The first press agency covering
the world of artificial intelligence.
GUIDES AND TOOLS / THE WORDS OF AI
Understanding the words.
Navigating the future.
The artificial intelligence glossary, from A to Z. Clear definitions, English terms and examples to read the news and use tools with greater awareness.
AccuracyAccuracyProportion of correct predictions out of the total cases examined. Alone it can be misleading when classes are very unbalanced.Practical example
An email classifier must also be assessed on errors that let spam through.
AI agentAI agent; agentic; agentic AISystem that uses a model and tools to perform actions towards a goal, with variable autonomy and defined limits.Practical example
An agent checks availability and prepares a booking, asking for confirmation before finalising it.
AlgorithmAlgorithmProcedure made of steps to solve a problem or transform data. Not all algorithms use artificial intelligence.Practical example
Sorting a directory by surname requires an algorithm, but not necessarily an AI model.
Algorithmic auditAlgorithmic auditSystematic examination of a system, data, and processes to evaluate behaviour, risks, and controls against explicit criteria.Practical example
It is checked whether a model changes the quality of responses for different user groups.
AlignmentAlignmentA set of methods to steer a system's behaviour towards desired goals, instructions and constraints. It does not guarantee perfect obedience.Practical example
An assistant is also evaluated on its ability to respect assigned operational limits.
APIApplication Programming InterfaceInterface that allows different programs to exchange requests and results according to defined rules.Practical example
A website sends a text to an AI service and receives a translation to display.
Artificial general intelligenceAGI; Artificial General IntelligenceConcept of AI with broad, transferable capabilities across tasks. There is no universally shared operational definition nor a single decisive test.Practical example
An announcement using the acronym AGI should be read asking which capabilities have been demonstrated.
Artificial intelligenceAI; AI; Artificial IntelligenceField of methods and systems that perform tasks such as prediction, recognition or generation. The term covers different technologies, not a single machine.Practical example
Recognising an object and generating a text are different AI applications.
AttentionAttention mechanismMechanism that weighs relationships between elements of the input during processing. It does not correspond to a person's conscious attention.Practical example
In a sentence, the model relates a pronoun to other elements of the text.
AutomationAutomationExecution of activities through rules or software, with reduced human intervention. It can operate without AI.Practical example
A rule files emails with a specific subject into a folder.
BenchmarkBenchmark testStandardised test used to compare performance. The result depends on the chosen tasks and test conditions.Practical example
A score in mathematics does not automatically demonstrate quality in journalistic writing.
BiasBias; algorithmic biasSystematic tendency in data, the model, or its use that can distort results and produce unfair effects.Practical example
Unrepresentative data can worsen the recognition of certain accents.
ChatbotConversational assistantSoftware that interacts through conversation. It may be based on rules, generative models, or a combination of methods.Practical example
A customer service answers questions; the ability to perform actions is an additional function.
ChunkingDocument segmentationDivision of extended material into smaller parts to facilitate indexing, retrieval, and processing.Practical example
A manual is divided into sections that the system can search separately.
ClassificationClassificationTask that assigns an input to one or more defined categories.Practical example
A request is classified as administrative, commercial, or technical.
ClusteringClusteringTechnique that identifies groups of similar elements without necessarily fixed categories in advance.Practical example
Thousands of reports are organised into thematic groups for review.
Computer visionComputer visionAI area dedicated to analysing images and videos for tasks such as recognition, localisation, and measurement.Practical example
A system detects cracks in photographs of an industrial component.
ContextContextInformation available to the model during a request, such as instructions, conversation, and provided documents. It does not coincide with all that the service retains.Practical example
Attaching the correct minutes helps formulate a relevant response to the meeting.
Context windowContext windowLimit to the amount of information a model can consider in one processing, often measured in tokens. A large limit does not guarantee uniform attention.Practical example
An overly long dossier may require selection or targeted retrieval of passages.
DatasetDatasetOrganised collection of data used to train, evaluate, or use a system.Practical example
A labelled image archive may be used to evaluate a recogniser.
Deep learningDeep learningFamily of learning methods based on neural networks with multiple processing layers.Practical example
A network learns features useful to distinguish objects in images.
DeepfakeRealistic synthetic contentImage, audio, or video generated or manipulated that can make a non-real representation appear authentic.Practical example
A synthetic voice is used to attribute words never spoken to a person.
Diffusion modelDiffusion modelFamily of generative models that learn a process of transformation between data and noise, also used to produce images.Practical example
A generator builds an image through successive steps starting from noise.
Digital twinDigital twinDigital representation of an object or process, linked to information about its real behaviour. It may integrate AI models but is not synonymous with AI.Practical example
A plant is simulated to compare maintenance scenarios.
EmbeddingVector representationNumerical representation of content that allows comparison of some features learned by the model.Practical example
A search finds documents on a concept even when they use different words.
ExplainabilityExplainabilityAbility to provide understandable explanations about a system's behaviour, suitable for those who need to interpret it. It does not always coincide with complete transparency.Practical example
An interface shows which factors contributed to a classification.
False negativeFalse negativeError in which the system fails to detect a case it should have reported.Practical example
An automated check does not identify a defect that is actually present.
False positiveFalse positiveError in which the system reports a condition as present that is actually absent.Practical example
A legitimate message is marked as spam.
Few-shotLearning from few examples within the promptRequest mode showing a few examples of the desired output without necessarily changing the model parameters.Practical example
Before classifying new news, three examples of headlines and categories are provided.
Fine-tuningModel adaptationAdditional training that modifies a model’s parameters to adapt it to specific data, tasks or behaviours.Practical example
A model is adapted to the style of a particular type of documents.
Foundation modelFoundation modelModel trained on broad data and usable as a base for multiple applications or adaptations.Practical example
The same model family can support summaries, classification and writing assistance.
Function callingFunction calling; tool callingGenerating a structured request for an external tool. The application executes the function and manages permissions.Practical example
The model suggests a calendar search and the software returns the appointments.
Generative AIGenerative AI; GenAISystems that produce content such as text, images, audio or code based on models learned from data.Practical example
An initial draft is generated from instructions and starting materials.
GPUGraphics Processing UnitProcessor suitable for many parallel operations, also used to train and run AI models. It is not exclusive to AI.Practical example
A server employs accelerators to process large amounts of matrix calculations.
GroundingGrounding to sourcesLinking the response to relevant data or documents. It can reduce some errors but still requires checks on sources and interpretation.Practical example
A response is built starting from a provided contract and cites its passages.
GuardrailSafety constraintControl that limits inputs, actions or outputs of a system. It can be technical or procedural and does not eliminate all risks.Practical example
An agent cannot send payments above a threshold without authorisation.
HallucinationHallucination; confabulationInaccurate or fabricated content produced by a model in a plausible form. Language fluency does not constitute fact-checking.Practical example
A text cites research that, upon checking author and journal, is found to be nonexistent.
High riskHigh-risk AIIn the AI Act, a legal classification linked to specific systems and uses. The assessment depends on the intended use and applicable regulations.Practical example
A use case in personnel selection requires verification of the specific regulatory framework.
Human-in-the-loopHuman supervision in the processWork organisation where a person intervenes at defined steps to assess, correct or authorise the system.Practical example
A draft produced by AI is verified before publication.
InferenceInferenceRunning a trained model to produce a prediction or content from new inputs.Practical example
When an assistant answers a question it is performing inference.
InteroperabilityInteroperabilityAbility of different systems to exchange data and use information according to compatible formats and rules.Practical example
An assistant consults an archive without requiring a manual copy of each document.
Knowledge distillationKnowledge distillationTechnique that transfers part of a model's behaviour to another, often smaller. Results depend on the task and data used.Practical example
A compact model is trained using examples produced by a larger system.
LatencyLatencyTime between a request and the system’s response or action. It should be distinguished from the overall request throughput.Practical example
A translation during a call requires low response times.
LLMLarge Language Model; large language modelLarge language model trained on data sequences to process and generate language. It is not an infallible repository of facts.Practical example
An LLM rewrites a text, but references to people and studies must be verified.
Machine learningML; machine learningMethods that learn regularities from data to perform tasks without manually defining every single response rule.Practical example
A model estimates demand using historical examples and relevant variables.
MCPModel Context ProtocolProtocol to connect AI applications to external tools and resources. It does not replace authentication, authorisations and application controls.Practical example
A connector exposes a document search with defined permissions to an assistant.
MemoryMemoryFunction that stores or retrieves information beyond a single request. Implementation and duration depend on the service and do not coincide with the model parameters.Practical example
An assistant retrieves a saved preference; the context window remains a separate limit.
ModelModelMathematical or computational system associating inputs to outputs through structure and parameters. It is distinct from the application using it.Practical example
Two services may use the same model with very different tools and rules.
Multi-agent systemMulti-agent systemSystem that coordinates multiple agents with different roles or tasks. More agents do not automatically mean higher quality.Practical example
One agent collects documents and another checks references, with supervision of the process.
MultimodalMultimodalCapable of processing multiple data types, e.g. text and images. The input and output modes depend on the model.Practical example
A system examines a photograph and answers a question about its content.
Neural networkNeural networkModel formed by connected units and learned parameters. The analogy with biological neurons does not imply functioning equivalent to the brain.Practical example
A network transforms input data into a prediction through multiple layers.
NLPNatural Language ProcessingField that studies automatic processing of human language, written or spoken. It includes generative and non-generative methods.Practical example
Classifying a text by topic is an NLP task.
OCROptical Character RecognitionRecognition of characters in images or scans to convert them into processable text. Corrections may be required.Practical example
A scanned letter becomes searchable text, to be checked where readability is poor.
Open modelOpen model; open weightsExpression indicating various degrees of availability of weights, code or documentation. Accessible weights do not automatically mean open source software or unrestricted use.Practical example
Before adopting it, licence, available materials and usage conditions are checked.
OverfittingOverfittingSituation where a model fits the training data too closely and performs worse on new cases.Practical example
The recogniser works well on seen images but fails on photos of a different environment.
ParametersParameters; weightsInternal values learned during training that contribute to the model's behaviour. Their number alone does not measure quality or utility.Practical example
A model with fewer parameters can be suitable for a specific task.
Personal dataPersonal dataInformation relating to an identified or identifiable person. Their processing with AI requires assessing purposes, legal bases, and relevant measures.Practical example
A document with names and contact details does not become anonymous by removing only the title.
Physical AIPhysical AI; embodied AIExpression used for AI systems that perceive and act in the physical environment, often through robots or machines.Practical example
A robot adapts a grip based on information received from its sensors.
PrecisionPrecisionAmong cases reported as positive, the proportion that are actually positive. It does not coincide with overall accuracy.Practical example
If ten defect reports include eight real defects, the precision is 80%.
Privacy by designData protection from the design stageAn approach that integrates data protection into the system's design choices, along with measures required by the context.Practical example
An assistant is designed using only the information necessary for its function.
PromptInstruction; requestInput provided to a model, which may include instructions, examples, and materials. The quality of the request helps but does not guarantee the result.Practical example
„Summarise the minutes and separate decisions, responsible parties, and deadlines.”
Prompt engineeringInstruction designDefinition and testing of instructions, examples, and context to obtain results suitable for a task.Practical example
Two requests are compared to check which produces summaries more faithful to the document.
Prompt injectionInstruction manipulationAttempt to have the system follow unauthorised instructions contained in input or external materials.Practical example
A page to be summarised contains a command attempting to disclose confidential data.
QuantisationQuantizationReduction of numerical precision used to represent certain components of the model, to save resources with possible effects on performance.Practical example
A compressed version may occupy less memory but must be re-evaluated on the actual task.
RAGRetrieval-Augmented GenerationArchitecture that retrieves information from an archive and provides it to the model during generation. It does not necessarily modify its parameters.Practical example
An assistant searches for the updated regulation and uses it to formulate the response.
RecallRecall; sensitivityAmong all truly positive cases, the proportion identified by the system. It should be considered together with precision.Practical example
If a check finds nine defects out of ten present, the recall is 90%.
Red teamingAdversarial evaluationTargeted tests to reveal problematic behaviours, vulnerabilities, and limitations of a system before or after release.Practical example
Evaluators try difficult requests and attempts to bypass controls.
Reinforcement learningReinforcement learningMethod in which a system learns a strategy based on rewards associated with actions and their outcomes.Practical example
An agent in a simulation tries strategies and receives a score for the result.
RLHFReinforcement Learning from Human FeedbackFamily of techniques that use human evaluations to guide model behaviour, often via a preference or reward signal.Practical example
Evaluators compare alternative responses to indicate which are more useful.
RobustnessRobustnessAbility to maintain appropriate behaviour when inputs or conditions change within a defined scope.Practical example
It is checked whether noise and different lighting worsen visual recognition.
Semantic searchSemantic searchRetrieval of content based on learned meaning relationships, beyond literal word matching.Practical example
Searching for „work absences” may retrieve documents discussing leave.
Supervised learningSupervised learningMethod using examples accompanied by the outcome or label to be predicted.Practical example
Photographs already identified as containing a defect help train quality control.
Synthetic dataSynthetic dataData artificially generated to simulate some properties of real data. They are not automatically anonymous or representative.Practical example
Test invoices are created without using the actual customer documents.
System cardSystem cardDocument describing capabilities, evaluations, limits, and distribution choices of an AI system. Contents and methods vary among producers.Practical example
The card helps identify which tests have been performed and by whom.
TemperatureTemperatureSampling parameter that can make generation more or less variable. It is not an indicator of the truth of responses.Practical example
Reducing variability does not automatically turn a response into a verified fact.
TokenProcessing unitUnit into which a model segments or represents data. In text, it may correspond to a word, part of a word, or punctuation.Practical example
Token length affects context limits and, in many services, consumption.
TrainingTrainingProcess that changes the parameters of a model through data and a learning criterion. It is distinct from using the already trained model.Practical example
A system learns to recognise invoices by observing example documents.
TransformerTransformer architectureNeural network architecture based on attention mechanisms, used in many language and multimodal models.Practical example
A model uses relationships between sequence elements to process context.
TransparencyTransparencyAvailability of relevant information on purpose, functioning, data, limits and responsibility of the system. It is not equivalent to just publishing the code.Practical example
Users of a response should be able to recognise AI involvement and its limits.
Unsupervised learningUnsupervised learningLearning that seeks structures in data without starting from correct answers already assigned to each example.Practical example
A system groups similar documents without first receiving their categories.
ValidationValidationEvaluation of the model during development on data distinct from those used to learn parameters. It should be distinguished from final testing.Practical example
A configuration is chosen using a validation set without adapting to the final test.
Vector databaseVector databaseStorage and retrieval system designed to compare vectors, often used with embeddings and semantic retrieval.Practical example
An archive finds the documents numerically most similar to a question's representation.
WorkflowWorkflowOrganised sequence of activities and controls. It may include AI, automations, and human steps with defined paths.Practical example
A text passes from source gathering to draft, verification, revision, and publication.
Zero-shotWithout examples in the promptRequest mode that describes the task without including specific examples to imitate. It does not mean the model was not trained.Practical example
A title is classified providing only permitted categories.
HOW TO READ THE GLOSSARY
A common language, without losing differences.
The definitions and examples are editorial explanations in English. Some words have different meanings depending on context: where necessary, we distinguish the technical concept from commercial or legal use. Updated 23 October 2026.
Propose an entry or report a correction →A project of the AELLE Media group
