn from it.

API (Application Programming Interface) — A way for software programs to communicate with one another.

Artificial neural network — A computer system loosely inspired by the brain, made of connected layers that learn patterns from data.

Attention — A technique that helps an AI model decide which parts of the input matter most.

Automation — Having a machine or computer do a task with little or no human involvement.

Autonomous system — A system that can act on its own within some limits.

Backpropagation — A training method that helps a neural network adjust itself after making mistakes.

Base model — A general-purpose AI model before it has been specialized for a particular task or behavior.

Batch — A group of examples processed together during training.

Benchmark contamination — When an AI has seen the test questions during training, making its score look better than it really is.

Bias — A systematic tilt in data, model behavior, or outcomes. Bias can be statistical, social, cultural, or political.

Big data — Extremely large datasets that can be analyzed to find patterns.

Black box — A system whose inner workings are hard for people to understand, even if its outputs are visible.

Bot — A software program that performs tasks automatically. A chatbot is a bot designed for conversation.

Calibration — How well an AI’s confidence matches reality. A well-calibrated AI is more confident when it is likely right and less confident when it is likely wrong.

Chain of thought — Step-by-step reasoning used by a model internally or displayed externally. In many systems, the full private reasoning is not shown to users.

Chatbot — A program designed to communicate through conversation.

Classification — Sorting something into categories, such as spam/not spam or benign/malignant.

Classifier — An AI system that assigns labels or categories.

Cloud AI — AI that runs on remote servers rather than directly on your device.

Clustering — Grouping similar items together without already knowing the categories.

Code generation — AI writing computer code from a prompt or description.

Computer vision — AI that works with images or video, such as recognizing faces, reading scans, or detecting objects.

Confidence score — A number showing how sure the AI is about its answer.

Context — The information available to the AI when it produces an answer.

Context engineering — Designing the information given to an AI so it performs better.

Context length / context window — How much information an AI model can consider at once.

Conversational AI — AI built to communicate with people using natural language.

Copilot — A kind of AI assistant that works alongside a person, often in coding, writing, or office software.

Corpus — A collection of texts or other data used for training or analysis.

Curation — Selecting, organizing, cleaning, or preparing information for use.

Data — Information used by computers, such as text, numbers, images, audio, clicks, records, or sensor readings.

Data augmentation — Creating modified versions of data to help a model learn better.

Data drift — When real-world data changes over time, causing an AI system to become less accurate.

Data labeling — Adding correct answers or tags to examples so a model can learn.

Data leakage — When information from the test set or future data accidentally gets into training, making performance look falsely strong.

Data mining — Searching large datasets for patterns.

Dataset — A structured collection of data.

Decision boundary — The line or rule a model uses to separate one category from another.

Deep learning — Machine learning using large neural networks with many layers.

Deepfake — AI-generated or AI-altered media that makes someone appear to say or do something they did not.

Deployment — Putting an AI system into real-world use.

Diffusion model — A common type of image or media-generation model that learns to create outputs by reversing a process of adding noise.

Dimensionality — The number of features or measurements used to represent something.

Distillation — Training a smaller model to imitate a larger model.

Domain adaptation — Adjusting an AI system so it works better in a new setting or field.

Downstream task — A specific task a general model is later used for, such as summarization, translation, or diagnosis support.

Edge AI — AI that runs on a local device, such as a phone, camera, car, or sensor, rather than in the cloud.

Embedding — A way of turning words, images, or other data into numbers so a computer can compare meaning or similarity.

Emergent behavior — A surprising ability or pattern that appears in a large AI system even though it was not directly programmed.

Encoder — Part of a model that turns input into an internal representation.

End-to-end learning — Training one system to go from raw input to final output without many hand-built intermediate steps.

Ensemble — A system that combines several models to get better results.

Evaluation (eval) — Testing an AI system to see how well it works.

Explainability — The ability to explain why an AI system produced a particular result.

Expert system — An older style of AI that uses human-written rules to imitate expert decision-making.

False negative — When an AI says something is not present, but it actually is. Example: saying a disease is absent when it is present.

False positive — When an AI says something is present, but it is not. Example: flagging a harmless email as spam.

Feature — A piece of information used by a model, such as age, word frequency, pixel color, or purchase history.

Feature engineering — Choosing or designing the information a model uses to make predictions.

Few-shot learning — Giving a model a few examples in the prompt so it can imitate the pattern.

Fine-tuning — Further training a general model on special data so it performs better for a particular task, style, field, or behavior.

Foundation model — A large general-purpose model trained on broad data and adaptable to many tasks.

Frontier model — A highly capable, cutting-edge AI model near the current limits of performance.

GAN (Generative Adversarial Network) — A kind of AI system where two models compete: one creates fake examples, and the other tries to detect them.

Generalization — A model’s ability to perform well on new examples, not just the data it was trained on.

Generative AI — AI that creates new content, such as text, images, audio, video, code, or designs.

Generative model — A model that produces new outputs rather than only classifying or predicting labels.

GPU (Graphics Processing Unit) — A computer chip often used to train and run AI because it can do many calculations at once.

Gradient descent — A method for gradually adjusting a model to reduce its errors.

Ground truth — The correct answer used to train or test an AI system.

Grounding — Connecting an AI’s answer to reliable sources, facts, data, documents, or real-world evidence.

Guardrails — Rules or systems that try to keep AI behavior safe and appropriate.

Hallucination — A false or unsupported AI output that sounds plausible.

Human-in-the-loop — A system design where humans review, approve, correct, or supervise AI decisions.

Human-on-the-loop — Humans monitor the AI system from above, but do not approve every individual action.

Human-out-of-the-loop — An AI system acts without meaningful human review.

Hyperparameter — A setting chosen before or during training, such as learning rate, model size, or batch size.

Image generation — AI creating images from text, sketches, reference images, or other inputs.

Inference — Using a trained model to produce an answer, prediction, or output.

Instruction tuning — Training a model to follow human instructions more usefully.

Interpretability — Understanding what parts of a model or input led to a certain output.

Iteration — Repeatedly improving a system through cycles of testing and revision.

Jailbreak — A prompt or attack designed to bypass an AI system’s safety rules.

JSON mode — A setting that makes a model output structured data in JSON format, useful for software systems.

Knowledge base — A collection of information an AI system can use or search.

Knowledge cutoff — The latest date through which a model’s built-in training knowledge is likely to be current.

Knowledge graph — A structured map of entities and relationships, such as people, places, events, and concepts.

Label — The answer attached to a training example.

Labeled data — Data that includes correct answers or categories.

Language model — An AI model trained to understand, predict, or generate language.

LLM (Large Language Model) — A large AI model trained on massive amounts of text and often able to write, summarize, translate, code, and answer questions.

Latency — How long it takes a system to respond.

Learning rate — A setting that controls how quickly a model changes during training.

Local model — An AI model running on your own device or server rather than through a cloud provider.

Loss — A measure of how wrong a model is during training.

Machine learning (ML) — A branch of AI where systems learn patterns from data rather than being programmed with every rule.

Machine translation — AI translation from one language to another.

Metadata — Information about information, such as author, date, file type, location, or source.

Model card — A document describing what a model does, how it was trained, intended uses, limitations, and risks.

Model collapse — A possible problem where future AI models trained heavily on AI-generated material become worse or less diverse.

Model drift — When an AI system’s performance changes over time because the world, data, or user behavior changes.

Model weights — The internal numbers learned during training that shape how a model behaves.

Multimodal AI — AI that can work with more than one kind of input or output, such as text, images, audio, and video.

Natural language — Ordinary human language, such as English, Spanish, or Mandarin.

NLP (Natural Language Processing) — AI that works with human language.

Neural network — A model made of connected artificial neurons that learn patterns from data.

Noise — Randomness, errors, irrelevant information, or messy variation in data.

Non-deterministic — Not always giving the exact same answer every time.

Objective function — The goal a model is trained to optimize.

One-shot learning — Giving a model one example and asking it to follow the pattern.

Open model — A model whose code, weights, or design may be publicly available, depending on the meaning of “open.”

Open-source AI — AI software released with permissions that let others inspect, use, modify, or share it.

Optimization — The process of improving a model by reducing errors or improving performance.

Output — What the AI produces.

Overfitting — When a model memorizes training examples too closely and performs poorly on new examples.

Parameter — A number inside a model that helps determine its behavior.

Pattern recognition — Finding meaningful regularities in data.

Personalization — Adapting AI behavior to a specific user, organization, style, or context.

Pipeline — A sequence of steps that data or tasks move through.

Pretraining — The first large training stage where a model learns general patterns from broad data.

Precision — Of the items the AI marked as positive, how many were actually positive.

Prediction — An AI system’s estimate of what is likely to be true or happen.

Predictive AI — AI used to forecast or classify rather than generate new content.

Prompt — The instruction or input given to an AI system.

Prompt engineering — Writing or structuring prompts to get better AI outputs.

Prompt injection — An attack where hidden or malicious instructions try to override the user’s or developer’s intended instructions.

Provenance — The origin or history of data, content, or an AI output.

Quantization — A technique that makes models smaller or faster by using less precise numbers.

Query — A question or request sent to a search system, database, or AI system.

RAG (Retrieval-Augmented Generation) — A method where an AI searches documents or databases before answering, so its answer can be based on supplied sources rather than memory alone.

Recall — Of all the truly positive items, how many the AI successfully found.

Recommendation system — AI that suggests items, such as movies, products, songs, articles, or posts.

Red teaming — Testing an AI system by trying to make it fail or behave unsafely.

Regression — Predicting a number, such as price, risk, temperature, or demand.

Reinforcement learning — Training through rewards and penalties.

RLHF (Reinforcement Learning from Human Feedback) — Training a model to behave better using human preferences or ratings.

Representation — The internal way a model encodes information.

Responsible AI — Building and using AI in ways that are ethical, safe, accountable, fair, and legally compliant.

Retrieval — Finding relevant information from a database, document collection, or search index.

Robotics — The field of building machines that can sense, move, and act in the physical world.

Safety filter — A system that blocks or changes unsafe, disallowed, or inappropriate outputs.

Scalability — The ability of a system to keep working well as usage, data, or complexity grows.

Semantic search — Search based on meaning rather than exact keywords.

Semi-supervised learning — Training with a mix of labeled and unlabeled data.

Sentiment analysis — AI that estimates emotion, attitude, or tone in text.

SLM (Small Language Model) — A language model smaller than large frontier models, often cheaper and easier to run.

Speech recognition — AI that turns spoken language into text.

Synthetic data — Artificially generated data used for training, testing, or privacy protection.

System prompt — High-level instructions that shape how an AI assistant behaves.

Temperature — A setting that affects how random or creative a model’s output is. Lower temperature is usually more predictable; higher temperature is usually more varied.

Test set — Data held back to evaluate how well a model performs after training.

Text-to-image model — AI that creates images from written descriptions.

Token — A chunk of text processed by a language model. A token may be a word, part of a word, punctuation, or another text fragment.

Token limit — The maximum number of tokens a model can process or produce.

Tool use — When an AI model calls outside tools, such as search, calculators, calendars, databases, code interpreters, or file systems.

Training — The process by which a model learns patterns from data.

Training data — The information used to teach a model.

Transformer — A neural network architecture that made modern large language models possible by using attention to process relationships among tokens.

Transparency — Making enough information available about an AI system for people to understand, evaluate, or challenge it.

Turing Test — A famous proposed test of whether a machine can imitate human conversation well enough to be mistaken for human.

Uncertainty — How unsure a model or system is.

Underfitting — When a model is too simple or poorly trained to capture the patterns in the data.

Unlabeled data — Data without correct answers attached.

Unsupervised learning — Learning patterns from data without labeled answers.

User prompt — The specific message or request from the user.

Validation set — Data used during development to tune and compare models before final testing.

Vector — A list of numbers that represents something, such as a word, sentence, image, or user preference.

Vector database — A database designed to store and search embeddings.

Vector search — Searching by mathematical similarity between embeddings.

Vision-language model — A model that can work with both images and text.

Weights — The learned internal values of a model.

Workflow automation — Using AI and software tools to complete multi-step tasks.

World model — A model’s internal representation of how things relate or behave. In AI discussion, this term can be technical, speculative, or metaphorical depending on context.

XAI (Explainable AI) — Methods for making AI decisions more understandable to people.

Zero-shot learning — Asking a model to do a task without giving examples in the prompt.