Lecture 1: What is Artificial Intelligence (AI)?

Part 1 (20 minutes): Basic explanation with analogies and examples

Duration: approx. 20 minutes.

The term "Artificial Intelligence" is used broadly in public discussion and essentially refers to methods and systems that take on tasks for which humans typically need intelligence. Various institutions and textbooks express this basic idea similarly but emphasize different aspects. The European Commission describes AI as technologies that process information from data and derive predictions, recommendations or decisions from it. This description emphasizes the data-driven nature of many modern systems [1]. At the same time, national standardization bodies note that there is no single universally binding definition; instead, characteristics such as automation, learning ability and adaptability are mentioned [4][10].

A helpful analogy to distinguish between classical software and AI is the following: Imagine you want to teach a cook how to salt a soup. With classical software you give the cook an exact recipe with fixed amounts and rules: "If water amount X, then Y grams of salt." The program follows these instructions unchanged. With an AI-based solution, by contrast, you would show the cook many examples of good and bad soups; the 'AI' learns from the examples which amount of salt under which conditions leads to a good soup, and generalizes rules that are not formulated as explicit instructions. This learning-based approach characterizes many modern AI systems and is explained similarly in popular accounts and company descriptions [2][3].

Concrete examples make the differences tangible. Image recognition systems are used today to label objects in photos or to assist with medical imaging data. Language models can generate texts, answer questions or create summaries. In games, AI programs have overcome tasks that long seemed the domain of humans: the well-known example is a program that plays the complex board game Go at world-class level; this result demonstrates that AI can surpass human experts in certain, well-defined problem spaces [5].

It is important that many AI applications are specialized: they work reliably in the narrower area for which they were trained, but can quickly lose performance outside that area. This is because they typically recognize and use patterns in data but do not possess a general understanding of the world like humans do. This limited generalizability is a recurring theme in research and regulatory discussions [4][9].

Part 2 (20 minutes): Deeper dive and introduction of key technical terms

Duration: approx. 20 minutes.

To systematically capture how AI works, we introduce some central terms: algorithm, model, training, machine learning, deep learning, supervised learning, unsupervised learning and generalization. An algorithm is generally a well-defined sequence of computational steps; in AI one often refers to the combination of algorithm and trained parameters as a model. In the field of machine learning, models are not programmed by explicit rules but are adjusted based on examples (data): the training process changes internal parameters so that the model produces desired outputs on known examples. This basic description can be found in introductory sources from major technology providers and overview organizations [2][3].

Machine learning can be divided into basic modes. In supervised learning, models are trained with input data and the correct outputs so that they can later assign similar inputs correctly. In unsupervised learning, the system itself searches for structure in data without predefined answers. Deep learning refers to models with many layers (neural networks) that are particularly good at processing images, speech and other complex data; the eponymous depth refers to the number of these layers. These terms are described consistently in the relevant specialist and popular literature and by technology providers [3][10].

Another central concept is overfitting: a model can reproduce the training data very accurately during training but then not generalize well to new, slightly different data. Generalization is the ability to make meaningful predictions for new situations from the learned examples and is regarded in research and practice as a key criterion for the quality of a model [4][10].

Explainability and transparency are also important terms in this context. Explainability refers to measures and methods that aim to make a model or its decisions more understandable to people. Research programs such as the US DARPA project on explainable AI work to make black-box models more transparent; at the same time it is well documented that full explainability for complex models is often difficult and requires active research effort [8].

Finally, it is useful to distinguish between weak and strong AI, a distinction from historical and scientific discussions: weak (or narrow) AI refers to systems that perform specialized tasks. Strong (or general) AI would possess a general, human-like understanding and flexible problem-solving ability. As of the time of the available sources there is no scientific consensus that strong AI has already been realized; most practical systems today are narrow, specialized applications [10][4].

Part 3 (10 minutes): Applications, limits and thought exercises

Duration: approx. 10 minutes.

In summary: AI can perform many useful tasks today, provided these tasks can be formulated as pattern recognition, prediction or optimization problems. Practical examples include automated image and speech processing, support for decision-making in medicine, automated translation and text generation as well as specialized game or optimization programs. Case studies and technical reports document these application cases and show practical successes, for example in the area of Go achievements or with large language models that can produce coherent texts [5][6].

At the same time, clear limits exist. AI models can "hallucinate" — i.e., produce false or factually incorrect outputs — especially when used outside their training distribution. Models also reflect biases in the training data and can produce unexpected or discriminatory outcomes. These problems are documented in scientific analyses and ethical guidelines and are considered central risks that require technological solutions and organizational measures [7][9].

Another limiting factor is the traceability of results in complex models: because many methods rely on large neural networks, the exact cause of a specific decision is not always easy to explain. Research programs on explainable AI are working on methods for better interpretation; at the same time the research situation here is dynamic and not all questions are solved [8].

For deeper reflection I suggest three short thought exercises that you can consider individually or discuss in small groups. First: you have a system for detecting skin cancer in skin photos. What risks arise if the system was trained only with images from a specific geographic or demographic environment? This question leads directly to data diversity and generalization [4][7]. Second: a chat-based system provides an incorrect medical recommendation. Who should bear responsibility in a real-world application: the developers, the operators or the users? Here you touch on regulatory and ethical guidelines as discussed in codes by the European Commission [9]. Third: imagine you have to explain a model's decision to an affected person. Which information would be relevant for that person to understand the decision? This task demonstrates the practical importance of explainability and transparency [8][9].

In conclusion, it should be noted that public and expert discussion about AI is increasingly interdisciplinary: technical advances, regulatory questions, ethical considerations and social impacts are discussed in parallel. Many institutions, including national and supranational bodies, are working on definitions, recommendations and technical frameworks to shape the development and use of AI responsibly [1][4][9].