A modern coffee machine makes decisions. It recognizes which beans have been added, selects the appropriate grind setting, measures the amount of water, and adjusts the temperature. Nobody calls that artificial intelligence. Nobody talks about “ChatGPT inside a coffee machine.”
At its core, ChatGPT works in a surprisingly similar way. Which word comes next? Which answer is the most likely? A system processes information and produces a response according to predefined methods. Technically, there is a world of difference between a coffee machine and a Language ModelNeural network specialized in processing text and representing language as probability calculations
Auch bekannt als: LLM, Large Language Model. At the conceptual level, however, the two are surprisingly similar.
The Wrong Question
Nobody debates whether coffee machines are intelligent. When it comes to language models, people talk about little else. Is ChatGPT really intelligent? Does it understand what it says? Those are interesting questions, but they miss the point. Machines have been making decisions for decades, long before anyone talked about artificial intelligence. A thermostat decides when the heating turns on. A traffic light controller decides how long a light stays green.
So the real question is not: Is it intelligent? The real question is: What decision is this system making, and who is affected by it?
What the AI Act Is Actually Looking At
This is exactly where the AI ActEU regulation governing AI systems according to risk levels - in force since August 2024, deadlines for High-Risk AI postponed to 2027/2028 by the Digital Omnibus
Auch bekannt als: EU AI Act, AI Regulation, Artificial Intelligence Act begins. It is not concerned with how advanced a technology is. It is concerned with the consequences of a decision.
A coffee machine decides on the grind setting. If it gets that decision wrong, the coffee tastes bad. A system that pre-screens job applications helps determine who gets invited for an interview. A system that evaluates creditworthiness helps determine whether someone can finance a home. At a fundamental level, all three systems may work in similar ways: recognizing patterns, weighing data, and producing an output. From a societal perspective, however, they are worlds apart.
Responsibility Instead of the Model
That is why the AI Act does not regulate models, parameters, or intelligence. It regulates applications. The same underlying technology may remain completely unregulated in one context while becoming a high-risk application in another. Not because the model itself has changed, but because the responsibility associated with its decisions has changed.
A language model that suggests recipes carries different responsibilities than one that helps support medical diagnoses. Not every system that works with language automatically falls under the same rules. And not every system that makes decisions is therefore inherently risky.
Why the Same Technology Can Be Harmless Today and Risky Tomorrow
Once you understand that, it also becomes clear why headlines about the AI Act often seem to contradict one another. One article warns about high-risk AI, while another reports on completely unregulated applications. Both may be talking about the very same underlying technology. The difference is not the system itself, but what it is being used for.
Tomorrow morning, you might press the start button on your coffee machine again. It will decide how finely to grind the beans, how long to extract the coffee, and when the coffee is ready. Nobody would suggest regulating it the same way as a system used to screen job applicants. That is exactly the distinction the AI Act is trying to make. Not perfectly. But perhaps more clearly than many headlines would suggest.
