AI glossary

AI glossary: the most important terms explained simply

Prompt, token, hallucination, agent: when you start using AI tools, you quickly stumble over terms that nobody explains. Here you will find the most important ones in plain language, each with a practical example and links to suitable tools and guides. The terms are sorted by topic so that you can spot connections more easily. The overview takes you straight to each entry.

By Gabriel Schaffrath41 termsUpdated November 2026

Basics

Artificial intelligence (AI)AI, machine intelligence

Artificial intelligence refers to computer programs that solve tasks which used to require human thinking: understanding language, recognising images, writing texts or preparing decisions. Such programs are not programmed step by step with fixed rules; instead, they learn patterns from a huge number of examples. You have probably been using AI for a long time already, for instance in your phone's face recognition, in film suggestions on a streaming service or in the spam filter of your inbox. Important to know when getting started: AI does not think like a human and has no consciousness of its own. It is more like an extremely diligent pattern finder that estimates from experience what fits next. The best-known variety today is generative AI, which creates new content on request.

Example: If your smartphone automatically sorts your photos into "beach" or "dog", AI is behind it. An AI assistant like ChatGPT goes even further: it understands your questions and answers in full sentences.

Generative AIGenAI, generative artificial intelligence

Generative AI creates new content instead of merely sorting or recognising existing content. From a short description, it produces texts, images, music, videos, voices or program code. For beginners, this is the part of AI you most often work with directly: chat assistants belong to it just as much as image generators or song tools. Imagine a chef who has studied thousands of recipes and puts together a new dish for you on request. He does not copy any single recipe but recombines what he has learned. That is exactly why the results often seem astonishingly good, but they can also contain errors or inconsistencies. So always check important content yourself before you publish it or pass it on.

Example: In Midjourney you type "a cosy café in the rain, watercolour style" and get several completely new images. With Suno, a few lines of text even become a complete song with vocals.

Large language model (LLM)Language model, AI language model

The abbreviation LLM stands for large language model: an AI program that has been trained on huge amounts of text and can therefore understand and generate language. At its core, it does something surprisingly simple. Piece by piece, it predicts which text building block is most likely to come next. Because it has learned countless connections in the process, this produces whole letters, summaries or program code. LLMs are the engine behind chat assistants such as ChatGPT, Claude or Google Gemini. It is comparable to the word suggestion feature on your phone keyboard, only unimaginably bigger and more powerful. For you, this means: an LLM writes fluently and convincingly, but does not automatically know whether a statement is actually true.

Example: If you ask Claude to rephrase an angry message to your landlord in a more objective tone, an LLM is at work in the background. It writes the new version building block by building block, based on the language patterns it has learned.

AI assistant (chatbot)Chatbot, AI chatbot, chat assistant

With an AI assistant, you talk in completely normal language, by keyboard or by voice, and it answers questions or completes tasks for you. Well-known examples are ChatGPT, Claude, Google Gemini and Microsoft Copilot. Unlike older chatbots on company websites, which only reel off ready-made answers, modern assistants also understand unusually phrased requests and take the previous course of the conversation into account. You can think of them as a versatile helper: they write, explain, translate, plan and sort ideas. Many assistants can now also search the internet, read files or create images. Which one suits you best depends on your tasks, your budget and your expectations regarding data protection.

Example: A typical request to Microsoft Copilot: "Write a friendly message declining the birthday party on Saturday, I'm ill." Seconds later you have a text that you only need to adjust and send.

Training dataTraining set, training dataset

Training data consists of the examples an AI learns from, such as texts, images, sound recordings or program code. Depending on the model, these are huge collections from publicly accessible websites, books, licensed content and other sources. For you, this has two practical consequences. First, a model only knows what was in its data up to a certain cut-off date; it only learns about newer events if it also searches the internet. Second, the data shapes what the AI is good at and which biases it adopts. Much like a pupil, it only knows what was in its textbooks. Some providers also use your chats for further training; depending on the provider and plan, you can switch this off.

Example: If you ask an AI assistant without web search for the result of yesterday's football match, it cannot know it. The match simply did not appear in its training data yet.

HallucinationConfabulation, AI hallucination

We speak of a hallucination when an AI states something false or completely made up as confidently as if it were an established fact. This happens because a language model does not query a database of verified knowledge but generates plausible-sounding phrases. Typical examples are invented quotes, wrong dates, studies that do not exist or links that lead nowhere. For beginners, this is the single most important pitfall, because the answer sounds fluent and convincing even when it is wrong. It is comparable to a party guest who would rather invent a story than admit that he does not know something. So check figures, names and sources before you use them any further.

Example: Suppose you ask ChatGPT for three specialist books on garden ponds, including author and publisher. One of them does not exist at all, even though the title and publisher sound completely credible. A quick search exposes the error.

BiasAI bias, algorithmic bias, prejudice

Bias means a systematic skew: an AI consistently favours or disadvantages certain groups, opinions or viewpoints. The cause usually lies in the training data: whatever appears frequently there, the AI considers normal, and whatever is missing, it easily overlooks. If the data reflects old clichés, the AI adopts them without noticing. In everyday use, this is particularly noticeable with image generators, but also with texts, job ads or summaries of controversial topics. A comparison: anyone who only reads a single newspaper easily gets a skewed picture of the world. So watch out for results that seem one-sided, and if necessary phrase your prompt in a deliberately more diverse way.

Example: If you ask an image generator for "a photo of a business leader", the results quite often show mainly men in suits. If you describe specifically whom you have in mind, the result becomes more diverse.

MultimodalMultimodal AI, multimodality

An AI is called multimodal if it understands or generates several kinds of information, not just text but also images, spoken language, audio or video. Early language models could only read and write. Today you can show many assistants a photo, send them a voice message or upload a PDF file and get a fitting answer. That is handy for you, because you no longer have to painstakingly put everything into words: a photo of a jammed bike lock sometimes explains more than three paragraphs of description. You can compare it to a person who sees, hears and reads at the same time. Exactly which formats a tool processes, and how much of that works in the free version, differs from provider to provider.

Example: A photo of the contents of your fridge is enough: Google Gemini recognises the food in the picture and suggests suitable recipes, without you having to type in every ingredient one by one.

Prompts and working with AI

PromptInput, instruction

Any input you use to tell an AI what to do is called a prompt: a question, a request or a detailed set of work instructions. It can consist of a single word or of several paragraphs with examples and attached files. The prompt is your most important tool, because the quality of the answer depends heavily on how clearly you phrase it. Think of hiring a tradesperson: "just make the bathroom nice" leads to surprises, while a precise description with measurements and your favourite colour leads to what you have in mind. With image, video and music tools, too, the prompt describes what should be created. Good prompts state the goal, the context and the desired format of the answer.

Example: Instead of "Write something about dogs", you type into ChatGPT: "Write a short, funny Instagram caption about a dachshund who hates rain, with three suitable hashtags." The result fits much better straight away.

Prompt engineeringPrompt design, prompting

Prompt engineering is the art of phrasing instructions to an AI in such a way that good results come out reliably. That sounds technical, but above all it is clear communication. A few building blocks have proven their worth: giving the AI a role, explaining the context, naming the task precisely, specifying the desired format and, if needed, providing an example. Refining is just as important. If the first answer is not right, you say specifically what should change instead of simply asking again. You do not need programming skills or secret tricks for this. It is similar to delegating at work: whoever gives a good briefing gets better work back. Successful prompts are worth saving and reusing.

Example: Role, context and format make the difference. To Claude you write: "You are a travel adviser. Plan five days in Lisbon for two people on a small budget, as a table with a daily plan and estimated costs."

System promptSystem instruction, system message

In the background of many AI chats there is a basic instruction, the system prompt. It defines how the AI should behave in general, even before you write your first message. Providers use it to set, for example, the tone, safety rules or the current date. You normally do not see this instruction, but it affects every answer. It is comparable to the onboarding of new employees: before they serve their first customer, they know the house rules and the tone to use. Interesting for you: in many assistants you can set something similar yourself, via settings such as personal instructions or projects. A custom assistant with your brand voice also works on this principle. That way, you do not have to explain who you are again in every chat.

Example: In ChatGPT's settings, you enter once: "I teach maths at a secondary school. Answer briefly, in English and without jargon." Every new conversation then takes these instructions into account automatically.

Context windowContext length, context size

The context window describes how much text an AI can keep in view at the same time during a conversation. This includes your messages, its answers and any uploaded files. This capacity is measured in tokens, and its size varies from model to model. Imagine a desk: the AI can take into account everything that lies on it. If it gets too full, older sheets slide off, and the AI forgets early details or contradicts itself. In long chats or with thick documents, you notice this, for instance, when rules agreed at the start are suddenly ignored. Then it helps to start a new chat with a short summary of what has happened so far. For large document collections there are also solutions such as RAG.

Example: If you work on your term paper in one single chat for weeks, at some point the AI stops sticking to the citation rules you set at the beginning. The long history exhausts the context window.

TokenTokens

Before an AI processes text, it breaks it down into small building blocks called tokens. A token can be a short word, a syllable or even just a punctuation mark. Long words such as "uncharacteristically" usually consist of several tokens. Tokens matter to you because many limits are measured in them: how much text fits into the context window, how long an answer may be and what usage via an API costs. You can think of them as Lego bricks from which sentences are assembled. Be careful with the wording, though: some tools such as Leonardo AI also call their free credits tokens. In that case, they mean billing units for images, not text building blocks.

Example: A long contract in Claude plus lots of follow-up questions: since the AI processes the entire history again with every answer, you use up a lot of tokens and hit your limit faster than with short questions.

TemperatureSampling temperature, randomness setting

Temperature controls how predictable or how varied an AI's answers are. At a low temperature, the model almost always chooses the most likely phrasing, and the answers become factual and uniform. At a high temperature, it reaches for unusual words more often. That seems more creative, but also increases the risk of nonsense and hallucinations. You can compare it to a musician: at a low temperature he plays the melody note for note, at a high one he improvises freely. In normal chat apps, the value is usually preset and not visible to you. In developer interfaces or via an API, you can often choose it yourself. In a chat, you achieve something similar with requests such as "answer soberly" or "be playful".

Example: If you ask ChatGPT several times for a slogan for your café, you get a different suggestion each time. This is partly due to temperature: a certain amount of randomness is deliberately built in.

See also: Prompt engineering, API, Hallucination

Usage limitQuota, rate limit

A usage limit is the upper limit on how much you may use an AI tool within a certain period. This can be a number of messages per time window, a fixed amount of credits per day or a restriction on particularly demanding features. In the free version of Claude, for example, there is a limit per five hours; with Perplexity it is three Pro searches a day. These limits exist because every AI answer costs real computing power. Think of it like the data allowance in your mobile contract: once it is used up, you have to wait or you get reduced performance. Paid plans usually raise the limits, but even they are rarely completely unlimited. So plan important tasks with a little buffer.

Example: With the free version of ChatGPT you create images for a birthday invitation. After about two or three images, the tool tells you that your quota for today is used up. Then your only options are to wait or to switch to another tool.

Deep ResearchIn-depth research, research mode

Deep Research is what several providers call a mode in which the AI does not answer a question in seconds but takes several minutes and researches independently. It formulates its own search queries, reads many websites, compares the information and finally delivers a structured report with source references. This is worthwhile for topics for which you would otherwise open ten browser tabs, such as a product comparison or a market overview. You can compare it to an assistant you send to the library who comes back with a dossier. The mode is offered in Google Gemini and ChatGPT, among others, and other research tools have similar features. In the free versions it can usually only be used to a limited extent. You should still check a sample of the sources.

Example: In Google Gemini's Deep Research mode you ask: "Compare a heat pump and a pellet heating system for a detached house built in 1985." After a few minutes you receive a structured report with pros and cons and links to the sources.

Models and technology

Reasoning modelThinking mode, thinking model, reasoning mode

Reasoning models think before they answer: they internally break a task down into intermediate steps, check possible ways of solving it and discard wrong approaches instead of starting to write straight away. This takes a few seconds to minutes longer, but usually delivers much better results for maths, logic, planning, code or tricky comparisons. It is comparable to a person who first reaches for a sheet of scrap paper before tackling a difficult calculation. In many AI assistants you encounter the principle as a switch or mode with names such as "Thinking" or "Reasoning". For a short email it is rarely worth it. Because the thinking uses more computing power, a stricter usage limit often applies on the free plan, depending on the provider.

Example: If you ask ChatGPT in thinking mode which of three mobile phone plans is the cheapest over two years for your usage, the model works through the options step by step instead of just estimating.

Fine-tuningModel fine-tuning, additional training

With fine-tuning, an AI model that has already been fully trained is trained further on your own examples so that it permanently masters a specific task or style better. Imagine a trained chef learning the house recipes in a new restaurant: she has known how to cook for a long time, and now she gets the finishing touches. For beginners, the distinction is what matters most. If you only give the AI rules, templates or files, for example via a system prompt, you do not change the model itself and are not doing any fine-tuning. Genuine further training usually requires many carefully prepared examples, technical know-how and a budget. In everyday use, you will therefore usually reach your goal faster with good instructions or with RAG.

Example: A custom GPT with your brand voice is not fine-tuning: you only give ChatGPT instructions and sample texts. It would only be fine-tuning if someone trained the model further on hundreds of your previous replies to customers.

Open-source and open-weight modelsOpen models, open-source AI, open weights

An open-weight model is an AI model whose fully trained weights, i.e. its learned knowledge in numerical form, are freely available for download. Open source in the strict sense goes further: in that case, the training code, information about the training data and a free licence are disclosed as well. In everyday use, the two terms are often mixed up. A comparison: open weight is like a finished cake to take away, while open source also gives you the recipe. Providers such as Meta, Mistral or DeepSeek publish such models. What matters most for you is that they can run on your own hardware without a subscription. Whether commercial use is permitted is governed by the respective licence. If you use the same model via the provider's app, your inputs still end up on its servers.

Example: You can download Stable Diffusion for free and use it on your own computer: no subscription, no credits and without your prompts going to a provider. That is exactly what an open model makes possible.

Local AIOn-device AI, offline AI, local LLM

We speak of local AI when an AI model runs directly on your own device, i.e. on a laptop, PC or smartphone, and not in a provider's cloud. The big advantage: your texts, photos and recordings never leave the device, and much of it even works without the internet and without a subscription. It is like cooking in your own kitchen instead of ordering from a delivery service: more control, but you need the right equipment. Local models are usually smaller than the big cloud assistants, so their answers are sometimes simpler, and depending on their size they need a lot of RAM or a good graphics card. This is made possible above all by open models. For sensitive documents, local AI is often the most privacy-friendly solution.

Example: With Whisper (local) you convert a confidential conversation with a client into text directly on your laptop. The recording is not uploaded, and there are no per-minute costs.

APIApplication programming interface, programming interface

API is short for application programming interface. It means a clearly defined way for two programs to talk to each other. A comparison from a restaurant: you do not go into the kitchen yourself, but give your order to the waiter, and he brings you the finished meal. With AI tools, the API comes into play whenever another app addresses a language model without a chat window, for example in an automation. Important for beginners: the API is usually billed separately from the chat subscription, mostly according to usage in tokens. You should treat the API key you need for this like a password and never share it publicly.

Example: If you use Zapier (Free) to pass every new contact request to ChatGPT automatically, this runs via the API: you store your API key once, and after that every draft reply is created without any chat window at all.

EmbeddingVector embedding, vector representation

An embedding translates text, images or sounds into a long series of numbers that represents their meaning. Content with a similar meaning gets similar numbers and therefore sits close together. Imagine a huge map of meanings: "dog" and "puppy" are almost next to each other, while "tax return" is far away. This lets an AI find relevant material even when completely different words are used. As a beginner, you will hardly ever need to create embeddings yourself. But you encounter them wherever a search works by meaning rather than by exact terms, for example when chatting with your own documents, in recommendations or with RAG.

Example: If you ask in a document chat "How far ahead do I have to tell them if I move out?", embeddings help the AI find the paragraph on the termination period in your tenancy agreement, even though completely different words are used there.

RAG (retrieval-augmented generation)Retrieval-augmented generation, grounded generation

RAG stands for retrieval-augmented generation, which roughly means: look things up first, then answer. The AI first searches for relevant passages in a knowledge source, for example in your documents or on the internet, and then formulates its answer on that basis, often with source references. It is like an open-book exam instead of one from memory. For beginners, this is important because it lets the AI use current or very specialised information that never appeared in its training data. It also reduces the risk of hallucinations, but they do not disappear entirely. So for important statements, check the cited source yourself.

Example: If you upload your lecture notes to Gemini Notebook and ask about the most important learning theories, the tool first picks out the relevant passages and answers with numbered references to where it found them.

Diffusion modelDiffusion models, latent diffusion

Diffusion models create images by gradually working a picture out of pure random noise, step by step. During training they see millions of images that are made noisier and noisier, and they learn to remove this noise again. When generating, the model then starts with a kind of snow flurry, and your prompt steers which shapes and colours emerge from it. It is similar to a steamed-up window that slowly clears until a scene becomes visible. Stable Diffusion even carries the principle in its name, and many other image and video generators use this or a related method. For you, this explains why the same prompt delivers a different image every time and why settings such as steps or seed influence the result.

Example: If you run the same prompt four times in Stable Diffusion, you get four different images, because each one starts with different noise. If you note down the seed value, you can reproduce a favourite result with the same settings.

Image, video and audio

Text-to-imageImage generation, AI image generator, T2I

Text-to-image means: you describe an image in words, and the AI creates a new picture from it within seconds, anything from photorealistic to hand-drawn. It works much like an illustrator to whom you explain over the phone what you have in mind. The more precise your description, the better the result: subject, style, lighting, perspective and format all belong in a good prompt. You will come across the technology in chat assistants that create images on the side and in specialised image generators. Not every tool is equally good at lettering within the image. Before any commercial use, also check the terms of your plan, and be careful with real people, logos and protected characters.

Example: If you type "Cosy café on a rainy evening, watercolour style, landscape format" in ChatGPT (image), a picture appears within a few seconds that you can adjust with a follow-up prompt, for example "more warm light in the window".

Inpainting and outpaintingGenerative Fill, remove object, expand image

With inpainting, you replace or remove a specific part of an image: you mark the area, describe what should appear there if necessary, and the AI fills it in to match its surroundings. Outpainting, on the other hand, extends an image beyond its edges, for example when a portrait format is to become a wide landscape format. This is comparable to a restorer who seamlessly repairs a damaged spot in a painting, or to a jigsaw puzzle that is continued beyond its frame. In everyday use, you will come across both in photo apps under names such as "Remove object", "Generative Fill" or "Expand image". This is ideal for holiday photos, product images and social media formats. Check hands, lettering and edges in the result carefully.

Example: In Cleanup.pictures you simply wipe a passer-by out of your holiday photo, and the AI fills in the beach and sky behind them. With Adobe Photoshop you then extend the same image into landscape format for a YouTube thumbnail.

UpscalingAI upscaling, AI image enlargement, super resolution

Upscaling, sometimes also called AI enlargement, means using AI to bring an image or video to a higher resolution. Classic enlargement merely stretches the existing pixels, and the result becomes blurry and pixelated. An AI upscaler, by contrast, adds plausible details such as strands of hair, fabric textures or leaves, because it has learned what such things normally look like. It is a bit like an experienced draughtsman who neatly redraws a small sketch in a large format. This is handy for old family photos, product images or AI images you want to print. But bear in mind: the new details are invented. With faces, lettering or photos that are meant to prove something, the result may therefore differ from the original.

Example: If you use Upscayl to enlarge a small phone photo from 2010 to four times its size, it often looks sharp enough afterwards for an A4 print, and at no cost at all.

Text-to-speech (TTS)Speech synthesis, voice output, read-aloud feature

Text-to-speech, TTS for short, uses AI to turn written text into spoken language. Modern voices hardly sound robotic any more: they insert pauses, stress words sensibly and, depending on the tool, can speak cheerfully, calmly or matter-of-factly, often in many languages. Imagine a narrator who is available at any time and reads out your script without a single slip. You can use TTS for voice-overs in videos, audio versions of blog articles, podcasts or as a reading aid when studying. Check the licence before publishing, because with many providers, whether you may use the recordings commercially depends on your plan. You should listen to technical terms and names in advance.

Example: If you paste the script for your explainer video into ElevenLabs (Free) and choose a voice, you get a finished voice-over as an audio file within seconds, which you can drag straight into your video edit.

Speech-to-text (transcription)Speech recognition, transcription, ASR

Speech-to-text converts spoken language into written text; this is also called transcription. The AI listens to a recording or a live conversation and writes along, often including timestamps and an indication of who is speaking. It is like a tireless minute-taker who never loses the thread. You come across the technology in your phone's dictation feature, in automatic subtitles, podcast transcripts and meeting tools that create summaries afterwards. The quality depends heavily on the recording: background noise, dialects or people talking over each other lead to errors, and names are often misspelt. Before you record a conversation, everyone involved should also be informed and agree to it.

Example: If Fathom joins your video meeting, it transcribes the conversation, assigns each contribution to the right speaker and afterwards sends you a summary with the tasks that were agreed.

Voice cloningVoice clone, voice replication, AI voice

Voice cloning creates a digital copy of a real voice. To do this, you upload voice recordings; depending on the tool, short samples are enough, but very natural results usually need longer, clean recordings. After that, the AI can speak any text in this voice, often even in other languages. It is comparable to an impersonator who imitates you with deceptive accuracy after listening briefly. This is practical for creators who want to voice their videos without recording themselves every time. But precisely this strength also makes the technology risky, for example for scam calls and deepfakes. So only clone your own voice or one for which you have explicit permission. Reputable providers usually ask you to confirm this.

Example: If you give ElevenLabs a few minutes of recordings of your own voice, the tool then reads out every new script in your voice, in Spanish or German too if you wish.

DeepfakeSynthetic media, audio deepfake

Deepfakes are videos, images or audio recordings that have been created or manipulated with AI and make real people appear, deceptively convincingly, to say or do things that never happened. The name combines "deep learning", a machine learning method, with the word "fake". Deepfakes are misused for fraud, fake news or compromising images, for example. Unnatural mouth movements, flickering edges, strange hands or mismatched lighting can give them away, but such errors are becoming rarer. So rely instead on healthy scepticism: check the source and context, and ask the person concerned directly. In the EU, the EU AI Act provides for labelling of such content.

Example: If a family member supposedly calls you in a panic and needs money immediately, it may be an audio deepfake. Hang up and call back on the number you know before you transfer anything.

Agents and automation

AI agentAgent mode, autonomous agent, agentic AI

An AI agent is an AI system that does not just answer a task but completes it independently over several steps. It plans, uses tools such as web search, a browser or files, checks interim results and keeps going until the goal is reached. A normal AI assistant is like an information desk, whereas an agent is more like a helper to whom you give an assignment and who then sets off. In everyday use, you encounter agents as agent mode in ChatGPT, Gemini or Claude and in tools for workflow automation. Important for beginners: agents make mistakes like any AI. Set them clear limits, have them ask you before making purchases or sending messages, and check the result.

Example: If you ask Manus to compare three hotels in Lisbon under €120 a night, the agent opens booking sites itself, collects prices and reviews and finally gives you a ready-made comparison table.

Workflow automationProcess automation, no-code automation

Workflow automation means that recurring processes run by themselves according to fixed rules: when something happens, a specific action follows automatically. You know the principle from a timer switch that turns on the light every evening. Tools such as Zapier, Make or n8n connect apps like email, spreadsheets and calendars for this purpose, usually with a few clicks instead of program code. With AI, such processes become much more flexible, because a language model understands, sorts or summarises texts along the way. Unlike an AI agent, a workflow does not decide freely but follows exactly the steps you have defined. That makes it predictable. It is best to start with a small routine and keep an eye on the tools' free quotas.

Example: When an invoice arrives by email, a workflow in Zapier detects the attachment, has ChatGPT extract the amount, date and sender and enters everything into a spreadsheet automatically, without you lifting a finger.

MCP (Model Context Protocol)Model Context Protocol, MCP server, connector

The abbreviation MCP stands for an open standard through which AI assistants can talk to other programs and data sources, such as a calendar, a notes app, a database or files. MCP is often compared to USB-C: a single standard port replaces many individual special cables. Anyone who offers a tool as an MCP server makes it usable for many AI apps at once. MCP was originally developed by Anthropic, the company behind Claude. Since the end of 2025, the Agentic AI Foundation under the umbrella of the Linux Foundation has looked after the standard, and ChatGPT, Claude, Cursor and Gemini, among others, support it. For you, MCP usually shows up as a connector or integration in the settings. Only connect services you trust, because the AI gets real access to them.

Example: If you connect Claude to your notes in Notion via an MCP server, it can search pages there and create new entries without you having to copy content back and forth between windows.

See also: AI agent, API, Agent Skills

Agent SkillsSkills, SKILL.md

Agent Skills are packages of expertise and work instructions that an AI agent loads when needed in order to complete a particular task in the same reliable way every time. At its core, a skill is a folder containing a file called SKILL.md. This file holds the name, a short description and the instructions, optionally supplemented by templates or small scripts. The agent initially reads only the name and description and fetches the full instructions only when a task matches them. This keeps the context window free. Imagine a folder of recipe cards from which a chef pulls out exactly the right one. The format comes from Anthropic and is now an open standard. While MCP provides access to programs, skills supply the knowledge of how to work with them.

Example: If you install the LinkedIn agent in Claude Code, the command /li-post with a short idea is all it takes: the skill delivers three hook variants and a finished draft in your own writing style.

Vibe codingProgramming by prompt, AI-assisted coding

Vibe coding means programming in everyday language: you describe to an AI what your app or website should be able to do, and it writes the code for it. Instead of understanding every line, you only check whether the result works and give feedback such as "Make the button bigger" or "Save the entries permanently". The term became known in early 2025 through the AI researcher Andrej Karpathy. Tools such as Lovable, Bolt.new or v0 quickly turn an idea into a prototype this way, even without programming skills. It is similar to an architect who starts sketching straight away while you talk. The downside: errors and security gaps easily go unnoticed. As soon as real customer data or payments are involved, someone with experience should review the code.

Example: If you type into Lovable "Build me a web app where my sports club can enter training times", a clickable first version is ready within a few minutes, which you then refine further via chat.

Costs, law and safety

FreemiumFree plan, free version, free tier

Freemium is a pricing model in which you use the basic version of a tool for free and only pay for more power. The word combines "free" and "premium". Almost all well-known AI tools work this way: the free tier has a usage limit, such as a fixed number of messages, credits or images per day, or it lacks the most powerful models and extras such as high-quality downloads. It is similar to a free sample in the supermarket that is meant to make you want the whole pack. Freemium is ideal for testing. Still, read the small print, because some free plans add watermarks, do not allow commercial use or, depending on the provider, use your inputs for training.

Example: Perplexity gives you 3 Pro searches a day for free. If you need more in-depth research every day, you switch to Perplexity Pro / Max, which starts at $20 a month.

Training opt-outAI training opt-out, turning off data use for training

Training opt-out refers to a setting with which you specify that your chats and uploaded files must not be used as training data for future AI models. Many providers improve their models using their users' conversations, often by default on free and personal plans. Depending on the provider, you will find the relevant switch in the privacy or data control settings. On business and team plans, training on customer data is usually ruled out from the start. It is comparable to the tick box against promotional mail when you place an order: often you have to tick it yourself. Important: an opt-out does not delete anything. Conversations may still be stored for a while, for example for abuse monitoring. Truly sensitive data therefore should not go into the chat in the first place.

Example: In ChatGPT you open Settings, go to Data Controls and turn off the option "Improve the model for everyone". According to the provider, new conversations will then no longer be used for training.

AI labelling and watermarksAI label, transparency notice, digital watermark

AI labelling means openly indicating that an image, video, sound or text comes wholly or partly from an AI. This is done visibly, for example with a note such as "AI-generated", or invisibly via digital watermarks and provenance data in the file that software can read later. In the EU, transparency obligations under the EU AI Act have applied since August 2026: providers of generative AI must generally mark outputs in a machine-readable way, and anyone who publishes deepfakes must disclose this. Major social media platforms also require such labels. Much like the label "Advertisement", it is about trust. But bear in mind: watermarks can sometimes be removed. A missing label therefore does not prove that something is genuine.

Example: If you post an image created with Midjourney that looks like a real photo, you activate the platform's AI label when uploading and also write "created with AI" in the text of your post.

See also: Deepfake, EU AI Act, Image generation

EU AI ActAI Act, Artificial Intelligence Act, AI Regulation

The EU AI Act, officially the Artificial Intelligence Act, is the first comprehensive EU law on artificial intelligence. It classifies AI by risk, similar to toys or medical devices: the greater the potential danger, the stricter the obligations. It has been in force since August 2024, and the rules apply in stages. Since February 2025, certain practices such as social scoring have been banned; since August 2025, obligations for large general-purpose models have applied; and since August 2026, transparency obligations such as AI labelling have been added. For high-risk areas such as applicant selection or lending, the EU institutions agreed in 2026 to postpone the strictest obligations to December 2027 and in some cases August 2028. The obligations fall mainly on providers and companies. For you, this means more transparency, but no guarantee of error-free AI.

Example: If you chat with a customer service bot on a shop website, the AI Act generally requires it to be clearly recognisable that you are talking to an AI and not to a human.

See also: AI labelling and watermarks, GDPR, Deepfake