A shared language for AI

AI, in plain language.

From agents and RAG to open weights. Look up a term, read a short explanation and unfold an everyday example. No technical background needed.

Find a termTopic

120 of 120 terms

Illustrative examples · not actual case reports

Responsibility & security

Access control

Rules and technical controls determining who or what may access data and functions. An AI agent should only receive the permissions needed for its task.

Example & further reading

Imagine this

An assistant could be allowed to read published guidance but not modify it or open confidential cases in the same system.

Search terms: Adgangskontrol · rettighedsstyring · access control · autorisation

Accuracy

In classification, accuracy is the proportion of all examples assigned to the correct category. A high score can hide poor performance on rare categories.

Example & further reading

Imagine this

If a model sorts ninety out of one hundred fictional enquiries correctly, its accuracy on that particular test is ninety percent.

Search terms: Andel korrekte svar · accuracy · nøjagtighed · klassifikationsnøjagtighed

Agent harness

A nonstandard umbrella term for software surrounding an agent model, such as tools, instructions, and execution controls. A harness is not the AI model itself; implementations vary.

Example & further reading

Imagine this

The same model can have different tools and approval requirements in two applications, changing which actions each application permits.

Search terms: Agent-harness · harness · agent scaffold · agentramme

Agentic workflow

A sequence of steps that uses AI to perform a task and may select subsequent actions. Some steps can follow fixed rules while others are model-directed.

Example & further reading

Imagine this

A request is categorized, relevant guidance is retrieved, and a draft answer goes to an employee before reaching the customer.

Search terms: Agentbaseret arbejdsforløb · agentic workflow · agentisk arbejdsgang

AI Act

The EU's common legal framework for AI, linking requirements to particular uses and risks. The specific assessment depends on the system, its purpose and the organisation's role.

Example & further reading

Imagine this

Before introducing an AI service, practitioners and legal colleagues could clarify its purpose and identify which requirements need closer examination.

Search terms: AI-forordningen · AI Act · EU AI Act · AI-forordning

AI agent

An AI system where a model selects next steps and uses available tools to complete a task. Its actions depend on application rules and user permissions.

Example & further reading

Imagine this

You request a meeting brief; the system searches approved documents, gathers relevant information, and creates a draft for your review.

Search terms: AI-agent · agent · kunstig intelligens-agent

AI governance

An organisation's arrangements for deciding, developing, using and overseeing AI. They cover roles, responsibilities, risk assessments and decisions throughout a system's lifetime, not only technical settings.

Example & further reading

Imagine this

A council could agree who approves new AI tools, who owns each service and who responds when something goes wrong.

Search terms: AI-governance · AI-styring · governance

AI literacy

The knowledge and skills needed to understand AI's possibilities, limitations and risks and use it thoughtfully. Relevant competence depends on the person's role and tasks.

Example & further reading

Imagine this

A staff member could learn to check an AI summary against the original and recognise when a task needs specialist help.

Search terms: AI-kompetence · AI literacy · AI-forståelse · AI-færdigheder

AI platform

A combined software environment for building, using, or managing AI applications. It may include model access, data connections, testing, and access controls; available features vary.

Example & further reading

Imagine this

An organization manages model access, answer testing, and employee permissions in one environment while developing a service for everyday questions.

Search terms: AI-platform · platform for kunstig intelligens

Algorithm

A sequence of computational steps that a computer can follow to complete a task. An algorithm can use fixed rules and does not have to involve AI.

Example & further reading

Imagine this

Sorting a list of meetings by date is an algorithmic task, even when the solution uses no artificial intelligence.

Search terms: Algoritme · algorithm

Annotation

Adding descriptions or labels to data, such as topics on texts or marked regions in images. These labels support training and checking and require clear guidelines.

Example & further reading

Imagine this

Two employees label topics in public news articles and compare their choices to keep categories consistent across the finished dataset.

Search terms: Annotering · datamærkning · labeling · labelling · data annotation

Anonymisation

Processing information so people can no longer be identified under relevant conditions. Removing names is not enough; combinations of information and possibilities for re-identification must also be assessed.

Example & further reading

Imagine this

A team could check whether a table still reveals individuals through rare combinations of age, location and medical conditions.

Search terms: Anonymisering · anonymization · anonyme data

API

A defined interface that applications use to exchange data or request functions. In AI, an API can provide access to a model from another application.

Example & further reading

Imagine this

A website's help box sends your question through an API to a model service and receives text to display on the page.

Search terms: API · Application Programming Interface · programmeringsgrænseflade

Artificial intelligence (AI)

A broad term for computer-based methods that perform tasks such as recognising patterns, processing language or suggesting actions. AI includes much more than chatbots.

Example & further reading

Imagine this

A library could use AI to suggest subject labels for public book descriptions, which a librarian then checks.

Search terms: Kunstig intelligens (AI) · AI · KI · artificial intelligence

Attention

A computational mechanism that weights how parts of an input relate to one another. It is a mathematical operation, not human attention or understanding.

Example & further reading

Imagine this

In a sentence about a building needing renovation, the mechanism can connect the word it with the building rather than other objects.

Search terms: Attention · self-attention · opmærksomhedsmekanisme

Audit log

A chronological record of important system actions, such as access, changes and approvals. It allows people to investigate events later and trace actions to responsible actors.

Example & further reading

Imagine this

When an editor publishes an article, the log could show its time, user and version so the change can later be examined.

Search terms: Revisionslog · audit log · revisionsspor · hændelseslog

Benchmark

A shared test that makes results from different systems more comparable. A strong score only describes performance on the tasks and conditions covered by that test.

Example & further reading

Imagine this

A council could compare two models on one document task before checking whether the better model also handles its own documents.

Search terms: Benchmark · sammenligningstest · benchmarking

Bias

Systematic distortions in data, development or use that can influence an AI system's results. They may affect groups differently and arise from more than training data alone.

Example & further reading

Imagine this

A speech tool may make more errors on certain dialects if development and testing have not covered linguistic variation adequately.

Search terms: Bias · skævhed · systematisk skævhed · AI bias

Central processing unit (CPU)

The computer's general-purpose processor, which runs programs and coordinates many kinds of tasks. It can also run AI, although specialised processors better suit some demanding calculations.

Example & further reading

Imagine this

A small model for sorting public documents might run on a server without an additional graphics card.

Search terms: Centralprocessor (CPU) · CPU · centralprocessor · central processing unit

Chunking

Splitting larger documents into smaller pieces that can be retrieved or provided to a model. Piece size and boundaries affect how much context remains available.

Example & further reading

Imagine this

A long association handbook is split at headings so a question about keys can retrieve the relevant section without the entire handbook.

Search terms: Tekstopdeling · chunking · chunk · tekstbidder

Citation (AI answer)

A reference from an AI answer to a source or specific passage. It helps readers check claims, but does not alone prove the source supports them.

Example & further reading

Imagine this

After stating an event detail, the answer provides a link to the programme so you can check its date and venue.

Search terms: Kildehenvisning i AI-svar · citation · source citation · kildehenvisning

Cloud

Network access to shared computing resources that can be allocated and adjusted on demand. Resources include servers, storage and applications, with responsibilities depending on the service type.

Example & further reading

Imagine this

A development team could rent extra computing capacity for a short model test and release it once the planned assessment finishes.

Search terms: Cloud · cloud computing · skyen · cloudtjenester

Context window

The limit on information a model can work with in one computation, measured in tokens. Conversation history and documents take space; more capacity does not guarantee better comprehension.

Example & further reading

Imagine this

If lengthy minutes and previous messages take too much space, the system must shorten the material or select relevant passages.

Search terms: Kontekstvindue · context window · kontekstlængde

Data card

A structured description of a dataset, including its collection, contents, purpose and limitations. It helps others understand what the material can and cannot reasonably support.

Example & further reading

Imagine this

A data card for fictional resident letters could describe their topics, the authors' choices and language variations missing from the current collection.

Search terms: Datakort · data card · datasætdokumentation

Data minimisation

The principle of limiting personal data to what is necessary for a clear purpose. Having access to more information does not mean it should all be used.

Example & further reading

Imagine this

If AI only needs to improve a standard letter's wording, the task might not require a resident's name, address or personal identifier.

Search terms: Dataminimering · data minimization · mindst mulige data

Data poisoning

Manipulating data used to train or adapt a model so that its behaviour changes in unwanted ways. Poisoning can introduce errors, biases or particular hidden responses.

Example & further reading

Imagine this

A testing team could examine whether false labels in a practice dataset cause a model to consistently misclassify certain enquiry types.

Search terms: Dataforgiftning · data poisoning · forgiftede træningsdata

Data protection impact assessment

A systematic examination of planned personal-data processing and its risks to people. It considers necessity, proportionality and measures intended to reduce those risks before proceeding.

Example & further reading

Imagine this

When planning an AI assistant, a council could map data flows, possible consequences for residents and safeguards that need examination.

Search terms: Konsekvensanalyse vedrørende databeskyttelse · DPIA · konsekvensanalyse

Data residency

Where data is physically stored or processed, such as particular countries or data centres. Location alone does not establish who can access it or whether processing meets every requirement.

Example & further reading

Imagine this

When considering an AI service, a council could examine the locations of active data, backups and the processing performed by the model.

Search terms: Dataplacering · data residency · datalokation · dataresidens

Dataset

A defined collection of data, such as a table, texts or images. AI work may use separate datasets for training, ongoing adjustments and independent testing.

Example & further reading

Imagine this

A spreadsheet containing dates, weather and total electricity use for a sports centre could form a dataset for a limited trial.

Search terms: Datasæt · dataset · data set

Deep learning

Machine learning using neural networks with many processing layers. Together, these layers can capture complex patterns in images, audio and text during training.

Example & further reading

Imagine this

A trial using road-surface images could apply deep learning to mark possible cracks for subsequent inspection by a person.

Search terms: Dyb læring · deep learning · DL

Diffusion model

A generative model trained to remove added noise from data. During generation, the process can start with noise and gradually produce something such as an image.

Example & further reading

Imagine this

An image generator could create a clearly labelled illustration of a green urban space for discussion, not evidence of an existing place.

Search terms: Diffusionsmodel · diffusion model · denoising diffusion

Distillation

A method that trains an often smaller model using a larger model's outputs or calculations. The aim is to retain useful capabilities while requiring fewer resources.

Example & further reading

Imagine this

A team could test whether a smaller model can take over simple public-text categorisation after training with a larger model.

Search terms: Destillation · knowledge distillation · model distillation

Embedding

A numerical representation a model calculates from data such as text or images. It can support content comparisons using patterns learned during the model's training.

Example & further reading

Imagine this

Two texts about bicycle repair may receive similar numerical representations even when one mentions punctures and the other describes flat tires.

Search terms: Embedding · embeddingvektor · vektorrepræsentation · indlejring

Encryption

A cryptographic transformation that makes data unreadable without the required key. Encryption can protect information during storage or transfer, but does not replace access management.

Example & further reading

Imagine this

A document could be encrypted while travelling to a server, while the organisation separately controls who may open it afterwards.

Search terms: Kryptering · encryption · kryptografi

Evaluation

A systematic assessment of how well an AI system performs a particular task, using selected examples, explicit quality criteria, and often both measurements and human judgement.

Example & further reading

Imagine this

A library could ask staff to assess chatbot answers to the same twenty questions about opening hours, events and borrowing.

Search terms: Evaluering · evaluering · model evaluation · evals

Explainability

The ability to give people an understandable and faithful explanation of an AI system's output or operation. A fluent explanation is not necessarily an accurate one.

Example & further reading

Imagine this

A document model could show which passages contributed to its classification so that a staff member can examine the result's basis.

Search terms: Forklarbarhed · explainability · XAI · explainable AI

False negative

A result where a model misses something that is actually present. This matters particularly when the system is meant to identify certain information or problems.

Example & further reading

Imagine this

A check of fictional documents could miss a personal identifier and incorrectly mark that document as ready for public sharing.

Search terms: Falsk negativ · overset fund · FN

False positive

A result where a model says something is present when it is not. Also called a false alarm, its importance depends on the task and consequences.

Example & further reading

Imagine this

A spam filter could move an ordinary message from a resident into the spam folder even though it is legitimate.

Search terms: Falsk positiv · falsk alarm · FP

Fine-tuning

Further training of an existing model on selected data to adapt it to a task. Its internal values change, unlike simply giving it a different prompt.

Example & further reading

Imagine this

A team could fine-tune a model using approved public examples to improve its handling of a particular document category.

Search terms: Finjustering · fine tuning · finetuning · eftertræning

Foundation model

A model trained broadly on large amounts of data to provide a starting point for many tasks. It can be adapted later and need not process only text.

Example & further reading

Imagine this

The same foundation model could underpin both a writing assistant and a tool suggesting topics for public documents.

Search terms: Grundmodel · foundation model · basismodel

Framework

A reusable structure of code and conventions for building an application. An AI framework may bring together common parts such as model calls, tools, and workflows.

Example & further reading

Imagine this

A development team uses the same basic structure for several help services, adapting each service's data sources, features, and interface.

Search terms: Framework · softwareramme · udviklingsramme

Function calling

A form of tool calling where the model supplies a programmed function's name and inputs in a defined format. The application must check and execute it.

Example & further reading

Imagine this

Your weather question becomes a function name with city and date fields; the application then fetches the relevant forecast.

Search terms: Funktionskald · function calling · structured function call

Generative AI

AI that produces text, images, audio or code using patterns from training data and the current input. Its output is not necessarily accurate or original.

Example & further reading

Imagine this

A communications officer could generate three drafts about library opening hours and check every detail before publishing the chosen version.

Search terms: Generativ AI · GenAI · generative artificial intelligence

Graphics processing unit (GPU)

A processor suited to performing many similar calculations at the same time. It is therefore often used for graphics and demanding AI training or model execution.

Example & further reading

Imagine this

An IT team could use a GPU server for a local language model and measure response time, memory needs and electricity use.

Search terms: Grafikprocessor (GPU) · GPU · grafikprocessor · graphics processing unit

Grounding

Connecting an AI answer to specific information from available sources. This can reduce invented content, but quality still depends on those sources and their use.

Example & further reading

Imagine this

An answer about opening hours uses the library's current website; you can check the answer by opening the source it used.

Search terms: Kildebasering · grounding · forankring i kilder

Guardrails

Rules and checks that constrain an AI system's inputs, responses, or actions. They may check sensitive information or require approval, but do not eliminate every risk.

Example & further reading

Imagine this

A check flags possible personal identification numbers in a draft and pauses sending so an employee can review the content.

Search terms: Sikkerhedsbegrænsninger · guardrails · sikkerhedsregler

Hallucination

When generative AI produces plausible-sounding information that is wrong, fabricated or unsupported by the material. The term describes an output failure, not a human experience.

Example & further reading

Imagine this

A draft refers to a report that does not exist. An employee discovers the error by checking the claimed sources.

Search terms: Hallucination · konfabulation · confabulation · fabrikerede svar

Hugging Face

A company and platform offering machine learning tools and a community that shares models, datasets, and demonstration applications. Individual models and resources have different usage terms.

Example & further reading

Imagine this

A developer explores models for categorizing incoming requests and reads their descriptions and usage terms on Hugging Face before testing them.

Search terms: Hugging Face · HF · HuggingFace

Human in the loop

A workflow where a person reviews, changes, or approves an AI system's proposals during execution. An approval step can pause selected actions pending that decision.

Example & further reading

Imagine this

The system prepares a message for association members, but an employee reviews and approves the text before the application sends it.

Search terms: Menneske i beslutningsforløbet · human-in-the-loop · HITL · menneskelig godkendelse

Hybrid search

Search that combines multiple methods, often keyword and vector search, and merges their results. It can capture both specific terms and content with related meaning.

Example & further reading

Imagine this

You search for a spare part using its product number and problem description; the system combines word matches with related issues.

Search terms: Hybridsøgning · hybrid search · kombineret søgning

Inference

Using a trained model on new input to calculate a response, category or prediction. The values established during training normally stay unchanged during this operation.

Example & further reading

Imagine this

Each time an employee asks an approved model to shorten a public text, the system performs a new inference.

Search terms: Inferens · inference · modelkørsel

Jailbreak

An attempt to make an AI model disregard its safety rules and respond to otherwise restricted requests. The term describes an attack attempt, not necessarily a successful one.

Example & further reading

Imagine this

During an authorised security test, a team could check whether a chatbot maintains its boundaries when someone tries to change its role.

Search terms: Jailbreak · jailbreaking · omgåelse af sikkerhedsregler

Keyword search

Searching documents for words or phrases. The system may process word forms and spelling variations, but starts from written terms rather than calculated similarity of meaning.

Example & further reading

Imagine this

You enter a specific product number in the search box to locate its manual among many documents describing similar products.

Search terms: Nøgleordssøgning · keyword search · ordbaseret søgning · lexical search

Knowledge base

An organized collection of documents or other information that people or applications can search. An AI application can use the collection as material for its answers.

Example & further reading

Imagine this

Association rules, practical guides, and contact information are collected in a knowledge base where employees and a help service can find information.

Search terms: Vidensbase · knowledge base · KB · videnssamling

Knowledge graph

A structured collection describing things and named connections between them, such as people, organizations, and memberships. These connections can support searches across related information.

Example & further reading

Imagine this

A knowledge graph connects members, working groups, and events so you can find which groups were involved in a particular meeting.

Search terms: Vidensgraf · knowledge graph · KG

LangChain

A software framework for building AI applications and agents with connections to models and tools. Developers configure the components; LangChain is not itself an AI model.

Example & further reading

Imagine this

A developer uses LangChain to connect a chosen language model with a search tool that retrieves practical information from association documents.

Search terms: LangChain · LangChain framework

LangGraph

A software framework for managing agent workflows as connected steps. It supports features including saved state, branching paths, and pauses for human review during execution.

Example & further reading

Imagine this

An application follows defined steps, can return for missing information, and pauses for an employee's review before continuing through the process.

Search terms: LangGraph · LangGraph framework

Large language model (LLM)

A large model trained on extensive text to calculate patterns in language. It can generate text and summaries, but it is not a dependable source of facts.

Example & further reading

Imagine this

A language model could rewrite public guidance in simpler language while a specialist checks its meaning and important qualifications.

Search terms: Stor sprogmodel (LLM) · LLM · large language model · sprogmodel

Latency

The delay a user or system experiences between a request and a response. For AI, this can distinguish time to the first output from time to completion.

Example & further reading

Imagine this

A chatbot could start displaying text after one second even though completing the answer takes another five seconds in total.

Search terms: Svartid · latency · forsinkelse · responstid

Machine learning

An approach to AI that adjusts a model's calculations using examples in data. Developers do not need to write a separate rule for every situation.

Example & further reading

Imagine this

A municipality could explore whether historical temperatures and energy readings help predict the heating needs of a swimming pool.

Search terms: Maskinlæring · ML · machine learning

MCP

Model Context Protocol is an open standard connecting AI applications with external tools or data sources. Access and permissions still depend on the particular setup.

Example & further reading

Imagine this

An AI application can use an MCP connection to search association documents when the connection and document permissions are configured.

Search terms: MCP · Model Context Protocol · modelkontekstprotokol

Memory (agent)

Stored information an agent system can reuse later, such as conversation history, preferences, or task notes. This involves storing and retrieving data, not human memory.

Example & further reading

Imagine this

With your permission, a writing assistant saves your preference for short answers and retrieves it when you start another conversation.

Search terms: Agenthukommelse · memory · agent memory · hukommelse

Model

The computational component that turns input into something such as a prediction or text. A model is not the complete application, interface, data sources or access controls.

Example & further reading

Imagine this

A municipal chatbot could switch language models while keeping the same interface, login system and approved document sources.

Search terms: Model · AI-model · ML-model

Model card

A structured description of an AI model, its intended uses, tests and limitations. It helps readers assess whether the model is suitable for a particular task.

Example & further reading

Imagine this

Before choosing a model for Danish letters, a team could examine its model card for language tests and uses it was not designed for.

Search terms: Modelkort · model card · model documentation

Model hub

A place collecting AI models with files and descriptions so they can be found, shared, or downloaded. Access, documentation, and usage terms vary between models.

Example & further reading

Imagine this

You can look for a Danish text model in a model hub and compare its documentation before a developer tests it.

Search terms: Modelkatalog · model hub · modelrepository · modelbibliotek

Model weights

Numerical values that determine the strength of connections or influences in a model's calculations. They are adjusted during training and are central to a trained model.

Example & further reading

Imagine this

When an IT team downloads model weights, it does not automatically receive the original training data or a complete user interface.

Search terms: Modelvægte · weights · vægte

Multi-agent system

A system with multiple AI agents performing parts of a shared task and exchanging results. Using several agents also introduces coordination and checking requirements.

Example & further reading

Imagine this

When planning an event, one agent checks venues and another checks transport; their results are combined for your further work.

Search terms: Multiagentsystem · multi-agent system · fleragentsystem

Multimodal AI

AI that processes or combines different kinds of input or output, such as text, images and audio. The supported combinations depend on the particular model.

Example & further reading

Imagine this

An employee could ask a question about a public chart and receive a written explanation of the trend it shows.

Search terms: Multimodal AI · multimodal · LMM · multimodal model

Neural network

A model built from connected calculation units that transform numbers through successive layers. Connection strengths can change during training; the model is not a human brain.

Example & further reading

Imagine this

When an app recognises a road sign in an image, a neural network may calculate which sign category it resembles.

Search terms: Neuralt netværk · kunstigt neuralt netværk · neural net · ANN

Observability

The ability to understand what is happening inside a system through measurements, events and connected traces. It helps people investigate failures and slow processes.

Example & further reading

Imagine this

When a chatbot is slow, the team could check whether waiting occurs in search, the model call or subsequent response processing.

Search terms: Observerbarhed · observability · driftsindsigt · telemetri

Ollama

Software for running and accessing AI models, including on your own computer; cloud features also exist. Where data is processed depends on the model and setup.

Example & further reading

Imagine this

A developer tests a downloaded model on a computer using Ollama and checks the settings before using real documents in the experiment.

Search terms: Ollama · Ollama runtime

On-premises

Running a system on hardware at the organisation's own physical premises or data centre. Location does not automatically ensure security; operations, updates and access still need management.

Example & further reading

Imagine this

A council could test a language model on a server in its own data centre and arrange the system's maintenance itself.

Search terms: On-premises · on-prem · lokalt datacenter · on premise

Open source AI

AI with freedoms to use, study, modify and share. OSI's definition also requires relevant materials such as code, model values and training-data information; free access is insufficient.

Example & further reading

Imagine this

When assessing an open AI solution, the team checks permissions and available development materials rather than relying on its website label.

Search terms: Open source-AI · OSAID

Open weights

Making trained model values available so the model can run outside its provider's service. Licences and supporting materials vary; open weights do not automatically mean open source.

Example & further reading

Imagine this

A municipality could investigate local operation with open weights while checking the licence, security and operating resources.

Search terms: Åbne modelvægte · open weights · open-weight models · tilgængelige modelvægte

OpenCode

An open-source AI coding agent available through a terminal, desktop application, or code editor. It connects to selected models and tools for work on software projects.

Example & further reading

Imagine this

A developer asks OpenCode to explain website code and propose a change, which is then reviewed and tested before use.

Search terms: OpenCode · OpenCode AI · opencode-ai

Orchestration

Coordinating which models, tools, or agents run when, and how their results move between steps. This can use fixed rules or choices produced by a model.

Example & further reading

Imagine this

A system starts the search, passes the findings to summarization, and then opens an approval step before anything is sent.

Search terms: Orkestrering · orchestration · agentorkestrering

Overfitting

When a model fits its training examples so closely that it performs poorly on new examples. It may capture incidental details instead of reliable patterns.

Example & further reading

Imagine this

A model predicts the training year's energy consumption precisely but fails the following year because it also fitted random fluctuations.

Search terms: Overtilpasning · overfitting · overfit

Parameters

Internal model values that are usually adjusted during training and influence its calculations. Their number indicates model size but does not determine quality by itself.

Example & further reading

Imagine this

Two language models with different parameter counts should be tested on the same task; the larger one may not perform best.

Search terms: Parametre · modelparametre · parameters · parameter count

Personal data

Information relating to an identified or identifiable person. It can include names, contact details or combinations of information that make it possible to recognise someone.

Example & further reading

Imagine this

A message may contain personal data without a name if its address, workplace and description together make the sender identifiable.

Search terms: Personoplysninger · persondata · personal information

Planning (agent)

Dividing a task into proposed steps and potentially revising their order after new results. A plan is working information, not proof that the task will succeed.

Example & further reading

Imagine this

For a newsletter, the system outlines steps to find sources, draft text, and check links before you review the result.

Search terms: Agentplanlægning · planning · agent planning · opgaveplanlægning

Precision

The proportion of a model's positive findings that are actually correct. Precision asks how dependable the flagged findings are, rather than how many relevant cases were missed.

Example & further reading

Imagine this

A document checker flags ten texts as incomplete; if eight really lack information, precision is eighty percent in this test.

Search terms: Præcision · precision · positiv prædiktiv værdi

Prompt

The input used to guide a model's response, such as the task, background information and desired format. A clear prompt does not guarantee an accurate result.

Example & further reading

Imagine this

A prompt could request five plain-language bullet points from an attached public text without adding any new information.

Search terms: Prompt · instruktion · AI-instruktion

Prompt injection

When instructions in user input or external content try to divert an AI system from its intended task. Such content can appear in documents or webpages.

Example & further reading

Imagine this

A document being summarised could contain hidden instructions to alter the conclusion; the system should treat them as document content.

Search terms: Prompt-injektion · prompt injection · indirect prompt injection

Provenance

Information about where data or content came from and how it was produced. A provenance trail helps people examine sources, changes and responsibilities along the way.

Example & further reading

Imagine this

For an AI-generated meeting summary, a team could record its source recording, model version and when a staff member reviewed the text.

Search terms: Oprindelsesspor · proveniens · data provenance · dataoprindelse

Pseudonymisation

Replacing direct identifiers with codes, for example, while keeping additional identifying information separately protected. The connection may still exist, so pseudonymisation is not the same as anonymisation.

Example & further reading

Imagine this

In a fictional study, participants could receive numbers while a separately protected key connects those numbers to their real names.

Search terms: Pseudonymisering · pseudonymization · pseudonymiserede data

Quantization

Representing a model's numbers with lower precision so they require less memory and sometimes less computation. The saving can affect quality and therefore needs testing.

Example & further reading

Imagine this

A quantised language model may fit on a smaller computer, but the team compares its answers with the original model first.

Search terms: Kvantisering · quantization · quantisation · 4-bit · 8-bit

RAG

A method where a system retrieves relevant sources and provides them to an AI model to inform an answer. These sources do not automatically retrain the model.

Example & further reading

Imagine this

You ask about association facilities; the system retrieves passages from the current handbook and uses them to formulate an answer.

Search terms: RAG · Retrieval-Augmented Generation · retrieval augmented generation · søgeunderstøttet generering

Rate limit

A limit on the number of requests or amount of usage a client can submit within a period. It protects capacity and can reduce abusive use.

Example & further reading

Imagine this

If a service allows sixty calls per minute, an import could queue its requests instead of sending them all at once.

Search terms: Forbrugsgrænse pr. tidsrum · rate limit · rate limiting · kaldsgrænse

Recall

The proportion of genuinely relevant cases that a model finds. High recall means few missed cases, but does not by itself reveal how many incorrect alerts it produces.

Example & further reading

Imagine this

If twenty guidance documents are relevant and the search finds fifteen of them, recall is seventy-five percent for that test.

Search terms: Genfinding · recall · sensitivitet · følsomhed

Red teaming

A planned, authorised exercise that actively looks for failures and harmful possibilities in an AI system. Its purpose is to discover weaknesses and improve safeguards.

Example & further reading

Imagine this

A multidisciplinary team could try difficult questions in a test environment and document situations where the chatbot gives misleading or inappropriate answers.

Search terms: Red teaming · AI red teaming · adversarial testing · modstanderprøvning

Reinforcement learning

Training in which a system tries actions and adjusts through rewards or penalties within an environment. The aim is effective sequences of actions; reward design matters.

Example & further reading

Imagine this

In a simulation, a system could try heating-control strategies and receive rewards for low energy use within comfort limits.

Search terms: Forstærkningslæring · RL · reinforcement learning

Reranking

Reassessing an existing list of search results and changing their order by relevance. An additional model may compare each result directly with the question.

Example & further reading

Imagine this

A search returns twenty guides; an extra step moves the text most relevant to your particular printer problem higher up.

Search terms: Genrangering · reranking · rerank · omrangering

Retrieval

Finding and fetching material from a collection that may be relevant to a query. This is the search stage; writing a new answer is a separate step.

Example & further reading

Imagine this

A question about summer events leads the system to retrieve relevant calendar entries before they are potentially used in a combined answer.

Search terms: Informationshentning · retrieval · information retrieval · genfinding

Robustness

How well an AI system maintains useful performance when inputs or conditions change. Robustness can be explored using spelling errors, different wording and unusual situations.

Example & further reading

Imagine this

A team could try the same question with typing errors and short sentences to see whether the chatbot still responds sensibly.

Search terms: Robusthed · robustness · robust AI

Sandbox

A separated environment where code or tools can run with limited access to the rest of a system. Protection depends on configuration; a sandbox is not risk-free.

Example & further reading

Imagine this

An AI-generated calculation runs on sample data copies in a test environment without access to the organization's actual documents or accounts.

Search terms: Afgrænset testmiljø · sandbox · sandkasse

SDK

A package of code, tools, and guidance for developers building for a particular service or platform. It can make an API easier to use.

Example & further reading

Imagine this

A developer uses a prepared code package to send questions to a model service without writing the entire connection from scratch.

Search terms: SDK · Software Development Kit · softwareudviklingspakke

Self-hosting

Taking responsibility for installing and operating an application on chosen infrastructure. This may be owned hardware or rented cloud servers, and it requires ongoing maintenance.

Example & further reading

Imagine this

A team could operate its own code repository on a rented server and also handle updates, backups and monitoring.

Search terms: Egen hosting · self-hosted · self-hosting · selvhosting · egen drift

Semantic search

Search that aims to find content with related meaning rather than only matching words. It often uses models to compare queries with potential results.

Example & further reading

Imagine this

You search for help with a punctured bicycle tire and also receive a guide describing how to replace an inner tube.

Search terms: Semantisk søgning · semantic search · betydningsbaseret søgning

Sensitive personal data

Specific categories of personal data with special protection, such as health, political opinions and religion. This is a defined legal category; not all confidential information belongs to it.

Example & further reading

Imagine this

A letter about an identifiable person's diagnosis contains sensitive data; their name alone does not belong to that category.

Search terms: Følsomme personoplysninger · særlige kategorier · special category data · følsomme oplysninger

Session

A bounded conversation or work period in an application, with associated messages and activities. Its duration and what gets stored depend on the application.

Example & further reading

Imagine this

You develop a meeting brief through several messages in one session; a new work period may have a different associated history.

Search terms: Session · arbejdssession · agentsession

Shadow AI

The use of AI tools or agents outside an organisation's known and managed arrangements. It can create blind spots around data, security, costs and responsibility.

Example & further reading

Imagine this

A staff member could start using a personal chatbot for work before the organisation knows about the tool or assesses its use.

Search terms: Skygge-AI · shadow AI · uautoriseret AI · uadministreret AI

State

The data describing where an application or agent currently stands in a workflow. This can include selected settings, available results, and steps still to complete.

Example & further reading

Imagine this

During registration, the application's state records that a name is entered, a time is selected, and final confirmation remains pending.

Search terms: Tilstand · state · agent state · systemtilstand

Supervised learning

Training with examples that already include the desired answer. The model is adjusted to relate inputs to answers so it can predict answers for new examples.

Example & further reading

Imagine this

Road-sign images with correct categories could train a model, which is then tested on a separate set of signs.

Search terms: Superviseret læring · overvåget læring · supervised learning

Synthetic data

Artificially produced data that imitate relevant properties of real data. They can support testing or training but do not automatically ensure anonymity or representative results.

Example & further reading

Imagine this

A team could test a form using invented addresses and events rather than copying real citizen case records into development.

Search terms: Syntetiske data · synthetic data · kunstigt genererede data

Temperature

A setting that affects randomness when a model selects the next piece of text. Higher temperature generally increases variation; lower temperature does not guarantee accuracy.

Example & further reading

Imagine this

More variation may help with brainstorming, while standard information drafts often need less variation and still require careful checking.

Search terms: Temperatur · temperature · sampling temperature

Test set

A separate collection of examples used to assess a model after development. These examples are kept out of training and ongoing tuning to make the assessment more meaningful.

Example & further reading

Imagine this

A development team could reserve fifty fictional resident enquiries that the model only encounters during the final assessment of the new service.

Search terms: Testsæt · testdata · test dataset

Throughput

The amount of work a system actually processes per unit of time, such as documents per minute. High throughput does not necessarily mean short waits for individual users.

Example & further reading

Imagine this

Two systems could take the same time to process one document while one handles many more documents at the same time.

Search terms: Behandlingskapacitet · throughput · gennemløb · gennemstrømning

Token

A small unit used by a language model to process text. It may be a word, word fragment or character; the division depends on model and language.

Example & further reading

Imagine this

A long Danish word may become several tokens, so a page's word count is not the same as its token count.

Search terms: Token · tokens · teksttoken

Tool calling

When an AI system requests information or an action from a connected tool. The application or service executes the request and can return results to the model.

Example & further reading

Imagine this

You ask whether an item is in stock; a connected lookup tool retrieves the shop's current inventory for the answer.

Search terms: Værktøjskald · tool calling · tool use · værktøjsbrug

Training

The process of adjusting a model's internal values using data and a task objective. Training is different from using a model that has already been trained.

Example & further reading

Imagine this

A development team trains a model on energy readings and known heating consumption before testing it on other periods.

Search terms: Træning · modeltræning · model training

Training data

The examples used to adjust a model during training. They should be separated from final test data so quality is also assessed on previously unseen examples.

Example & further reading

Imagine this

Energy readings from earlier years could support training, while a separate dataset is reserved for checking the model's predictions afterwards.

Search terms: Træningsdata · training set · træningssæt

Transfer learning

Using patterns from an already trained model as a starting point for another task. This can reduce new training needs but requires testing on the new task.

Example & further reading

Imagine this

A general image model could provide the starting point for a waste-sorting trial instead of training everything from scratch.

Search terms: Overførselslæring · transfer learning · genbrug af modeltræning

Transformer

A model architecture that calculates relationships between parts of its input through a mechanism called attention. It is widely used for language and other kinds of data.

Example & further reading

Imagine this

A translation model can use transformer architecture to consider a sentence's context rather than translating every word in isolation.

Search terms: Transformer · transformer-model · transformer architecture

Transparency

Making understandable information about an AI system's use, purpose, limitations and responsibilities available to relevant people. This does not necessarily mean publishing all code or data.

Example & further reading

Imagine this

A council chatbot could clearly explain that answers are AI-generated, which topics it covers, and how to contact a staff member.

Search terms: Gennemsigtighed · transparency · åbenhed

Underfitting

When a model fails to capture important patterns in its training data. It may be too simple or insufficiently trained and performs poorly even on training examples.

Example & further reading

Imagine this

An overly simple model predicts almost identical heating consumption every day and misses clear differences between winter and summer.

Search terms: Undertilpasning · underfitting · underfit

Unsupervised learning

Methods that find patterns or groups in data without predefined correct answers. Results need interpretation and checking; a discovered group does not automatically have practical significance.

Example & further reading

Imagine this

A team could group public articles by linguistic similarity, then investigate whether those groups correspond to meaningful professional topics.

Search terms: Usuperviseret læring · ikke-overvåget læring · unsupervised learning

Vector

An ordered list of numbers that can represent properties of something. In AI, vectors are used for tasks including comparisons of text, images, and other data.

Example & further reading

Imagine this

In a simplified example, a bicycle is represented by numbers for weight, price, and size so software can compare bicycles.

Search terms: Vektor · vector · talvektor

Vector database

A database that stores numerical representations and supports searches for similar representations. It can also store source information; similarity is not the same as truth.

Example & further reading

Imagine this

An association stores numerical representations of its guides so searches can locate related passages and show which documents they come from.

Search terms: Vektordatabase · vector database · vector DB

vLLM

An open-source software library for running language models and serving their responses to applications. It handles tasks such as multiple requests; the model is selected separately.

Example & further reading

Imagine this

A technical team uses vLLM on a server to make a selected model available to the organization's internal help service.

Search terms: vLLM · vLLM serving · vLLM inference

Access control

Imagine this

Read more at the source

Accuracy

Imagine this

Read more at the source

Agent harness

Imagine this

Read more at the source

Agentic workflow

Imagine this

Read more at the source

AI Act

Imagine this

Read more at the source

AI agent

Imagine this

Read more at the source

AI governance

Imagine this

Read more at the source

AI literacy

Imagine this

Read more at the source

AI platform

Imagine this

Read more at the source

Algorithm

Imagine this

Read more at the source

Annotation

Imagine this

Read more at the source

Anonymisation

Imagine this

Read more at the source

API

Imagine this

Read more at the source

Artificial intelligence (AI)

Imagine this

Read more at the source

Attention

Imagine this

Read more at the source

Audit log

Imagine this

Read more at the source

Benchmark

Imagine this

Read more at the source

Bias

Imagine this

Read more at the source

Central processing unit (CPU)

Imagine this

Read more at the source

Chunking

Imagine this

Read more at the source

Citation (AI answer)

Imagine this

Read more at the source

Cloud

Imagine this

Read more at the source

Context window

Imagine this

Read more at the source

Data card

Imagine this

Read more at the source

Data minimisation

Imagine this

Read more at the source

Data poisoning

Imagine this

Read more at the source

Data protection impact assessment

Imagine this