Generative AI.

Classic AI systems classify: is this email spam? Does this image show a crack in the concrete? Generative AI, by contrast, produces something new. A language model writes a text, an image model creates an illustration, a code model suggests a program function. Technically, most of these systems are based on large neural networks that have learned from vast numbers of examples which words, pixels or lines of code are likely to follow one another.

The breakthrough for the wider public came at the end of 2022 with ChatGPT. Since then generative AI has found its way into office software, search engines, design tools and development environments. For businesses it is above all an accelerator: first drafts of texts, variants of ads, summaries of long documents and software prototypes take minutes instead of hours. It is changing software development too — see vibe coding.

The limits matter. Generative AI produces plausible content, not automatically correct content. Facts, figures and sources have to be checked. Image models produce photos of things that do not exist; such images must not be passed off as real photos of projects or people. And texts without first-hand experience, examples and a point of view sound interchangeable — to readers and search engines alike.

Legally, not every question has been settled yet, particularly regarding copyright in training data and in generated works. In Switzerland, the prevailing legal view is that purely machine-generated content does not enjoy copyright protection. For businesses this means: use generative AI, but label it transparently where it is relevant, and always plan human editing for important content.