General Purpose AI

Most computer programs can do exactly one thing. A translation program translates, a chess program plays chess. General Purpose AI works differently: a single system translates, writes texts, answers questions, programs, and summarizes documents. The English expression literally means “AI for general purposes.” Such systems are not trained for a specific task, but instead learn general patterns from huge amounts of text, images, or code. Only afterward do users decide what to use the system for. Well-known examples are the systems behind ChatGPT, Gemini, or Claude.

A tool that doesn’t know its own areas of application

Economically, this versatility is the real breakthrough. In the past, a company had to have a dedicated system developed for each application. That cost months and a lot of money. Today, one takes an existing general-purpose system and adapts it with just a few instructions. That is why so many AI products have emerged in such a short time.

But this very openness is also a problem. At the time of training, the manufacturer does not know what will later be done with the system. A model that can summarize legal texts can also write deceptively realistic false reports. With a chess program, misuse is hard to imagine; with general-purpose AI, it is not.

That is why General Purpose AI has been a firm legal term since 2024. The EU’s AI Act, the AI Act for short, addresses such systems in a dedicated chapter. Providers must, among other things, document what data was used for training and observe copyright law. For particularly capable models, additional testing and reporting obligations apply. Anyone who merely builds an app on top of such a model has significantly fewer obligations than the manufacturer of the model itself.

Why one model can do so much at once

The foundation is training on very large amounts of data. A language model in the process learns only one simple exercise: predicting the next word in a text. Repeating this billions of times with texts from almost every field of knowledge produces a great deal of world knowledge along the way. The model has seen program code, cooking recipes, court rulings, and physics problems. Capabilities that no one deliberately built in emerge on their own as a result.

You can think of it like a broad general education. Vocational training as an electrician makes you very good in one field. A high-school diploma does not make you an expert in any field, but it opens up many further paths. General-purpose AI is the high-school diploma; the specialized application follows afterward.

This adaptation happens in two ways. With fine-tuning, the finished model is retrained on a small, specialized dataset, for instance medical reports. Or the task is simply described in the prompt one enters, without changing the model at all. The second way is cheaper and therefore the normal case.

From the chat app to the debate in Brussels

In everyday life, one usually encounters general-purpose AI through a chat or a search function. Behind it is often the same model, whether it concerns homework, an email, or trip planning. Office software, image editing, and programming tools are now also drawing on such models. Often one doesn’t even notice that the same system is working in the background every time.

In business news, the term comes up mainly in connection with regulation and money. When obligations for OpenAI, Google, or Mistral are reported on, it is almost always about their general-purpose models. A common misconception is to confuse General Purpose AI with a thinking, human-like AI. That would be Artificial General Intelligence, a goal that to this day remains purely hypothetical. General Purpose AI, by contrast, is something down-to-earth: a very broadly applicable tool that you can buy today.