AI glossary: 50 basic terms simply explained

The most important AI terms at a glance

Anyone dealing with artificial intelligence in the company quickly encounters a mix of technical vocabulary, English technical terms, and legal terms from the EU AI Act. This AI glossary explains the more than 50 most important terms briefly, in everyday language, and with a view to day-to-day business practice. It is intended as a reference work: the terms are organized into six topic blocks – from the technical basics to generative AI and language models and on to the key terms of the EU AI Act.

You can find more detailed articles on individual topics in the Bridgly knowledge hub, for example on the training obligation under Art. 4 EU AI Act. Legal terms reflect the status of the EU AI Act (Regulation (EU) 2024/1689); as of 4 Sep 2026.

AI basics: how machines "learn"

These terms form the foundation. They describe what is technically behind the umbrella term "AI" and how a system derives patterns from data.

• Artificial intelligence (AI): Umbrella term for computer systems that take on tasks for which human intelligence is usually assumed to be required – for example understanding language, recognizing images, or suggesting decisions.
• AI system: Under Art. 3 of the EU AI Act, a machine-based system that operates with a certain degree of autonomy and infers from inputs outputs such as predictions, recommendations, or decisions that can influence the environment.
• Algorithm: A defined set of instructions – a sequence of computational steps that produces a result from an input. All AI is based on algorithms, but not every algorithm is AI.
• Machine learning: Subfield of AI in which a system is not programmed rule by rule but learns patterns independently from example data and draws conclusions about new cases from them.
• Deep learning: A variant of machine learning with multilayered neural networks. It is particularly powerful for language and images but requires very large amounts of data.
• Neural network: A computational model modeled on the brain and made up of many connected "nodes". Through training, the strengths of the connections are adjusted so that the network solves a task as well as possible.
• Model: The result of training – the "learned" file with all parameters, which is then used for new inputs. Colloquially, an "AI model" is the trained system itself.
• Training data: The examples from which a model learns. Their quality and balance largely determine how well and how fairly the resulting AI works.
• Training: The process in which a model derives patterns from training data and adjusts its parameters. Training is computationally intensive and takes place before productive use.
• Inference: The application phase: the fully trained model processes a new input and generates a result. When you use a chatbot, inference runs in the background.
• Supervised learning: Training method with pre-labeled examples (input plus correct answer). The model learns to reproduce the known assignment – for example "invoice" or "not an invoice".
• Unsupervised learning: Method without predefined answers. The model searches for structures in the data by itself, for example groups of similar customers.
• Reinforcement learning: Learning through trial and feedback: the system tries out actions and is "rewarded" for good results. Among other things, it is behind the fine-tuning of modern language models.
• Parameters: The countless internally adjustable control variables of a model that are adjusted during training. Large language models have billions of them; their number is a rough indicator of model size.
• Overfitting: A model has "memorized" the training data instead of capturing general patterns. It shines on known examples but fails on new cases.
• Weak and strong AI: Today's systems are "weak" AI – specialized in individual tasks. A "strong" AI or AGI with human-like general intelligence is so far a theoretical concept, not an available product.

Generative AI and language models

Since ChatGPT, "generative AI" has been the most visible part of the topic. These terms explain how tools that generate texts, images, or code work.

• Generative AI: AI systems that generate new content – text, images, audio, video, or code – instead of merely classifying existing data. They are usually based on large, pre-trained models.
• Large Language Model (LLM): A large language model that has learned from huge amounts of text to predict the most likely next word in each case. On this basis, it formulates coherent answers.
• GPT: Abbreviation for "Generative Pretrained Transformer" – a widespread LLM architecture. Well-known products such as ChatGPT are based on it; "GPT" is therefore a model family, not a synonym for AI as a whole.
• Transformer: The technical architecture behind most of today's language models. It can capture relationships between words that are far apart and made the current generation of generative AI possible in the first place.
• Prompt: The input or instruction that a user gives to a generative model. How precisely the prompt is worded significantly influences the quality of the answer.
• Prompt engineering: The art of designing prompts so that a model reliably delivers usable results – for example through clear roles, examples, and format specifications.
• Token: The smallest processing unit of a language model – a word, part of a word, or a character. Computing costs and text lengths are usually measured in tokens.
• Context window: The maximum amount of text (in tokens) that a model can keep "in view" at the same time. If a document is longer, the beginning drops out of the context.
• Hallucination: A convincingly worded but factually incorrect or invented output of an AI model. It is a key quality and compliance risk and makes human control indispensable. More on this: AI hallucinations and their compliance risk.
• Embedding: The translation of words, sentences, or documents into numerical vectors so that a model can calculate similarity in content. The basis of many search and recommendation functions.
• Retrieval-Augmented Generation (RAG): A method that has a language model look things up specifically in verified documents before answering. This reduces the risk of hallucination and makes statements verifiable.
• Fine-tuning: Retraining an existing model with your own specific data so that it better fits a specialist field or style.
• Multimodal AI: Models that process several types of data at the same time – for example text and images – and can thus, for example, describe a photo or evaluate a chart.
• Chatbot: A dialog-oriented application that responds to user input in natural language. Modern chatbots are often based on LLMs.
• General-purpose AI (GPAI): General-purpose AI models that can be used for many different tasks. The EU AI Act contains its own transparency and documentation obligations for GPAI models.
• Foundation model (base model): A very large, broadly pre-trained model that serves as the basis for many applications. GPAI models are the regulatory term for this category.
• AI agent: A system that combines a language model with tools and room for action so that it carries out tasks in multiple steps and partly independently. Autonomy increases both benefit and the need for control.
• Guardrails: Technical and organizational barriers intended to prevent undesired AI outputs or actions – for example filters, blocklists, or approval steps.

The key terms of the EU AI Act

The AI Regulation (EU) 2024/1689 – KI-VO or EU AI Act for short – is the European legal framework for AI. It applies in stages; the bulk of its provisions has applied since August 2, 2026. These terms come up in almost every compliance discussion.

• EU AI Act / AI Regulation (KI-VO): Regulation (EU) 2024/1689, the world's first comprehensive legal framework for AI. It follows a risk-based approach: the higher the risk of an application, the stricter the obligations.
• Risk-based approach: Core principle of the EU AI Act. Applications are divided into tiers according to their risk – from prohibited to high-risk and subject to transparency obligations to minimally regulated.
• Prohibited practices (Art. 5): AI applications with unacceptable risk that are prohibited – including social scoring or emotion recognition in the workplace. These prohibitions have already applied since February 2, 2025.
• High-risk AI: Systems that pose significant risks to health, safety, or fundamental rights, for example in personnel selection, lending, or critical infrastructure (Annex III of the EU AI Act). The strictest obligations apply to them. The article on AI risk classes explains the four tiers in detail.
• Transparency obligation (Art. 50): A labeling obligation applies to certain systems with limited risk: users must be able to recognize that they are interacting with an AI or that content is AI-generated.
• AI literacy (Art. 4): The obligation of providers and deployers to take measures to support the development of their staff's AI literacy (version of Regulation (EU) 2026/1744, in force since July 27, 2026; previously: ensure a sufficient level "to their best extent"). It has applied since February 2, 2025 and is a key reason for training.
• Provider: Anyone who develops an AI system or has it developed and places it on the market under their own name. Providers bear the most far-reaching obligations of the EU AI Act.
• Deployer: Anyone who uses an AI system under their own responsibility – for example a company that uses a purchased recruiting tool. Most organizations are deployers, not providers.
• Annex III: The list of areas of use in which AI is considered high-risk – including employment, education, creditworthiness, and critical infrastructure.
• Human oversight (Art. 14): The requirement to use high-risk AI in such a way that humans can effectively monitor it and intervene. It is the most important answer to errors and hallucinations.
• Conformity assessment: The procedure by which a provider demonstrates, before placing it on the market, that a high-risk system meets the requirements of the EU AI Act.
• CE marking: The mark with which a provider makes the conformity of a high-risk AI system with EU requirements visible – analogous to other product regulations.

Data, data protection, and fairness

AI and data protection are intertwined. These terms are important because AI projects almost always also touch on personal data and questions of fairness.

• Personal data: All information relating to an identifiable person. As soon as AI processes such data, the General Data Protection Regulation (GDPR) applies in addition to the EU AI Act.
• Bias (distortion): A systematic imbalance in data or model that can lead to unfair or discriminatory results – for example when a recruiting model learns from one-sided historical data.
• Data quality: A measure of how complete, correct, and representative data is. Poor data quality is a main cause of errors and bias in AI systems.
• Anonymization: Changing data so that it can no longer be attributed to any person. Effectively anonymized data no longer falls under the GDPR.
• Explainability: The ability to make it comprehensible how an AI arrived at a result. It is a prerequisite for control, trust, and proof of fair decisions.

Risks, security, and operations

The last block collects terms that matter in ongoing operations and in the safe handling of AI – from typical forms of attack to organizational building blocks.

• Deepfake: Deceptively real-looking but artificially generated or manipulated media content. Realistic AI content is subject to a labeling obligation under Art. 50 EU AI Act.
• Prompt injection: An attack in which manipulated inputs or hidden instructions induce a language model to behave in an undesired way. A key security topic when using AI assistants.
• Shadow AI: The use of unapproved AI tools by employees without the knowledge of IT. It creates uncontrolled data protection and confidentiality risks.
• Automation bias: The tendency to trust machine suggestions uncritically. It undermines human oversight and is an important training topic.
• Robustness: The resilience of an AI system to faulty, unusual, or manipulated inputs. Robust systems deliver reliable results even when disrupted.
• AI policy: An internal rule that specifies which AI tools may be used in the company and how. It creates clarity and limits shadow AI.
• AI inventory: A register of all AI systems used in the company. It is the basis for being able to assign obligations under the EU AI Act in the first place.
• AI governance: The organizational framework of roles, rules, and processes for the responsible use of AI – from approving new tools to ongoing monitoring.
• Audit trail / documentation: The traceable recording of use, decisions, and training relating to AI. It is the basis for being able to prove compliance when it matters.

Building AI literacy in the team

Many of these terms belong to the basic knowledge that the EU AI Act addresses with the AI literacy obligation under Art. 4: employees should be able to assess the opportunities and risks of AI. A shared vocabulary is the first step toward this – from there, the path leads to the roles, risk classes, and obligations that apply to the specific use. Which topics effective training should cover is shown in the overview of the content of AI training. A note on our own behalf: For practical implementation in the company, Bridgly bundles these fundamentals in ready-made digital mandatory training courses whose delivery can be traceably proven via the learning platform, provided the participation logs are maintained and retained. The Artificial Intelligence topic hub collects further articles.

What is the difference between AI and machine learning?
Machine learning is a subfield of artificial intelligence. "AI" is the umbrella term for systems that exhibit intelligent behavior; machine learning specifically refers to the approach in which a system learns patterns from data instead of being hard-coded. Almost all of today's AI applications are based on machine learning.

What does "hallucination" mean in AI?
A hallucination is a convincingly worded but factually incorrect or entirely invented output of an AI model. It arises because language models calculate probabilities for words and do not carry out any fact-checking. That is why AI results must always be checked by humans in important contexts.

Which AI terms are particularly important for the EU AI Act?
Key terms are AI system, provider and deployer, the four risk classes with prohibited practices and high-risk AI, and AI literacy under Art. 4. These terms determine which obligations specifically apply to a company.

• Regulation (EU) 2024/1689 (EU AI Act), German version – Art. 3, 4, 5, 6, 50, 113; EUR-Lex. As of 4 Sep 2026 (as amended by Regulation (EU) 2026/1744).
• European Commission, "AI literacy – questions and answers". As of 19 Nov 2025.
• Bundesnetzagentur, guidance paper „KI-Kompetenz nach Artikel 4 KI-Verordnung“, June 2025.
• Regulation (EU) 2026/1744 ("Digital Omnibus"), Art. 1 no. 5 – new version of Art. 4 EU AI Act: EUR-Lex, accessed 4 Sep 2026.
• Cover image: created with generative AI and labeled in line with Art. 50 EU AI Act ("Created with AI"). As of July 2026.