AI Glossary

A method for evaluating AI models or prompts in which two variants are tested simultaneously and the results compared. Important for the systematic optimization of AI applications.

An open communication standard developed by Google in collaboration with the Linux Foundation and presented at Google Cloud Next in 2025. It enables AI agents from different vendors to interact with one another and delegate tasks without proprietary interfaces. Complementary to MCP.

A deliberate manipulation of an AI system's inputs in order to provoke faulty outputs. Example: minimal changes to an image, invisible to humans, cause an image-recognition system to classify a stop sign as a speed-limit sign.

A hypothetical form of AI with intellectual capabilities equal or superior to those of humans across all domains. Not yet realized. Leading AI researchers are deeply divided over when it might be achieved.

Regulation (EU) 2024/1689 – the world's first comprehensive law regulating AI systems. It entered into force on 1 August 2024. Its central instrument is the risk-based approach with four risk levels.

An AI system that independently perceives, plans, acts and uses tools in order to achieve a given goal autonomously – often without continuous human guidance.

The ability to understand AI systems, evaluate them critically and use them responsibly. A legal obligation for companies under Art. 4 of the EU AI Act since February 2025.

The competence to understand and work with AI. Enshrined as a legal obligation for companies in the EU AI Act (Art. 4): they must ensure their staff possess sufficient AI competence.

AI-generated direct answer summaries that search engines (such as Google) display above the classic blue links. A core aspect of the new search reality in the field of GEO.

A proposed EU directive on AI-specific liability with eased burden-of-proof provisions. The EU Commission withdrew the proposal from its work programme on 11 February 2025; AI liability is since covered primarily by the new Product Liability Directive (PLD 2024/2853).

The alignment of an AI system with human values, intentions and goals. The alignment problem – how do you ensure an AI system does what humans really want? – is the central open challenge of AI safety research.

Weak or narrow AI – systems specialized in specific tasks. All AI systems in existence today, including ChatGPT, Claude and Gemini, fall into this category.

A programming interface through which developers can integrate AI models into their own applications. All major AI providers make their models available via APIs (e.g. the OpenAI API, the Anthropic API).

A hypothetical stage of AI that far surpasses human intelligence in virtually all domains. Currently science fiction.

The central mechanism of the transformer architecture. It computes how strongly each element in an input should "attend" to every other. It enables the understanding of contextual relationships across large amounts of text.

A concept in which AI is understood not as a replacement for humans but as an amplifier of human capabilities. It emphasizes collaboration between humans and machines.

A neural network trained to transform its input into a compressed representation and to reconstruct the input from it. Variants (VAE – Variational Autoencoder) are used for generative AI.

A self-driving vehicle according to the SAE classification with five levels of autonomy. Level 2 (partial automation) is widespread today; Level 4 (fully automated without a driver) is available in pilot projects; Level 5 (fully autonomous under all conditions) remains a development goal.

The fundamental learning algorithm of neural networks. It propagates the output error backward through the network and adjusts the weights to minimize the error.

Processing multiple inputs at once rather than individually. In the AI context, relevant for efficient inference and cost reduction in API usage.

A standardized test for measuring the performance of AI models under defined conditions. Well-known benchmarks: MMLU (general knowledge), HumanEval (code), GPQA (expert knowledge at PhD level).

A systematic distortion in the outputs of an AI system that disadvantages certain groups. It can arise from data (data bias), annotations (labeling bias), under-representation (representation bias) or self-reinforcing feedback loops (feedback-loop bias).

A term for AI systems whose internal decision-making processes are not comprehensible to humans. Deep neural networks are typical black-box systems. Opposite: Explainable AI (XAI).

The use of private AI tools and accounts for professional purposes. One of the main sources of shadow AI and data-protection breaches in companies.

A standard for the cryptographic labeling of digital content with provenance information. It makes it possible to trace whether content was created by humans or AI. Developed by Adobe, Microsoft, Intel and others.

A prompting technique in which the model is asked to lay out its reasoning step by step before giving an answer. It considerably improves output quality on complex tasks.

A family of AI language models from the company Anthropic. As of Q1/2026 this comprises Claude Opus (reasoning-focused) and Claude Sonnet (balanced). Known for a long context window (up to 1 million tokens) and Constitutional AI as a safety approach.

A neural network specialized in processing spatially structured data, particularly images. Its filtered view of sub-regions (the convolution operation) makes CNNs especially effective for image recognition.

The ability of an AI agent to operate a computer desktop independently – moving the mouse, clicking, typing, scrolling. Introduced by Anthropic for Claude, it enables fully automatic desktop workflows.

An approach developed by Anthropic that aligns AI models through explicitly defined principles (a "constitution"). Instead of relying solely on human feedback, the model evaluates its own outputs against these principles.

The maximum amount of text (in tokens) that a language model can process at once. Current frontier models have context windows of 128,000 to 2,000,000 tokens.

A systematic model for creating high-quality prompts. It stands for: Context, Role, Action, Format, Tone.

A change in the statistical properties of input data over time that can cause an AI model to lose accuracy. It requires continuous monitoring in production systems.

A central data-storage system that stores large amounts of structured and unstructured raw data in its original form. The starting point for many AI training datasets.

An attack on AI systems in which training data is deliberately manipulated to steer the model in a desired direction or to corrupt it entirely.

A subfield of machine learning based on deep neural networks with many layers. The foundation of most modern AI systems, made possible by scaling laws.

An AI-generated or AI-manipulated image, audio or video that shows a person in a situation that never took place. It originated as a portmanteau of "deep learning" and "fake". Subject to labeling obligations under the EU AI Act.

A Chinese AI company (DeepSeek-V3, released on 26 December 2024; DeepSeek-R1, released on 20 January 2025) that achieved frontier performance with its reasoning model R1 for a fraction of the training cost of its Western competitors. It triggered an international debate about cost-efficiency in AI training.

The provisioning of a trained AI model in a production environment where it can be used by end users. Deployment decisions are highly relevant to the AI Act.

A generative model that learns to gradually "denoise" a noisy image, thereby creating new, high-quality images from noise. The basis of Midjourney, Stable Diffusion and the GPT-Image series (formerly DALL-E 3).

An EU legislative package proposed by the EU Commission on 19 November 2025; among other things it is intended to postpone the deadlines for high-risk AI systems (Annex III) to December 2027 and to introduce a single reporting point for companies (trilogue negotiations under way since March 2026).

Technologies that prevent confidential data from leaving the corporate network. In the AI context, used to stop sensitive data from being pasted into public AI tools.

A mandatory risk analysis under Art. 35 GDPR for AI systems likely to pose a high risk to the rights and freedoms of data subjects. Closely interlinked with the AI Act's risk management.

A training method introduced in 2023 by Rafailov et al. (Stanford) as a simpler alternative to RLHF. Instead of training a separate reward model, DPO optimizes the language model directly on the basis of preference pairs (preferred vs. rejected outputs). It considerably reduces the complexity and instability of the RLHF process.

Running AI models directly on end devices (smartphones, IoT sensors, industrial machines) rather than in the cloud. Advantages: lower latency, more data protection, offline capability.

A numerical representation of text, images or other data as vectors in a high-dimensional space. Semantically similar content lies close together. The basis of vector databases and semantic search.

Abilities that are not observable in small models but appear suddenly above a certain model size or training depth. Examples: multilingualism, logical reasoning, in-context learning.

A mechanism in Claude Opus and comparable reasoning models in which the model explicitly "thinks" before it answers – the internal reasoning process is partly visible to the user.

The machine readability of content for AI crawlers and parsers. A core GEO metric that measures how easily a model can extract structured facts, entities and relationships from a web page (e.g. through server-side rendering and a clear HTML structure).

A strategic component of GEO. The targeted build-up of AI visibility and "AI availability" by placing brands or products in the primary third-party sources (Reddit, Wikipedia, specialist media) that models weight most highly during data training or retrieval.

The process of manually selecting, transforming and creating features from raw data in order to improve the performance of ML models. With modern deep-learning models it is largely replaced by automatic learning.

A prompting technique in which the model is given one or more examples of the desired output before the actual task is posed.

Further training of a pre-trained base model on a smaller, domain-specific dataset in order to improve its performance for a specific task. More resource-efficient than pre-training from scratch.

A framework for classifying AI development by degree of autonomy. Stage 0: rule-based software (no learning, no AI Act). Stage 1: chatbots (passive question-answer systems). Stage 2: reasoning models (AI that thinks and plans). Stage 3: autonomous agents (AI acting independently). Stage 4: innovator AI/AGI (researching AI that creates new things). Stage 5: organization AI/ASI (organization-steering superintelligence, hypothetical).

A synonym for base model: a large, pre-trained AI model that serves as the basis for many different downstream applications. Current GPT models, Claude Sonnet and Gemini Pro are foundation models.

A mechanism through which a language model invokes external tools, APIs and databases. A central prerequisite for the functioning of AI agents.

A generative model consisting of two competing neural networks: the generator (creates synthetic data) and the discriminator (distinguishes real from generated data). Introduced in 2014 by Ian Goodfellow.

European Regulation (EU) 2016/679 on the protection of personal data, which also applies to processing by AI systems. The GDPR and the EU AI Act complement and largely overlap each other.

A family of AI models from Google DeepMind. As of Q1/2026 this comprises Gemini Pro (frontier), Gemini Flash (fast and cost-efficient) and Gemma 4 (open-source variant). Natively multimodal.

The strategic discipline of optimizing content so that it is cited and reproduced in more detail as a preferred source by generative AI applications, search engines (AI Overviews) and chatbots. It combines SEO with extractability, fame engineering and entity optimization.

A fundamental principle: the quality of an AI system's output depends directly on the quality of its inputs (training data and prompts). It applies to model training and to everyday prompting alike.

A general-purpose AI model that can be used for a wide range of tasks. Current GPT models, Claude Sonnet, Gemini Pro and Llama fall into this category. Subject to its own rules in the AI Act (Art. 51–56).

A specialized processor, originally developed for computer graphics, which thanks to its massively parallel compute architecture is especially efficient for training and inference of neural networks. NVIDIA's H100 and H200 are the industry standards in 2026.

Linking the outputs of an AI model to verified, up-to-date data sources. It reduces hallucinations by basing the model's answers on concrete reference documents.

A situation in which a language model produces plausible-sounding but factually incorrect outputs. It arises because models are optimized for statistical plausibility, not truth. Current frontier models hallucinate on 5–15% of factual questions depending on the domain.

A system design in which a human is involved in the decision-making process and retains final control. Mandatory for high-risk AI systems under Art. 14 of the EU AI Act.

A system design in which the AI decides independently but a human supervisor can intervene and stop it at any time. An intermediate position between fully autonomous AI and human-in-the-loop.

A configuration parameter of an AI model that is set before training (e.g. learning rate, number of layers, batch size). Adjusted not by the learning process itself but by the developer.

The ability of LLMs to learn from examples in the prompt without adjusting the model weights. It enables few-shot and zero-shot prompting.

The process in which a trained AI model is applied to new inputs and produces outputs. Unlike training, inference is required every time the model is used.

Fact density; the amount of new, dense and verifiable entities or concepts that occur in a document compared with the contextual norm. Regarded as a decisive ranking or selection factor for systems that synthesize sources.

The international standard for AI management systems (AIMS), adopted at the end of 2023. It defines requirements for the responsible development, deployment and monitoring of AI systems in organizations. It harmonizes with the EU AI Act.

The German national implementing act for the EU AI Act, with which Germany regulates the responsibilities of the national market-surveillance authorities (primarily the Federal Network Agency) for the AI Act. In preparation as of Q1/2026.

A structured knowledge base that represents entities (people, places, concepts) and their relationships as a graph. It can be combined with RAG systems to improve AI answers.

AI-controlled weapons systems that can identify, select and engage targets independently, without a human making the final decision. The subject of intense international debate; no binding ban exists as of 2026.

A family of open-source AI models from Meta AI. As of Q1/2026 Llama is available. It enables local execution and fine-tuning without dependence on proprietary cloud APIs.

A large language model with billions to trillions of parameters, trained on huge text corpora. Examples: current GPT models, Claude Sonnet, Gemini Pro, Llama.

A user-friendly desktop application for Windows, Mac and Linux that enables the easy download and local execution of open-source AI models without programming skills.

A special RNN variant with gates for the selective storing and forgetting of information. It dominated language processing before the transformer era; largely replaced in modern LLMs.

A protocol released by Anthropic in 2024 as an open standard for the standardized connection of AI models to external tools, data sources and services. Comparable to a USB standard for AI integrations.

A concept that requires a human to retain effective control over critical AI decisions at all times – particularly relevant for autonomous weapons systems and high-risk AI.

A technique in which the AI model is used to generate better prompts. The model helps to improve the quality of its own input.

A French AI company that develops powerful open-source models (Mistral 7B, Mixtral) and the AI assistant "Le Chat". A pioneer of European AI development with a focus on data protection to EU standards.

A model architecture with many specialized sub-networks (experts), of which only a fraction is activated per input. It enables very large overall models with efficient inference. GPT-4 and Mixtral are based on this principle.

A collection of practices, processes and tools for operationalizing machine-learning models in an enterprise environment. It covers the entire pipeline from model development through training and testing to deployment, monitoring and retraining in production environments. Analogous to DevOps in classic software development.

An important benchmark for measuring the general knowledge of AI models across 57 subject areas, from mathematics to law.

An AI architecture in which several specialized AI agents work together in parallel or sequentially to solve complex tasks that a single agent could not handle. Each agent takes on a clearly defined role (e.g. planner, researcher, coder, reviewer). Orchestrated by frameworks such as LangGraph, AutoGen or CrewAI and standardized by the A2A protocol.

AI systems that can process and generate different data types (text, image, audio, video, code) simultaneously. Gemini Pro and current GPT models are natively multimodal.

A computational model loosely conceived as an analogy to the biological brain. It consists of layers of nodes (neurons) and weighted connections. Learning takes place through adjustment of the weights.

A subfield of AI concerned with the processing and understanding of human language. The basis of all language models.

An open-source command-line tool for the local management and execution of AI models. Popular among developers for integrating local models into their own applications.

AI systems operated in one's own data center or on one's own servers rather than on cloud infrastructure. It offers maximum data control but requires one's own IT infrastructure.

A collective term for OpenAI's AI agent system, which navigates web browsers independently, fills in forms and can make bookings. The original "Operator" service (January 2025) was discontinued on 31 August 2025 and was merged into the "ChatGPT Agent" available since July 2025.

A problem in which a model learns the training data too precisely, including its errors and randomness, and consequently performs poorly on new, unknown data.

The internal weights of a neural network that are adjusted during training through backpropagation. Large models have billions to trillions of parameters. The parameter count is an (incomplete) approximation of model complexity.

The revised EU Product Liability Directive (2024/2853), which since October 2024 explicitly includes software and AI systems as products and thus subjects them to strict (no-fault) product liability.

The input (instruction, question, context, examples) that a user gives to an AI system in order to produce an output.

The systematic design of prompts in order to guide AI systems to optimal results. It includes techniques such as CRAFT, chain-of-thought, few-shot and metaprompting.

A security attack in which hidden instructions are embedded into a user input or an external document in order to manipulate the behavior of an AI system. Particularly dangerous with AI agents.

A structured collection of proven, reusable prompts for various tasks. A fixed element of professional enterprise AI usage.

A technique for reducing a model's memory footprint by representing the weights with lower numerical precision (e.g. 4-bit instead of 32-bit). It enables the local execution of large models on consumer hardware.

An architecture in which, before generating an answer, a language model retrieves relevant information from an external, up-to-date knowledge base and incorporates it into the answer. It considerably reduces hallucinations on fact-based tasks.

An AI model that goes through a multi-step internal reasoning process before answering (System 2). Examples: OpenAI thinking models, Claude Opus. Particularly effective in mathematics, logic and complex analyses.

A systematic attack simulation in which a team deliberately tries to manipulate or corrupt an AI system. Mandatory under the EU AI Act for GPAI models with systemic risk.

A training method in which human ratings of AI answers are used as a reward signal for the model in order to align it with human preferences.

A type of neural network with feedback connections that processes sequential data. Largely superseded by transformers.

Empirically established regularities according to which the performance of LLMs scales predictably and continuously with model size, amount of data and compute power. The basis of the major AI companies' investment decisions.

The unauthorized use of AI tools by employees outside the corporate infrastructure controlled by IT. It poses considerable data-protection and compliance risks. According to a Microsoft study in 2024, 78% of AI users in companies use private AI accounts for work tasks.

A method of explainable AI (XAI) that calculates the contribution of each input feature to a model's prediction.

A compact language model with 1–14 billion parameters, optimized for specific tasks and executable on consumer hardware. Examples: Phi-4 (Microsoft), Gemma 3 (Google), Llama.

The evaluation of people on the basis of their social behavior or personal characteristics by AI systems. Prohibited by the EU AI Act since February 2025 (Art. 5).

A term coined by the linguist Emily Bender for LLMs: systems that imitate linguistic patterns from training data in a highly complex way without understanding the underlying meanings.

A distinction borrowed from Daniel Kahneman's psychology between fast, intuitive AI models (System 1 – e.g. smaller GPT models, Claude Sonnet) and slow, analytical reasoning models (System 2 – e.g. OpenAI thinking models, Claude Opus).

An invisible preliminary instruction that defines the basic behavior, personality and constraints of an AI system. Used by companies to customize AI applications.

The automated analysis of large amounts of copyright-protected works by AI systems. In the EU it is regulated by the DSM Directive (Art. 3 and 4): permitted, provided the rights holder has not objected (opt-out principle).

Additional compute power used during inference (not during training) in order to achieve better results. The basis of reasoning models.

The fundamental processing unit for text in language models. A token typically corresponds to 3–4 characters or about three-quarters of an English word. 1,000 tokens correspond to roughly 750 words.

An algorithm that splits text into tokens before it is fed into a language model. Different models use different tokenization strategies.

An approach in which a model pre-trained on a large, general amount of data is further trained on a smaller, domain-specific amount of data (fine-tuning). The basis of practically all modern LLMs.

The fundamental neural-network architecture, presented in 2017 by Vaswani et al. (Google) in the paper "Attention Is All You Need". It is based on self-attention mechanisms and is the foundation of all modern language models.

An advanced prompting technique in which the model pursues several different solution approaches in parallel, evaluates them and selects the best one.

A thought experiment proposed by Alan Turing in 1950 for measuring machine intelligence: in a written conversation, can a human fail to tell whether their counterpart is a machine or a human?

A problem in which a model is too simple and fails to capture the underlying patterns in the training data. The opposite of overfitting.

A specialized database for the efficient storage and querying of embeddings (vectors). A core component of RAG systems. Well-known examples: Pinecone, Weaviate, Chroma.

A term for the paradigm shift in software development in which developers (or even non-programmers) describe their architectural intent in natural language instead of writing code semantically line by line. Tools such as GitHub Copilot Agent Mode, Lovable or Bolt.new turn prompts into fully deployed applications. The term was coined in early 2025.

An AI model that can process text and images simultaneously. It enables tasks such as image captioning, visual question answering and document analysis.

An invisible or visible marking in AI-generated content (text, image, audio) that makes a machine origin verifiable. Technically related to C2PA. Provided for in the EU AI Act as a transparency measure for synthetic content.

Methods and approaches that make AI decisions comprehensible and interpretable for humans. Methods: LIME, SHAP, attention visualization, chain of thought. Enshrined in the EU AI Act as a prerequisite for high-risk systems.

A prompting technique in which the model is given no examples – only the task description. The model solves the task solely on the basis of its prior knowledge.

A security principle that assumes no implicit trustworthiness for AI systems or agents. Every action of an AI agent must be explicitly authorized.