Free tools Windows power users keep installed
One-click scans. No signup required.
AI terminology gets confusing because people use one label for several different things: a field of research, a kind of model, a training method, or a way to connect a model to outside information. This glossary separates those ideas into 63 plain-language definitions. The selection is a practical guide, not a canonical list; terms and product labels can vary across organizations.
AI foundations
1. Artificial intelligence (AI)
A broad field focused on building computer systems that perform tasks associated with human intelligence, such as recognizing patterns, understanding language, or making decisions. Machine learning is one approach within AI, not another name for the entire field.
As an Amazon Associate I earn from qualifying purchases.
2. Machine learning (ML)
A way to build AI systems by having them learn patterns from data rather than relying only on hand-written rules. A system that learns to classify incoming messages as spam is using ML; it need not generate new content.
3. Deep learning
A branch of machine learning that uses neural networks with many layers to learn complex patterns. Deep learning is behind many modern image, speech, and language systems.
#1 Best Overall
4. Neural network
A machine-learning model made of connected computational units that transform input data. The name is inspired by brains, but a neural network is a mathematical system, not a digital human mind.
5. Algorithm
A defined procedure for carrying out a computation or solving a problem. An algorithm can be part of an AI system, but the word does not by itself mean that the system learns or is intelligent.
6. Dataset
A collection of data used to train, test, or operate a model. A dataset might contain labeled photographs, documents, or records of past transactions.
7. Label
The answer or category attached to an example in supervised learning. In a set of product photos, labels might identify each image as “shoe,” “bag,” or “hat.”
8. Feature
An input attribute a model uses to make a prediction, such as a transaction amount or an image’s pixel values. In some modern systems, useful features are learned automatically rather than selected by a person.
9. Supervised learning
Training with examples paired with known answers. A model can learn to recognize damaged parts from images labeled “damaged” or “undamaged.”
10. Unsupervised learning
Learning patterns from data that has not been assigned answer labels. For example, a system may group similar customer records without being told the groups in advance.
11. Reinforcement learning
Training in which a system takes actions and receives rewards or penalties, learning to improve its choices over time. It is often used for sequential decisions, such as selecting moves in a game.
12. Classification
A task that assigns an input to one or more categories. Marking a message as spam or not spam is classification; generating a new reply is a different task.
13. Prediction
An estimate of an unknown or future value based on available information. A model might predict delivery time from route and traffic data; a prediction is not necessarily a generated explanation.
14. Generative AI
AI that produces new content, such as text, images, audio, or code, in response to an input. This distinguishes it from systems whose main job is to classify or predict a value, although one product may combine both kinds of capability.
Models, data, and learning
15. Model
A computational system that has learned patterns from data or has been configured to carry out a task. A trained model can take an input and return an output, such as a category, a forecast, or a generated passage.
Rank #2
16. Foundation model
A model trained on broad data that can be adapted or prompted for a range of tasks. Foundation models can handle different modalities; the term is broader than large language model.
17. Large language model (LLM)
A model trained to process and generate language, typically by learning patterns in text. An LLM is text-focused, while a foundation model may also work with images, audio, or other data.
18. Multimodal model
A model that can process or generate more than one kind of data, such as text and images. For example, a multimodal model may answer a question about a picture sent with the prompt.
Recommended Free Tools
19. Training
The process of adjusting a model using data so that it learns patterns or improves at a task. Training is different from inference, when a trained model is used to produce an output.
20. Inference
The use of a trained model to make a prediction or generate an output from an input. Asking a deployed language model to draft a paragraph is inference, not a new training run.
21. Parameter
A value inside a model that is adjusted during training and influences its outputs. A parameter count describes part of a model’s structure, but by itself does not establish how well the model performs.
22. Pretraining
An initial training stage in which a model learns broad patterns from a large collection of data. A language model may learn how words and passages relate before it is adapted for a particular use.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 1123. Fine-tuning
Further training of an existing model on a more specific dataset or task. Fine-tuning a general model on carefully prepared support examples can adapt its behavior, but it is distinct from simply giving the model new information in a prompt.
24. Transfer learning
Reusing knowledge learned for one task or dataset to help with another. Starting with a pretrained image model and adapting it to identify a new set of objects is an example.
25. Overfitting
When a model learns training examples too closely and performs poorly on new, unseen data. A model that memorizes its practice questions but fails on different questions is overfit.
26. Training data
The data used to adjust a model during training. Its contents and quality can affect what patterns a model learns, but a model’s answer does not automatically reveal which specific training example it used.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →27. Synthetic data
Artificially generated data used in place of, or alongside, collected real-world examples. It can help create additional training examples, but generated data may reproduce errors or biases and still needs evaluation.
28. Data leakage
A problem in which information that should be unavailable during training or evaluation accidentally influences the result. If a test set’s answers leak into training, a high score may not reflect performance on genuinely new cases.
Working with language models
29. Prompt
The input or instruction given to a generative model. “Summarize this report in three bullet points” is a prompt; the report itself may also be included as prompt context.
30. Prompt engineering
The practice of shaping instructions and context to help a model produce a useful response. Specifying the audience, format, and constraints can make a request clearer, though it cannot guarantee a correct answer.
31. System instruction
An instruction supplied to guide a model’s behavior across a conversation or task, such as a requirement to answer in a particular style. It is distinct from the user’s immediate request, though exact handling varies by system.
32. Token
A unit a model processes, often a word, part of a word, punctuation, or another text fragment. A token is not necessarily a whole word: an uncommon word may be split into multiple tokens.
33. Tokenization
The process of dividing input into tokens a model can handle. Different models can tokenize the same text differently, so a token count is not the same as a word count.
34. Context window
The amount of input and conversation content a model can consider at one time, measured in tokens. It is a working limit for a given interaction, not automatically a permanent memory of everything a user has said.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
35. Temperature
A generation setting that affects how varied or predictable a model’s output is. A lower setting generally favors less variation, while a higher one can produce more varied choices; it does not make an answer more factually reliable.
36. Embedding
A numerical representation of data that captures patterns or relationships in a form useful to computational systems. Search systems can use embeddings to find passages related in meaning, even when they do not share the exact same words.
37. Vector
An ordered list of numbers. An embedding is commonly represented as a vector, which can be compared with other vectors to estimate how related their underlying items are.
38. Vector database
A database designed to store and search vectors, often to retrieve items with similar embeddings. A document system can use one to find passages that are semantically related to a query.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →39. Semantic search
Search that aims to find results related to the meaning of a query, rather than relying only on exact word matches. Searching “ways to reduce home heating costs” may find material about insulation even if the query’s words are absent.
40. Memory
Information a system retains or retrieves across interactions, if that capability is implemented. Memory is different from a context window: the window is what the model can consider in one interaction, while persistent memory requires a system to store and later supply information.
41. Fine-grained instruction following
A model’s ability to respond to multiple constraints in a request, such as a length limit and a required format. A request for a short, friendly answer in a table tests instruction following, not factual knowledge alone.
Retrieval, tools, and AI agents
42. Retrieval
Finding relevant information in a source such as a document collection, database, or search index. Retrieval selects material; it is not the same as having a generative model write an answer from that material.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minute43. Retrieval-augmented generation (RAG)
A method that retrieves relevant information and adds it to a language model’s prompt before the model generates a response. In a document assistant, the system can retrieve passages from the collection and provide them as context. RAG can ground a response in relevant material, but it does not guarantee that the response is correct. Google Cloud’s RAG overview describes this retrieval-then-generation workflow.
44. Grounding
Connecting a model’s response to information supplied from a source, such as retrieved documents. Grounding can make an answer more relevant to that material, but it is not proof that the answer faithfully represents it or is true.
45. Tool calling
A model’s request for an external system to perform an action or return information, such as checking a database or running a calculation. The tool performs the operation; the model can then use the result in its response.
46. Function calling
A structured form of tool calling in which a model supplies arguments for a defined function, often in a format software can validate. For example, it might provide a city name to a weather function; this is different from executing arbitrary code.
47. AI agent
A system that uses a model to pursue a task by choosing actions, often including calling tools and responding to results. An agent may take several steps, but the term alone does not establish how autonomous, reliable, or safe it is.
48. Workflow
A sequence of steps that moves a task through one or more systems. A workflow may include AI, but a fixed sequence—retrieve a document, summarize it, save the result—does not necessarily make the system an agent.
49. Orchestration
The coordination of models, tools, data sources, and steps in an AI application. An orchestration layer might route a question to search, pass results to a model, and format the final response.
50. API
An application programming interface: a defined way for software systems to request services or exchange data. An AI application may use an API to send a prompt to a model or to call an external tool.
Reliability, evaluation, and responsible AI
51. Hallucination
A plausible-sounding model output that is unsupported or incorrect. A fluent answer can still contain invented details, so style is not evidence of truth.
Best Value
52. Evaluation
The process of assessing a model or system against defined criteria, such as accuracy, usefulness, or safety. Results depend on the test data and method; a score on one evaluation does not prove that a system will perform equally well in every setting.
53. Benchmark
A defined dataset or set of tasks used to compare or measure systems. Benchmarks can be useful for a specific capability, but they may not reflect the conditions or priorities of a real deployment.
54. Bias
A systematic skew in data, model behavior, or outcomes that can disadvantage some people or groups. Checking only average performance may miss important differences across groups or contexts.
55. Fairness
A goal of treating people or groups equitably in a particular application. Fairness has multiple definitions and can involve trade-offs, so it should be assessed in relation to the system’s use and affected people.
56. Explainability
The extent to which people can understand why a system produced an output. A short explanation may be useful, but it is not automatically a faithful account of a complex model’s internal process.
57. Interpretability
How readily a person can understand a model’s internal operation or behavior. It is related to explainability, but often refers more directly to how understandable the model itself is.
58. Transparency
The availability of information about a system, such as its purpose, limitations, data practices, or evaluation. Transparency can help people assess a system, but disclosure alone does not ensure safe or fair outcomes.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →59. Privacy
Protection and appropriate handling of information about people. AI systems can raise privacy concerns when they collect, retain, infer from, or expose personal data.
60. Robustness
The ability of a system to keep working reliably when inputs or conditions change, including cases that differ from expected examples. A model robust to ordinary spelling mistakes may still fail on unusual or adversarial inputs.
61. Security
Protection of an AI system, its data, and connected tools against misuse or attack. Security concerns can include unauthorized access, manipulated inputs, and unsafe tool actions.
62. Safety
Reducing the risk that a system causes harm in its intended and foreseeable uses. Safety depends on the application and deployment context, not only on the model’s output in isolation.
Free tools Windows power users keep installed
One-click scans. No signup required.
63. AI governance
The policies, roles, processes, and oversight used to manage AI systems and their risks. Governance can include deciding when a system may be used, monitoring outcomes, and assigning responsibility for decisions.
Terms that are easy to mix up
- AI and machine learning: AI is the broad field; machine learning is one way to build AI systems by learning patterns from data.
- Generative AI and prediction: Generative AI creates content; a predictive or classification system estimates a value or assigns a category. Some applications combine these tasks.
- LLM and foundation model: An LLM is focused on language. A foundation model is a broader category that can include models built for multiple modalities.
- Token and word: Tokens are model-processing units, not a dependable one-token-per-word count.
- Context window and memory: The context window is the information available to the model at one time. Persistent memory depends on separate storage and retrieval features.
- Retrieval and generation: Retrieval finds relevant material; generation produces an output. RAG connects them by supplying retrieved information to a model before it responds.
- Grounding and truth: Grounding connects an answer to supplied sources; it does not guarantee that the answer is accurate.
- Training and inference: Training adjusts a model using data; inference uses the resulting model to produce an output.
Further reading
For terminology beyond this practical selection, see Google Cloud’s Generative AI glossary. For responsible-AI vocabulary, NIST’s The Language of Trustworthy AI: An In-Depth Glossary of Terms is a reference intended for use with the NIST AI Risk Management Framework or on its own. As with any glossary, consult current guidance when a precise definition matters because terminology and system capabilities change.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
