AI glossary.

AI (artificial intelligence)

Systems that solve tasks that would otherwise require human intelligence, for example understanding language, recognising images or preparing decisions.

Generative AI

AI that creates new content, such as text, images, audio, video or program code.

Language model (LLM)

Large language model: an AI model trained on very large amounts of text that generates text by calculating the most likely next word.

Prompt

The input or instruction you give an AI model.

Hallucination

Invented but plausible-sounding output from an AI, such as false facts, quotes or sources.

Prompt injection

An attack in which text manipulates the behaviour of a language model. Indirectly via content in emails, websites or documents that the model reads.

OWASP Top 10 for LLM

List of the ten most important security risks for applications built on language models, published by the non-profit OWASP. Number one is prompt injection.

Jailbreak

An attempt to bypass the safety rules of an AI model so that it produces unwanted content.

AI agent

A language model with tools and a planning loop that largely carries out tasks on its own, for example sorting emails or booking appointments.

MCP (Model Context Protocol)

An open standard that lets AI applications connect tools and data sources. Every connected server is a decision of trust.

RAG (Retrieval-Augmented Generation)

The model accesses a knowledge base to support answers with current or your own data.

Shadow AI

Use of AI tools without the company's approval or knowledge, often with private accounts.

Deepfake

Deceptively realistic, artificially generated or manipulated images, videos or audio recordings.

C2PA (Content Credentials)

Open standard for provenance information in images, video and audio. It records who created a file and whether AI was involved. A missing credential does not mean that something is fake.

Voice clone (voice cloning)

AI-based imitation of a voice, often from just a few seconds of audio, used for example in fraudulent calls.

Data poisoning

An attack in which training or knowledge data is manipulated so that a model responds wrongly or harmfully.

Red teaming (AI)

Targeted attack exercises on an AI system to find vulnerabilities and misbehaviour.

Bias (Verzerrung)

Systematic distortion in data or models that can lead to discriminatory results.

AI Act (KI-Verordnung)

EU Regulation 2024/1689, which regulates AI by risk: from prohibited through high-risk and transparency to minimal. It takes effect in stages.

High-risk AI

AI systems in sensitive areas such as recruitment, credit, education or critical infrastructure, for which the AI Act sets strict obligations.

AI literacy (Art. 4)

Obligation under the AI Act: providers and deployers must ensure that their staff can handle AI sufficiently well. Applies since 2 February 2025.

Provider and deployer (AI Act)

Providers develop or distribute AI systems. Deployers use them professionally. Both have different obligations depending on the risk class.

GPAI model

General-purpose AI model, such as large language models. Separate obligations have applied to them since 2 August 2025.

KI-MIG

German act implementing the AI Act, in force since 28 July 2026. It makes the Federal Network Agency the central AI supervisory authority.

Digital Omnibus (AI)

EU Regulation 2026/1744, which amends the AI Act and postpones the high-risk obligations to December 2027 and August 2028.

ISO/IEC 42001

International standard for an artificial intelligence management system. Certifiable.

NIST AI RMF

Voluntary AI risk management framework of the US institute NIST with the functions Govern, Map, Measure and Manage.

Token

The smallest unit of text a language model works with: a word, part of a word or a character. Input and output are counted in tokens and often billed accordingly.

Context window

The amount of text a language model can take into account at once. Anything beyond that no longer feeds into the answer.

Fine-tuning

Further training of an existing model with your own example data so that it fits a task or style better.

Foundation model

A large, broadly trained model that serves as the basis for many applications. Legally often classified as a general-purpose model.

Guardrails

Technical and organisational barriers intended to prevent an AI from producing unwanted or dangerous output or actions.

Embedding

A series of numbers describing the meaning of a text or image. It can be used to find similar content, for example in RAG systems.

Vector database

Database that stores texts as number vectors (embeddings) and searches by meaning rather than exact wording. It is the memory of many RAG applications.

Open-weight model

A model whose trained weights can be downloaded publicly. This is not automatically open source, as licence and training data may differ.

System prompt

A hidden basic instruction that controls the behaviour of an AI assistant. It should not contain secrets, because it can sometimes be extracted.

Multimodal AI

AI that processes several types of input, such as text, images, audio and video.

Inference

Using a trained model to calculate an answer. The counterpart is training.

Model card

A profile of a model with purpose, training data, limitations and known risks.

Labelling obligation (Art. 50)

Obligation under the AI Act to inform people when they interact with AI and to label AI-generated content.

Go to the full glossary