A/B testing
A way to compare two versions of a page, message, prompt, or workflow with similar audiences to see which one performs better. Change one important thing at a time when possible.
Explore this term →PLAIN-ENGLISH AI GLOSSARY
AI vocabulary without the acronym asteroid field.
A practical field guide to 157 important AI terms, from first prompts to technical systems. Every definition is short, human-readable, and built to help you understand—not merely survive—a jargon duel.
HOW TO USE THIS CONSOLE
Type what you heard—even if you only remember half of it.
Search matches titles, acronyms, and related plain-English words. Try “fake video,” “AI privacy,” “voice,” “training,” “search,” or “how AI makes pictures.”
Fast safety ruleWhen a term affects money, hiring, health, legal matters, or customer data, pause and verify before acting.
A
Accuracy
How often an AI system gives correct results for a defined task. A model can be accurate on one task and weak on another.
Explore this term →Active learning
A training approach where a model asks for labels or feedback on the examples that would help it learn most.
Explore this term →Adversarial attack
An input deliberately designed to confuse or mislead an AI system, such as a specially altered image or a malicious instruction hidden in text.
Explore this term →Agent
An AI system designed to take steps toward a goal, often by using tools, following rules, or working through multiple tasks. Think “intern with a checklist,” not “magic robot employee.”
Explore this term →Agentic workflow
A workflow in which an AI can plan, use tools, check intermediate results, and continue through several steps toward a goal. Human guardrails still matter.
Explore this term →AGI
Artificial General Intelligence: a hypothetical AI with broad, flexible abilities across many kinds of tasks. It is not a settled benchmark or a product label you should accept uncritically.
Explore this term →AI alignment
The effort to make an AI system behave in ways that match intended goals, rules, and human values.
Explore this term →AI assistant
Software that uses AI to help with tasks such as drafting, summarizing, planning, searching, or analyzing information.
Explore this term →AI governance
The policies, roles, review processes, and controls used to manage how an organization selects, deploys, and monitors AI.
Explore this term →AI literacy
The practical ability to understand what AI can and cannot do, use it thoughtfully, and spot when human judgment is needed.
Explore this term →AI model
A mathematical system trained to recognize patterns and produce outputs such as text, images, predictions, or classifications.
Explore this term →AI safety
The practice of reducing the chance that AI causes harm through errors, misuse, privacy failures, bias, or poor controls.
Explore this term →Algorithm
A defined set of steps for solving a problem. AI systems use algorithms, but not every algorithm is AI.
Explore this term →Alignment
See AI alignment. In everyday terms, it means getting a system to follow the intended mission rather than merely optimizing the easiest shortcut.
Explore this term →Analytics
The process of examining data to understand patterns, trends, or performance. AI can assist with analytics but does not replace checking the underlying numbers.
Explore this term →Annotation
A label or note added to data so humans or models can understand what it represents, such as marking a photo as containing a dog.
Explore this term →API
Application Programming Interface: a structured way for software systems to talk to each other. It is the docking port between apps—not the spaceship itself.
Explore this term →API key
A secret credential used by software to authenticate to an API. Treat it like a password: do not paste it into public chat tools or commit it to public code.
Explore this term →Artificial intelligence
A broad field of computing focused on systems that perform tasks associated with learning, perception, language, reasoning, or decision support.
Explore this term →ASR
Automatic Speech Recognition: technology that turns spoken audio into text.
Explore this term →Attention mechanism
A technique that helps many modern language models weigh which parts of the input are most relevant while producing an output.
Explore this term →Attribute
A field or characteristic attached to a record, object, person, or item in data.
Explore this term →Audit trail
A record of actions, changes, prompts, approvals, or system events that helps people investigate what happened later.
Explore this term →Authentication
The process of proving who or what is requesting access, often with passwords, passkeys, tokens, or multi-factor authentication.
Explore this term →Automation
A repeatable process where software performs steps for you, such as moving form submissions into a spreadsheet or sending a follow-up after an appointment.
Explore this term →Autonomous system
A system that can operate with limited human input within a defined environment. “Autonomous” does not mean it is always correct or safe.
Explore this term →B
Backpropagation
A training method that adjusts a neural network after it compares its prediction with the correct answer.
Explore this term →Benchmark
A standardized test used to compare models or systems on specific tasks. A strong benchmark score is useful evidence, not a guarantee of real-world performance.
Explore this term →Bias
A systematic pattern that can produce unfair, distorted, or unrepresentative results. Bias can come from data, model design, prompts, or how people use a system.
Explore this term →Big data
Very large, fast-moving, or complex data that requires specialized storage, processing, or analysis methods.
Explore this term →Binary classification
A prediction task with two possible categories, such as spam versus not spam.
Explore this term →Black box
A system whose internal reasoning is hard to inspect or explain clearly. The phrase does not mean the system is mysterious magic; it means visibility is limited.
Explore this term →Bot
Software that performs automated actions or conversations. A bot may be simple rules, AI-powered, or both.
Explore this term →C
Chain of thought
A model’s internal step-by-step reasoning process. It can be useful to ask for a concise explanation or a checked solution, but do not treat a detailed-looking rationale as proof.
Explore this term →Chatbot
A conversational software interface that responds to messages. Some chatbots are AI-powered; some follow fixed scripts.
Explore this term →Classification
Assigning an item to one of several categories, such as labeling a support ticket by topic.
Explore this term →Classifier
A model or rule system that performs classification.
Explore this term →Cloud AI
AI services delivered through internet-hosted infrastructure rather than running entirely on your own device.
Explore this term →Computer vision
AI methods that analyze images or video to detect, describe, classify, or measure things.
Explore this term →Confabulation
Another word for an AI hallucination: an invented or unsupported claim delivered with apparent confidence.
Explore this term →Confidence score
A number expressing how strongly a model favors a prediction. It is not the same as a guarantee that the prediction is correct.
Explore this term →Content moderation
Rules and systems used to detect or restrict content that may violate safety, legal, or platform policies.
Explore this term →Context
The information an AI receives to help it answer or act, including instructions, conversation history, documents, examples, and tool results.
Explore this term →Context window
The amount of information an AI model can consider at one time in a conversation or task. More context can help, but it does not guarantee the answer is correct.
Explore this term →Copilot
A name often used for an AI assistant that helps a person work faster. It should suggest, draft, or organize—not silently take responsibility for important decisions.
Explore this term →Copyright
A legal right that can protect original creative works. AI use can raise copyright questions, especially when training, uploading, or publishing content; get qualified advice for high-stakes situations.
Explore this term →CRM
Customer Relationship Management software: a system for tracking customer interactions, leads, and follow-ups.
Explore this term →Cross-validation
A way to test a model by training and evaluating it on different slices of the same dataset.
Explore this term →D
Data augmentation
Creating modified versions of training examples to help a model handle more variation, such as rotating images or rephrasing text.
Explore this term →Data drift
A change over time in the real-world data a model sees, which can reduce performance after deployment.
Explore this term →Data governance
The rules and responsibilities for collecting, storing, using, sharing, protecting, and retaining data.
Explore this term →Data labeling
The process of adding useful tags or answers to data for training or evaluation.
Explore this term →Data leakage
Sensitive, private, or restricted data being exposed to people or systems that should not receive it.
Explore this term →Data lineage
A record of where data came from, how it changed, and where it was used.
Explore this term →Data minimization
A privacy practice of collecting and using only the data genuinely needed for a purpose.
Explore this term →Dataset
An organized collection of data used for analysis, training, testing, or reporting.
Explore this term →Decision tree
A model or rule structure that makes decisions through a series of branching questions.
Explore this term →Deepfake
Synthetic or altered media made to resemble a real person, voice, event, or recording. Treat unexpected high-pressure media as something to verify independently.
Explore this term →Deep learning
A branch of machine learning that uses multi-layer neural networks to learn complex patterns from data.
Explore this term →Deployment
Putting a model, application, or workflow into real use. Deployment should include monitoring, access controls, and a rollback plan.
Explore this term →Diffusion model
A generative model often used for images that learns to create content by gradually reversing a process of adding noise.
Explore this term →Document AI
AI that extracts, classifies, summarizes, or searches information in documents such as PDFs, forms, invoices, or contracts.
Explore this term →E
Embeddings
Numeric representations of text, images, or other data that capture useful similarities. They are often used for semantic search and RAG.
Explore this term →Emergent behavior
A capability or behavior that appears when a system becomes larger or is used in a new way, even if it was not directly programmed as a feature.
Explore this term →Endpoint
A specific address or function exposed by an API that software can call.
Explore this term →Entity extraction
Identifying things such as names, organizations, dates, places, or product codes in text.
Explore this term →Evaluation
A structured way to test whether an AI system performs well enough for a defined task and risk level.
Explore this term →Explainability
How well a person can understand why a system produced a result.
Explore this term →F
False negative
A result that incorrectly says something is absent or safe when it is actually present or risky.
Explore this term →False positive
A result that incorrectly flags something as present, suspicious, or likely when it is not.
Explore this term →Feature
An input variable or measurable property used by a model, such as purchase amount, time of day, or word count.
Explore this term →Feature engineering
Selecting, transforming, or creating useful input features for a model.
Explore this term →Fine-tuning
Additional training that adjusts a base model for a narrower style, task, or domain. It is different from simply giving an AI a detailed prompt.
Explore this term →Foundation model
A large, broadly trained model that can be adapted to many tasks through prompting, fine-tuning, or tools.
Explore this term →G
Generalization
A model’s ability to work on new examples that were not part of its training data.
Explore this term →Generative adversarial network
A model architecture in which one network generates examples and another tries to detect them, historically used for synthetic media.
Explore this term →Generative AI
AI that creates new content such as text, images, audio, video, software code, or summaries.
Explore this term →GPU
Graphics Processing Unit: hardware that can perform many calculations in parallel and is widely used for AI training and inference.
Explore this term →Guardrail
A technical, procedural, or policy control that limits unsafe, unauthorized, or unwanted AI behavior.
Explore this term →H
Hallucination
When an AI produces information that sounds believable but is wrong, unsupported, or invented. Treat high-stakes answers as drafts requiring verification.
Explore this term →Human in the loop
A design where a person reviews, approves, corrects, or can stop important AI outputs or actions.
Explore this term →Hyperparameter
A setting chosen by people before or during model training, such as learning rate, batch size, or number of layers.
Explore this term →I
Inference
Using a trained model to produce an answer, prediction, classification, or generated content.
Explore this term →Instruction hierarchy
The order of importance among instructions given to an AI system, such as system rules, developer rules, and user requests.
Explore this term →Intellectual property
Creations of the mind that may be protected by laws such as copyright, trademark, patent, or trade-secret law.
Explore this term →Intent
The goal or meaning behind a user request, such as “cancel an order” or “compare tools.”
Explore this term →J
Jailbreak
An attempt to bypass a model’s intended safety rules or restrictions through clever instructions or other inputs.
Explore this term →K
Knowledge base
An organized collection of approved information used to answer questions or support work.
Explore this term →L
Label
A category, answer, or tag attached to a data example.
Explore this term →Large language model (LLM)
A type of AI trained on enormous amounts of text to predict and generate language. Chat-based assistants are commonly built on this kind of model.
Explore this term →Latency
The time it takes for a system to respond.
Explore this term →Learning rate
A training setting that controls how aggressively a model changes after each round of feedback.
Explore this term →M
Machine learning
A branch of AI in which systems learn patterns from data rather than relying only on hand-written rules.
Explore this term →Machine translation
Automatic translation from one language to another. It can be useful, but important legal, medical, or cultural nuance still needs review.
Explore this term →Metadata
Data that describes other data, such as a file’s author, date created, source, or access permissions.
Explore this term →Model
The underlying AI system that has learned patterns from data. A tool or app may use one model, several models, or change models over time.
Explore this term →Model card
A document that describes a model’s intended use, limitations, evaluation results, and important risks.
Explore this term →Model drift
A decline or change in model performance over time, often caused by data drift, changing user behavior, or a changing environment.
Explore this term →Model weights
The learned numerical values inside many AI models that encode patterns from training.
Explore this term →Multimodal
A system that can work with more than one kind of input or output, such as text, images, audio, video, or files.
Explore this term →N
Named entity recognition
A task that identifies named items in text, such as people, companies, locations, dates, or products.
Explore this term →Natural language processing
AI techniques for working with human language, including classification, extraction, translation, summarization, and generation.
Explore this term →Neural network
A machine-learning model made of connected layers that can learn complex patterns from examples.
Explore this term →No-code
Building software or automations mainly through visual tools instead of writing traditional code.
Explore this term →O
OCR
Optical Character Recognition: technology that reads printed or handwritten text from images or documents.
Explore this term →On-device AI
AI that runs mainly on a phone, computer, or other local hardware rather than sending every request to a cloud service.
Explore this term →Open source
Software whose source code is made available under a license that allows inspection and, depending on the license, reuse or modification.
Explore this term →Optimization
Improving a system toward a defined objective, such as lower error, lower cost, faster response, or higher user satisfaction.
Explore this term →Overfitting
When a model learns its training examples too specifically and performs poorly on new, real-world examples.
Explore this term →P
Parameter
A learned numerical value inside a model. People often use parameter count as a rough size measure, but bigger is not automatically better for every job.
Explore this term →Personal data
Information that identifies or could reasonably be linked to a person. Rules vary by location and context; protect it carefully.
Explore this term →PII
Personally Identifiable Information: data that can identify or help identify a person, such as a full name combined with a contact detail or government identifier.
Explore this term →Plugin
An add-on that gives an application extra features or connects it to another service.
Explore this term →Prediction
A model output estimating a likely category, value, event, or next item.
Explore this term →Prompt
The instructions, question, context, and examples you give an AI. Better prompts make the task clearer; they do not turn an unreliable claim into a fact.
Explore this term →Prompt chaining
Breaking a larger task into several prompts or model calls where each step feeds the next.
Explore this term →Prompt engineering
Designing prompts, examples, constraints, and output formats so an AI can complete a task more reliably.
Explore this term →Prompt injection
A malicious or untrusted instruction intended to make an AI ignore its rules, reveal data, or take an unsafe action.
Explore this term →Provenance
Information about where data, content, or an output came from and how it was produced.
Explore this term →R
RAG
Retrieval-Augmented Generation: a setup where an AI looks up relevant documents or data before answering. It can improve grounding, but you still need to check the source material.
Explore this term →Ranking
Putting items in order by a score or predicted relevance, such as ordering search results.
Explore this term →Recall
For classification, the share of truly relevant items that the system successfully finds.
Explore this term →Recommendation system
A system that suggests products, videos, articles, or actions based on patterns in data.
Explore this term →Red teaming
Authorized testing that tries to find weaknesses, unsafe behavior, or bypasses before a system causes harm in the real world.
Explore this term →Regression
A prediction task where the output is a number, such as estimated cost, temperature, or sales amount.
Explore this term →Reinforcement learning
A training approach where a system learns by receiving rewards or penalties for actions or outcomes.
Explore this term →Responsible AI
A practical approach to building and using AI with attention to safety, fairness, privacy, accountability, and transparency.
Explore this term →Retrieval
Finding relevant information from a collection of documents, records, or data.
Explore this term →RLHF
Reinforcement Learning from Human Feedback: a method that uses human preferences or ratings to help guide model behavior.
Explore this term →Robotic process automation
Software that automates repetitive, rule-based business tasks by interacting with applications much like a person would.
Explore this term →S
Safety filter
A system that detects or limits certain unsafe, harmful, or policy-violating inputs or outputs.
Explore this term →Sandbox
An isolated environment used to run code, tools, or experiments with reduced risk to other systems or data.
Explore this term →Semantic search
Search that tries to match meaning, not only exact words.
Explore this term →Sentiment analysis
Estimating the emotional tone or opinion expressed in text, such as positive, negative, or neutral.
Explore this term →SLM
Small Language Model: a language model designed to be lighter, cheaper, faster, or easier to run locally than a very large model.
Explore this term →Speech synthesis
Technology that converts text into spoken audio.
Explore this term →Structured data
Information organized in consistent fields or rows, such as a spreadsheet or database table.
Explore this term →Supervised learning
Training a model with examples that include known answers or labels.
Explore this term →Synthetic data
Artificially generated data used for testing, training, or simulation. It can help with privacy or scarcity, but it can also carry unrealistic patterns or bias.
Explore this term →System prompt
High-priority instructions that define an AI assistant’s overall role, boundaries, and behavior.
Explore this term →T
Temperature
A setting that influences how varied or predictable some generative-model outputs are. Higher temperature often increases variety and risk of wandering; lower temperature is usually more consistent.
Explore this term →Text-to-image
AI that creates images from written instructions.
Explore this term →Token
A small chunk of text that an AI model processes. A token might be a word, part of a word, punctuation, or a short sequence of characters.
Explore this term →Tool calling
A model capability that lets an AI request a defined action from external software, such as looking up a calendar or running a calculation.
Explore this term →Training data
The material used to help a model learn patterns. Training data influences what a model can do, but does not make it a reliable source for current or private facts.
Explore this term →Transformer
A neural-network architecture that uses attention mechanisms and powers many modern language and multimodal models.
Explore this term →TTS
Text-to-Speech: technology that turns written text into spoken audio.
Explore this term →U
Unstructured data
Information that does not live neatly in fixed rows and columns, such as documents, emails, images, audio, or video.
Explore this term →Unsupervised learning
Training methods that look for patterns or groups in data without supplied answer labels.
Explore this term →User interface
The screens, controls, and interactions through which a person uses software.
Explore this term →V
Vector database
A database designed to store and search embeddings, often used for semantic search and RAG.
Explore this term →Vector search
Searching for items with similar embeddings, which can help find related meaning even when wording differs.
Explore this term →Version control
A system for tracking changes to files and code over time, making collaboration and rollback easier.
Explore this term →Voice cloning
Creating synthetic speech that imitates a person’s voice. Treat unexpected audio requests involving money, codes, or urgent action as something to verify through another channel.
Explore this term →W
Workflow
A repeatable series of steps that moves work from a trigger to a result.
Explore this term →Z
Zero-shot prompting
Asking a model to perform a task without providing examples in the prompt.
Explore this term →Still foggy?
Ask the AI to explain a term in your own language.
Try: “Explain [TERM] to me as if I am smart but new to AI. Use one everyday analogy, one example, and one thing people commonly misunderstand.”
DOWNLOADABLE FIELD GUIDES
Take the glossary with you.
Compact cheat sheets for quick reference, training sessions, and those moments when someone casually drops “vector embedding” into a meeting.
Article details