Generative AI

Generative AI is a group of computer programs that create content themselves. You enter an instruction, such as "write a cover letter" or "draw a cat in space," and receive a finished result. The word "generative" comes from the Latin generare, meaning to create. The difference from older programs is important: earlier systems mainly sorted, recognized, or evaluated things. A spam filter only decides whether an email is advertising. Generative programs, by contrast, deliver something new that did not previously exist anywhere in exactly that form.

Why suddenly everyone is talking about it

Until around 2020, text production by machines was uselessly clumsy. With ChatGPT in the fall of 2022, this became useful in everyday life for millions of people for the first time. The service reached around 100 million users within two months. This turned a technology previously known only to experts into a mass-market product.

Economically, this is a upheaval. Texts, images, and code were always expensive because people needed time to produce them. Now these costs are dropping sharply. That is why corporations like Microsoft, Google, and Amazon are investing double-digit billions in data centers. Chipmaker Nvidia became one of the most valuable companies in the world largely because of this.

At the same time, new problems are emerging. Invented but convincingly worded statements are called hallucinations. Fake videos of real people are called deepfakes. And many creators are asking whether their works were used as training material without permission. Numerous lawsuits are currently underway over such questions.

From training material to finished answer

It all starts with training. The program is shown gigantic amounts of data, such as large portions of the public internet. For text, the training task is: predict which word comes next. Repeated millions of times, the system develops a statistical feel for language in the process. What it stores is not sentences, but billions of numerical values, so-called parameters.

When in use, only this prediction process runs. The model chooses, word by word, the most probable next piece. This is exactly why a chatbot appears to type its answer in real time. With images, it works differently: the program starts with pure image noise and removes it step by step until the desired subject becomes visible. This process is called diffusion.

A helpful comparison is a musician who has heard thousands of songs. He can improvise in a particular style without knowing any piece by heart. But the comparison has a limit. The model does not understand the content, it calculates with probabilities. That is why it can phrase a wrong answer just as confidently as a correct one. A common misconception is that a chatbot looks up the answer somewhere. Without a connected search function, it does not.

Generative AI in everyday life and in products

The best known are chat assistants such as ChatGPT, Google Gemini, or Claude. Image generators like Midjourney and Stable Diffusion create graphics from a text description. Programmers use tools like GitHub Copilot, which suggest lines of code. In many cases, the technology runs invisibly in the background, for example in translations, summaries in email programs, or customer service chats.

In school and at work, how to handle this has become a question in its own right. Many universities now require that use be disclosed. The EU has passed rules with the AI Act that, among other things, provide for labeling artificially generated content.

Certain keywords regularly appear in business news on this topic. "Model" means the trained program itself. "Prompt" is the user's input. "Tokens" are the word building blocks by which providers calculate their prices. Anyone who knows these four terms understands most reports on the topic without further explanation.