What is artificial intelligence? Artificial intelligence (AI) is the ability of computer systems to perform tasks that normally require human intelligence, such as learning, reasoning, understanding language and solving problems. The term covers a collection of techniques: spam filters and speech recognition fall under it, as do language models such as ChatGPT and Claude.
All the AI that exists today specialises in a defined kind of task. Systems with broad, human-like reasoning do not exist yet. The distinction between the two determines what you can expect from AI and where you need to check its output.
What does artificial intelligence mean?
Artificial intelligence literally means intelligence that does not occur naturally but has been built by people. Where ordinary software follows fixed, hand-written rules, an AI system learns patterns from data and adjusts its behaviour accordingly.
A spam filter that learns which emails are unwanted is a simple example. A language model that generates text based on patterns from enormous amounts of text it was trained on is a more complex example of the same principle. The term artificial intelligence was coined by the American computer scientist John McCarthy in the proposal for the 1956 Dartmouth conference. The field is therefore much older than the current wave.
How does AI work?
To understand how AI works, one principle is enough: an AI system receives large amounts of data, recognises patterns in it and adjusts its internal settings until its predictions are correct. That method is called machine learning.
Deep learning is a form of machine learning that uses neural networks: layers of connected computing units that together recognise increasingly complex patterns. Language models and image generators are based on it.
What are the types of AI?
There are three main types of AI: narrow AI, artificial general intelligence (AGI) and artificial superintelligence (ASI). Only narrow AI exists at the moment. Within narrow AI, you can classify systems further by how they learn and how they handle memory.
Narrow AI
Narrow AI, also called weak AI, is designed for one specific task or a limited set of tasks. Voice assistants, the recommendation systems of streaming services and spam filters are narrow AI. ChatGPT, Claude and Google Gemini fall under it too: they seem broadly applicable because they can handle text, code and analysis, but they work with patterns from their training data and operate within its limits. If you want to learn to work with tools like these yourself, start with an AI course for beginners.
Generative AI is the branch of narrow AI that creates new content, such as text, images, code and audio, rather than only classifying or predicting. Predictive AI estimates which customers are likely to cancel; generative AI writes the email to keep them. How generative AI arrives at an answer and where it goes wrong is covered on a separate page.
AGI and ASI
AGI and ASI are hypothetical types of AI. Artificial general intelligence would be able to perform any intellectual task a human can, and adapt to new situations without specific training. There is no scientific consensus on when, or even whether, AGI will be achieved.
Artificial superintelligence goes a step further: AI that surpasses human intelligence in every area. Both concepts mainly feature in the public debate about risks.
Classification by learning method
The learning method determines what an AI system is suitable for. In supervised learning, a model trains on labelled examples, such as emails marked as spam or not spam. In unsupervised learning, the model looks for patterns in unlabelled data itself, for example to group customers by behaviour. In self-supervised learning, the model learns from the structure of the data itself; large language models were trained this way, by repeatedly predicting which word comes next.
Reactive and limited memory AI
Reactive AI responds to input without remembering anything, like a chess computer that calculates every move afresh. Limited memory AI uses stored information to make better decisions, such as fraud detection that learns from historical transactions or a chatbot that keeps context within a conversation. Most AI in professional applications falls into this second group.
Which AI terms do you come across most often?
The AI terms you come across most often are LLM, prompt, bias and agentic AI. If you can tell them apart, you will read AI news and product information much more critically.
LLM
LLM stands for large language model: a model trained on enormous amounts of text, which enables it to process and produce language. ChatGPT, Claude and Gemini run on LLMs. The word large refers to the scale of the training data and of the number of parameters, the internal settings the model calculates with.
Prompt
A prompt is the instruction or question you give an AI system, and therefore the input its output is based on. A vague instruction produces a generic answer; an instruction with context, a goal and the desired format produces usable work.
Bias
Bias in AI is a systematic distortion in the output, caused by skewed training data. A selection system that learns from historical hiring data from a period when mostly men were hired may structurally score women lower. The system has no intent; it repeats the pattern that was in the data.
Agentic AI
Agentic AI refers to systems that carry out several steps of a task on their own: they receive a goal, make a plan and take actions. A chatbot answers one question and waits; an agent takes the next steps itself. That is exactly why oversight becomes more important, because there are more points at which something can go wrong without anyone watching.
What is AI used for?
AI is used in almost every sector: in healthcare to analyse medical images, in financial services for fraud detection and in marketing to personalise recommendations. For office professionals, the most direct applications are writing text, summarising documents, analysing data and answering questions in plain language.
AI is strongest at tasks that are repetitive, time-consuming or data-driven. Not every task calls for AI: a fixed, predictable action is often better served by ordinary automation. Which kind of AI suits which work is the core question of any AI implementation within an organisation, because the right use depends on the tasks a team actually carries out.
What are the limitations of AI?
The main limitation of AI is that a system can produce convincing output that is factually wrong, without signalling any doubt. Such an error is called an AI hallucination. In my own work with language models, I see it mostly with sources and figures: models cite reports and quotes that turn out not to exist when checked.
A language model recognises patterns in language, but does not understand the content the way a person does. It also adopts the bias in its training data. Verifying the output is therefore part of any serious use, especially in legal, financial or medical applications.
If you use AI at work, an understanding of these limitations is more useful than technical knowledge of the models. The AI Literacy Course covers that foundation: how a language model arrives at an answer and how you check the output, with a certificate on completion.
Frequently asked questions about artificial intelligence
Beyond the technical explanation, people also ask how European law defines artificial intelligence.
How does the EU AI Act define an AI system?
The EU AI Act defines an AI system as a machine-based system that operates with varying levels of autonomy and infers from the input it receives how to generate output, such as predictions, content, recommendations or decisions. The recitals name this capability to infer as a key characteristic. According to recital 12, systems that only execute rules defined entirely by people fall outside the definition.
What is the difference between an AI model and an AI system?
An AI model is the trained core, and an AI system is the application you use, with components around that model. Recital 97 of the AI Act states that AI models are important components of AI systems, but do not constitute AI systems on their own. A model needs additional components, such as a user interface, to become a system. ChatGPT is such a system, with OpenAI’s GPT models as the underlying model.
Do you have to be able to tell that you are dealing with AI?
Yes, since 2 August 2026, Article 50 of the AI Act requires that you can tell you are dealing with AI when you talk to an AI system, unless this is already obvious to an observant person. The same article requires providers to mark audio, images, video and text created by AI in a machine-readable way, as far as technically feasible. For generative systems already on the market before 2 August 2026, a transition period for this marking runs until 2 December 2026. Anyone who publishes a deepfake must disclose that the content has been artificially generated or manipulated.
