Generative AI

Most computer programs sort, calculate, or check things. Generative AI, on the other hand, creates something new: a text, an image, a piece of music, a video, or a piece of program code. You enter an instruction, such as “write a summary of this article,” and receive a result that didn’t exist anywhere before. This is possible because the program has previously worked through enormous amounts of examples—millions of texts or images from the internet. From this, it has learned which words, sentences, or pixels typically go together. The result is not a copy of an example, but a new combination of learned patterns.

From sorting to creating

Until a few years ago, AI was mainly good at classifying. A program could recognize whether a photo showed a cat. It could mark emails as advertising or as important. Such systems select the appropriate answer from a set of predefined options. Generative systems have broken through this boundary: they don’t deliver an answer from a list, but formulate it themselves.

Economically, this is the reason for the enormous investments of recent years. A great deal of office work consists of producing content: reports, customer emails, advertising copy, translations, software. A tool that provides a first draft in the process saves time in many professions. That’s why terms like generative AI now appear in corporate quarterly reports and no longer just in research papers.

At the same time, this creates new problems. When anyone can produce convincing texts and realistic-looking photos in seconds, it becomes harder to distinguish the real from the fabricated. It is also disputed whether the training data was even allowed to be used in the first place. Numerous newspapers and artists have therefore filed lawsuits.

Patterns instead of knowledge: what happens during generation

With text, the systems work piece by piece. The model estimates which word is most likely to follow next, appends it, and estimates again. In this way, an entire text is built sentence by sentence. It knows no meaning in the human sense, only statistical relationships between words. This explains a well-known weakness: when the model doesn’t know something, it invents the most probable continuation. Such freely invented but confidently phrased statements are called hallucinations.

With images, it works differently. So-called diffusion models are widespread. They start with a surface of pure image noise, i.e., random colored dots. Then they remove this noise in many small steps. At each step, they orient themselves according to the user’s text instruction. After several dozen rounds, the noise has turned into a suitable image.

A common misconception is that the program looks things up on the internet with every request. It normally does not do this. Its knowledge is stored in billions of internally saved numerical values, the parameters, which were set during training. Only when a search function is explicitly activated do current sources come into play.

Where generative systems are used today

The best known are chat programs like ChatGPT, Gemini, or Claude. You type a question and receive a fully formulated answer. Image generators like Midjourney or Stable Diffusion do the same with graphics. In software development, assistants suggest suitable lines of code while the programmer types.

Often you no longer even notice the technology because it’s built into familiar products. Word processors offer writing assistance. Video conferencing systems write a summary after the meeting. Phone cameras fill in image areas that are missing from the photo. Translation services and customer hotlines use the same underlying technology.

In the news, you usually encounter the term in three contexts: in connection with billion-dollar investments in data centers, in disputes over copyright, and in regulation such as the EU’s AI Act. Among other things, this requires that artificially generated content be labeled. Anyone who understands the principle can put such reports into much better perspective.