In almost every Friday Weekly, a term comes up at some point that causes half the participants to nod briefly while the other half secretly looks it up. Token, RAG, agent, hallucination: The world of AI has its own language, and it’s growing faster than any dictionary.

The problem here isn't the technical term itself. The problem arises when everyone uses the same word but means something different. Management talks about “our AI agent,” IT refers to a chatbot with access to three documents, and the data protection officer wonders whether an impact assessment is needed for this.

This glossary establishes a common language. Each term is explained in two to four sentences, with an example and, where appropriate, a reference to Switzerland. The terms are organized by topic so you can easily find related entries.

Basics: What AI Actually Is

Before we get into models and laws, it’s worth taking a look at the basic concepts. They come up in every discussion and are often used in a vague way.

  • Artificial Intelligence (AI): An umbrella term for computer systems that perform tasks that previously required human thought, such as understanding language, recognizing images, or preparing decisions. In English, it is called Artificial Intelligence (AI). Today, most people use the term to refer to generative AI like ChatGPT, but the concept is much older and broader.
  • Machine learning: A branch of AI in which a system does not have rules programmed into it, but instead derives them from examples. A spam filter learns what spam is by analyzing thousands of flagged emails, rather than working through a fixed list of keywords.
  • Neural Networks and Deep Learning: A neural network is a computational model consisting of many interconnected layers, loosely inspired by the brain. Deep learning refers to networks with a very large number of layers. Nearly all of today’s AI breakthroughs—from image recognition to language models—are based on this technology.
  • Generative AI: AI that generates new content: text, images, music, videos, or program code. It differs from traditional AI, which primarily classifies or makes predictions. According to the Federal Statistical Office, 47 percent of 16- to 74-year-olds in Switzerland were already using generative AI in 2025; for more information, see the page “AI in Switzerland by the Numbers.”
  • Language Model (LLM): A Large Language Model is a large neural network that has been trained on vast amounts of text and can process and generate language. Essentially, it always predicts the most likely next piece of text. Well-known examples include GPT, Claude, Gemini, and the Swiss model Apertus.
  • Multimodal: A model is multimodal if it processes multiple types of input, such as text, images, audio, or video. For example, you can feed a multimodal model a photo of an invoice and ask whether the sales tax is shown correctly.
  • Training data: The data from which a model learns. Its quality and source determine what a model is capable of, which languages it supports, and what biases it inherits. For most commercial models, the training data is not disclosed.
  • Parameters: The learned control variables of a model. Their number roughly indicates the model’s size: A model with 8 billion parameters (8B) runs on a good laptop, while one with 70 billion (70B) requires a server. More parameters often mean greater capability, but also greater computational effort.
  • Token: The unit in which a language model reads and writes text. A token is usually a part of a word; a German word often consists of more than one token. Many providers charge by the million tokens.
  • Context window: The amount of text a model can process in a single pass, measured in tokens. A large context window allows the model to process an entire contract or a long transcript all at once. The model “forgets” anything that lies outside the window.
  • Hallucination: When a model provides information that is convincingly presented but incorrect—such as a made-up source or an incorrect date. This happens because the model generates probable answers rather than verified ones. Therefore, always verify facts, figures, and quotes against the original source.
  • Bias: A systematic deviation in a model's results, usually carried over from the training data. A recruitment tool trained on data from a male-dominated industry may disadvantage women without anyone intending it to.
  • Benchmark: A standardized test used to compare models, for example in mathematics, programming, or general knowledge. Benchmarks are useful, but they say little about how well a model solves your specific task. A test you design yourself using real-life examples from your daily life is often more meaningful.

In a nutshell: A language model is a very good probability engine for text. This explains its strengths and its typical errors.

Models: What's Behind the Tools

When comparing AI tools, you’ll quickly come across terms related to the models themselves. They determine where your data is processed and how much control you have.

  • General-purpose AI model: This is how the EU AI Act refers to models such as GPT or Claude, which can be used for many different tasks (General-Purpose AI, GPAI). As of August 2, 2025, their providers will be subject to specific obligations in the EU, such as the requirement to provide technical documentation.
  • Open Weights: A model's training weights are publicly available and can be downloaded. This means you can run the model yourself, but you don't necessarily know what it was trained on. Examples include Meta's Llama and Mistral.
  • Fully open model: In addition to the weights, training data, code, and methods are also disclosed. This makes it possible to trace how the model was developed. The Swiss model Apertus is one of the few models that goes this far.
  • Apertus: The fully open, multilingual language model developed by the Swiss AI Initiative, a collaboration between ETH Zurich, EPFL, and the Swiss National Supercomputing Center (CSCS). Version 1.0 was released on September 2, 2025, and Version 1.5 in July 2026. It was trained on over 1,800 languages, including Swiss German and Romansh.
  • Reasoning Model (Thinking Mode): A model that works through intermediate steps before arriving at an answer. This takes more time and computing power, but yields significantly better results for multi-step tasks, such as calculations or planning.
  • Fine-tuning: An existing model is further trained using custom examples so that it can better master a specific style or subject area. For most companies, this is more time-consuming than necessary; often, good prompting or RAG is sufficient.
  • Inference: The moment when a fully trained model computes a response. Training happens once and is expensive; inference occurs with every request. With cloud services, you typically pay for inference.
  • On-premises model: A model that runs on your own device or server without any data leaving your premises. Tools like Ollama or LM Studio make this possible even without programming knowledge. The AI Models comparison shows which models run on which hardware.
  • Quantization: A process that reduces the size of a model by storing its numbers with less precision. Here’s an example: Instead of recording each number with ten decimal places, you round it to two. The model becomes smaller and faster, usually with only a slight loss of quality.

In a nutshell: The more open a model is, the more control you have over its operation and data. The larger it is, the more computing power it requires.

Application: How AI Works in Everyday Life

In practice, you rarely encounter AI as a standalone model, but rather as a tool with an interface, connections, and rules. These terms describe how they all work together.

  • Chatbots and AI assistants: Applications that let you communicate with a model—such as ChatGPT, Claude, Copilot, or Le Chat—using natural language. The comparison of Swiss AI chatbots shows which Swiss alternatives are available and where their data is stored.
  • Prompt: The input you use to tell a model what to do. A good prompt specifies the role, task, context, and desired format. “Summarize” produces something; “Summarize the minutes for senior management in five points, listing pending decisions first” produces something useful.
  • System prompt: A message sent along with every conversation that defines an assistant's behavior, such as tone, permitted topics, or language. Users usually do not see it.
  • RAG (Retrieval Augmented Generation): A method in which the model first searches your own documents and then responds using the passages it finds. This allows a chatbot to answer questions about your personnel regulations without the model needing to be retrained. The assistant on swissai.ch works according to this principle.
  • Embedding and Vector Database: An embedding translates a text into a sequence of numbers that represents its meaning. Texts with similar meanings are then located close to one another. A vector database stores these sequences of numbers and uses them to find suitable passages for RAG.
  • AI agent: A system that independently pursues a goal in several steps: planning, using tools, checking interim results, and continuing. For example, an agent can compare quotes, create a spreadsheet, and draft an email. The more autonomy it is allowed, the more important clear boundaries and controls become.
  • Tool Usage (Function Calling): A model’s ability to call external functions, such as a web search, a calendar, or a CRM. Only then can AI not only write text but also interact with systems.
  • MCP (Model Context Protocol): An open standard introduced by Anthropic in November 2024 that is now supported by many providers. It defines how AI assistants connect to tools and data sources. To put it another way: MCP is to AI what USB is to computers—a universal connector.
  • Human-in-the-Loop: A person reviews and approves important steps before they take effect. For sensitive decisions—such as those involving human resources or payments—this is the most important safeguard.
  • Deepfake: An image, video, or audio clip generated or altered using AI that appears authentic and depicts real people, places, or events. As of August 2, 2026, deepfakes must be labeled as AI-generated in the EU; Switzerland does not yet have its own regulation on this matter.
  • Watermarks and labeling: Technical markings that identify AI-generated content, whether visible or invisible within the data. The EU AI Act requires these for generative systems; existing systems have a transition period until December 2, 2026, to comply.
  • Shadow AI: AI tools that employees use without the company’s knowledge or approval, often with personal accounts. Experience has shown that bans are largely ineffective. A clear policy and approved tools that are good enough that no one wants to use alternatives are more effective.

In short: Value rarely comes from the model alone, but rather from its integration with your data, tools, and processes.

Law and Data Protection: What Applies in Switzerland

Switzerland does not yet have its own AI law. Nevertheless, there are already many rules in place that you should be aware of. You can find a detailed overview in the article “AI Law in Switzerland.”

  • Data Protection Act (DSG): The revised Federal Data Protection Act has been in effect since September 1, 2023, and is worded in a technology-neutral manner. It therefore also applies to AI whenever personal data is processed. The abbreviation “nDSG,” standing for “new DSG,” is still commonly used to refer to it.
  • FDPIC: The Federal Data Protection and Information Commissioner is the supervisory authority for data protection in the private sector and at the federal level. In November 2023, the Commissioner stated that the current Data Protection Act applies directly to AI.
  • Personal data and sensitive personal data: Personal data refers to any information relating to an identified or identifiable individual. Sensitive personal data includes, for example, health information, religious beliefs, or biometric data. Stricter rules apply to this type of data, even if you enter it into an AI tool.
  • Data Processing on Your Behalf: If a provider processes personal data on your behalf—such as a cloud-based AI service—you need a contract that specifies the purpose, security measures, and any subcontractors. For business versions offered by major providers, this is usually included in the terms of service, but it is often missing from free versions.
  • Disclosure to Foreign Countries: Personal data may only be transferred to countries with adequate data protection or with additional safeguards, such as standard contractual clauses. For the United States, the Swiss-U.S. Data Privacy Framework has been in effect since September 15, 2024, but only for certified companies.
  • Data Protection Impact Assessment: A mandatory assessment required when data processing may pose a high risk to data subjects. It is often necessary when using AI with personal data, such as when evaluating job applications.
  • Council of Europe AI Convention: The first binding international agreement on AI, human rights, democracy, and the rule of law. Switzerland signed it on March 27, 2025. The Federal Council intends to ratify it and, to that end, amend legislation on a sector-by-sector basis wherever possible.
  • Public consultation: The process through which cantons, political parties, associations, and interested parties can provide feedback on a federal government draft bill before it goes to Parliament. According to the Federal Council’s plans, the public consultation for the implementation of the AI Convention is scheduled for the first half of 2027. You can follow the latest developments in our article on the public consultation regarding AI regulation.
  • EU AI Act: The EU’s AI regulation, which took effect on August 1, 2024, and is being phased in gradually. It also applies to Swiss companies if their AI systems are offered in the EU or if the results of those systems are used there. The article “EU AI Act for Swiss Companies” explains what this means in practice.
  • High-risk AI: A category defined by the EU AI Act for AI used in sensitive areas such as personnel selection, education, lending, or critical infrastructure. Such systems are subject to strict requirements regarding risk management, documentation, and human oversight, effective December 2027 and August 2028, respectively, under the Digital Omnibus.
  • Digital Omnibus: Regulation (EU) 2026/1744, in effect since July 27, 2026. It streamlines the AI Act and, most notably, extends the deadlines for high-risk AI.
  • AI literacy: The knowledge and skills needed to use AI competently and responsibly. According to Article 4 of the EU AI Act, providers and operators must take measures to promote AI literacy among their staff. You can find relevant courses in the overview of AI training programs.
  • FINMA Supervisory Notice 08/2024: The Financial Market Supervisory Authority’s Expectations for Banks, Insurance Companies, and Other Institutions Regarding the Use of AI, published in December 2024. It addresses, among other things, governance, an inventory of AI applications, data quality, testing, and explainability.
  • Copyright and AI: In Switzerland, it remains unclear whether and how copyrighted works may be used to train AI. Parliament has passed the Gössi motion (24.4596) on the protection of intellectual property from AI; how it will be implemented remains to be seen. Under current law, only intellectual creations by humans are protected; therefore, content generated solely by AI is generally not considered a work.

In short: Anyone who uses AI with personal data is already operating within a regulated framework. The main new additions are transparency requirements and, through the AI Act, rules for high-risk applications.

The Swiss AI Landscape

Some terms are specific to Switzerland or are often confused here. They help you understand announcements and offers correctly.

  • Swiss AI Initiative: A joint initiative by ETH Zurich and EPFL, in collaboration with other Swiss universities, that develops open AI models for science, business, and society, including Apertus. It utilizes the computing power of the Alps supercomputer. Learn more on the Swiss AI Initiative and Apertus pages.
  • swissAI: The independent Swiss association for artificial intelligence, formerly known as KImpact. swissAI is not part of the Swiss AI Initiative, but supports its goals. The " Organization" page lists the people behind it.
  • CSCS and Alps: Switzerland's National High-Performance Computing Center in Lugano operates the Alps supercomputer, one of Europe's most powerful systems for AI. Apertus was trained on Alps.
  • Digital sovereignty: The ability of a country, organization, or individual to exercise control over its own data, systems, and digital dependencies. In the context of AI, the specific question is whether models, computing power, and data are under one’s own control or in the hands of a few foreign corporations.
  • Data Location: Switzerland: A service processes and stores data in data centers in Switzerland. This is an important criterion, but it does not replace the need to verify who has access and what laws the provider is subject to.

Work and Continuing Education

Finally, here are the terms that come up most often in companies and in personal professional development.

  • AI Policy: Internal rules governing which AI tools may be used for what purposes, which data is off-limits, and who approves what. It serves as the foundation for combating "shadow AI" and is a key component of AI competence as defined by the AI Act.
  • AI Strategy: A plan outlining where AI creates value within the company, the order in which applications will be implemented, and what is needed to make it happen. According to EY, only 9 percent of respondents in Switzerland view AI as an integral part of their strategy. The AI Strategy Package offers a structured approach to getting started.
  • Federal Certificate in AI: A new federal credential in advanced vocational education for professionals who implement AI in companies. It is supported by the industry and recognized by the federal government; swissAI is supporting its development. The article “AI Courses and Training Programs in Switzerland” provides an overview of this and other pathways.
  • CAS (Certificate of Advanced Studies): A part-time continuing education program at a college or university worth at least 10 ECTS credits, which equates to approximately 300 hours of study. Many universities of applied sciences and traditional universities offer CAS programs in AI, data science, or AI in marketing.
  • Certified AI trainers: Experts who provide practical training on generative AI in accordance with swissAI standards. You can find the directory under "AI Trainers."

Frequently Asked Questions

What is the difference between AI and generative AI?

AI is the umbrella term for systems that solve problems using what appears to be human intelligence. Generative AI is a subset of this: it generates new content such as text, images, or code, whereas traditional AI primarily classifies and predicts.

What is a token?

A token is the smallest unit in which a language model processes text; it is usually a part of a word. Input length, context windows, and prices are measured in tokens.

What does RAG stand for?

RAG stands for Retrieval Augmented Generation. The model first searches through stored documents and then formulates the response based on the passages it finds. This allows a chatbot to respond using its own company knowledge without needing to be retrained.

Is there an AI law in Switzerland?

Not yet. However, the Data Protection Act, the Unfair Competition Act, and copyright law, as well as industry regulations, are already in effect. The Federal Council intends to ratify the Council of Europe’s AI Convention; the public consultation on the necessary legislative amendments is scheduled for the first half of 2027.

What is the difference between swissAI and the Swiss AI Initiative?

The Swiss AI Initiative is a research program run by ETH Zurich, EPFL, and other universities that, among other things, is developing the Apertus language model. swissAI is the independent association for artificial intelligence that connects people and companies in the use of AI and provides them with training.

Conclusion

A shared vocabulary is the simplest AI measure a company can implement. Anyone who knows what a context window is understands why the chatbot left out the beginning of the contract. Anyone who can distinguish between RAG and fine-tuning chooses the more cost-effective option. And anyone familiar with data protection terminology asks the right questions before data is fed into a tool.

In a nutshell:

  • Language models predict likely continuations; this gives rise to both strengths and hallucinations.
  • Tokens and context windows determine how much a model processes at a time and how much it costs.
  • RAG combines a model with its own knowledge without retraining it.
  • Agents operate in multiple steps and therefore require clear boundaries and human oversight.
  • In Switzerland, AI is currently governed primarily by the Data Protection Act; a separate AI law is currently being drafted.
  • The EU AI Act applies to Swiss companies as soon as their AI is used in the EU.

Is there a term you're missing? Email us at human@swissai.ch or submit it to Friday Weekly. We're constantly updating the glossary.

As of October 6, 2026.

Sources: Data Protection Act on FedLex · FDPIC on AI and Data Protection · EU AI Regulation on EUR-Lex · Digital Omnibus on EUR-Lex