The AI glossary.
Thirty-five terms that make AI coverage easier to evaluate. The definitions focus on what each idea means in practice—and what it does not prove.
Agent
An AI system that can plan steps and use tools to pursue a goal. An agent still needs boundaries, permissions, monitoring and a clear point for human review.
Algorithm
A defined procedure for turning inputs into outputs. Machine-learning algorithms learn some of their behaviour from data instead of relying only on rules written by people.
Alignment
Work intended to make an AI system behave consistently with human intentions and constraints. Alignment is not a one-time setting and does not guarantee error-free behaviour.
API
An application programming interface: a documented way for software systems to exchange requests and results. Many organizations access AI models through APIs rather than running the models themselves.
Artificial intelligence
A broad term for computer systems that perform tasks associated with perception, language, prediction, reasoning or decision support. Different AI methods have very different capabilities and risks.
Benchmark
A standardized test used to compare models or systems. A benchmark is useful evidence, but it may not represent a real workload, population or operating environment.
Context window
The amount of information a model can consider in one interaction, usually measured in tokens. A larger window does not ensure that every detail will be used accurately.
Distillation
A method for training a smaller model to reproduce useful behaviour from a larger model. The goal is often lower cost or latency with acceptable quality on a defined set of tasks.
Embedding
A numerical representation that places semantically related content near each other in a mathematical space. Embeddings are commonly used for similarity search and retrieval.
Evaluation
A repeatable process for measuring whether an AI system meets defined requirements. Strong evaluations use representative examples, clear acceptance criteria and checks after the system changes.
Fine-tuning
Additional training that adapts an existing model using selected examples. Fine-tuning can shape behaviour, but it does not automatically add current facts or reliable citations.
Foundation model
A broadly trained model that can be adapted to many tasks. Language, image and multimodal foundation models often become components inside larger products.
Function calling
A structured way for a model to request that software run a defined tool or function. The surrounding application—not the model—should validate permissions, arguments and results.
Generative AI
AI that produces new text, images, audio, video, code or other content in response to input. Its output is generated from learned patterns and can be plausible without being correct.
Grounding
Connecting a model response to supplied evidence, data or constraints. Grounding can improve relevance and traceability, but the system still needs to verify that sources support the answer.
Guardrail
A technical or procedural control intended to prevent, detect or contain unwanted AI behaviour. Effective guardrails are tested against realistic failures and are not treated as perfect barriers.
Hallucination
A fluent AI output that is unsupported, fabricated or wrong. Hallucinations are a system-quality problem, not evidence that the model consciously believes something.
Inference
The process of running a trained model to produce a prediction or output. Inference cost, speed and hardware requirements shape whether an AI product is practical at scale.
Latency
The time between a request and the system’s response. Low latency matters for interactive tasks, while slower responses may be acceptable when deeper processing creates better results.
Large language model (LLM)
A model trained on large collections of text and other data to predict and generate language. An LLM can be one component of a product that also includes retrieval, tools and safety controls.
Machine learning
Methods that find patterns in data so a system can make predictions or generate outputs. Performance depends on the data, objective, evaluation and conditions in which the system is used.
Model
A learned mathematical system that maps inputs to outputs. The word can refer to anything from a small classifier to a large multimodal foundation model.
Multimodal model
A model that can work across more than one type of information, such as text, images, audio or video. Multimodal ability should be evaluated separately for each input and combination.
Open weights
A release approach that makes a model’s learned parameters available under stated licence terms. Open weights are not necessarily open source, and users still need data, code and infrastructure to operate the model.
Prompt
The instructions and context supplied to a generative model. Clear prompts help, but reliable workflows also require good source material, examples, evaluation and review.
Quantization
A technique that represents model values with lower numerical precision to reduce memory use and speed up inference. The trade-off can include some loss of quality or stability.
Red teaming
A structured attempt to discover how an AI system can fail, be misused or bypass its controls. Findings are valuable when they lead to product, policy or monitoring changes.
Retrieval-augmented generation (RAG)
A system design that retrieves relevant material and supplies it to a generative model before it answers. RAG quality depends heavily on document preparation, search and citation checks.
Synthetic data
Artificially generated examples used for training, testing or privacy-sensitive analysis. Synthetic data can fill gaps, but it can also reproduce or amplify the assumptions of the system that made it.
System prompt
Higher-priority instructions supplied by the application operating a model. A system prompt can shape behaviour but should not be treated as a secure boundary by itself.
Temperature
A setting that changes how variable a model’s generated output is. Lower temperature can make output more consistent, but it does not make unsupported claims factual.
Throughput
The amount of AI work a system can process over a period of time. It is often measured alongside latency and cost when planning production capacity.
Token
A unit into which a model breaks text for processing. Tokens can be parts of words, whole words or punctuation, and they are commonly used to measure context and API usage.
Training data
The examples used to fit a model’s parameters. Data quality, coverage, permissions and documentation all influence what the model can do and where it can fail.
Vector database
A system designed to store and search numerical embeddings. It is often used in RAG systems to retrieve semantically related passages rather than exact keyword matches.