Sample · Part I · Sections 1 – 3 · Fundamentals

The term and its delimitation

The term “artificial intelligence” is used differently in everyday language, in law and in official statistics. Where the levels are mixed, misjudgements arise as to the extent of one’s own obligations.

1.1 Three levels of meaning

In everyday language, “AI” usually denotes a text-based chat system such as ChatGPT. This usage is widespread, but has neither legal nor statistical significance.

In law, the definition in Article 3(1) of the AI Act applies. Under that provision, an AI system is a machine-based system that is operated with varying degrees of autonomy, that may be adaptive after deployment and that infers, from the input it receives, how to generate outputs — predictions, content, recommendations or decisions that can influence physical or virtual environments.

The decisive element of the definition is inference. A system that derives from inputs how it is to respond falls within the definition. A system that executes exclusively rules laid down in advance does not, as a rule.

Official statistics, in turn, measure technologies: text analysis, image recognition, speech recognition, language generation, machine learning. The survey unit is enterprises with ten or more employees. This classification is not a legal definition and was never intended as one.

1.2 What is regularly not an AI system

In July 2025 the European Commission published guidelines on the definition of the term. According to those guidelines, the following are typically not regarded as an AI system: rules laid down exclusively by human beings and executed automatically; simple filtering or sorting; analytical overviews without a recommendation for action of their own; simple average-based forecasts. Knowledge-based expert systems, by contrast, may be AI systems, provided that they infer on a logic basis.

For practice it follows that a payroll accounting system does not become an AI system by virtue of calculating automatically. A merchandise management system does not become one by virtue of the manufacturer’s product designation. Conversely, a chat system on a company website is one, even if it is not designated as such internally.

1.3 How generative language models function

Large language models — the technical basis of the widely used chat systems — generate text by repeatedly selecting the most probable next sequence of characters. The German Federal Office for Information Security describes this process as stochastic: no examination of substantive correctness takes place. What is generated is what is linguistically plausible.

Two effects of considerable operational significance follow from this:

• Confabulation. The model generates substantively incorrect information in a linguistically convincing form. References are affected in particular: case numbers, standards, studies, prices and contact details are output in the correct format but entirely invented.
• Automation bias. A fluently formulated output is adopted without examination, precisely because it appears assured. This is a documented effect, not individual misconduct — and it is the actual risk factor in operational use.

1.4 Prevalence

According to the harmonised survey by Eurostat, in 2025 20.0 per cent of enterprises in the European Union with at least ten employees used AI technologies; in the previous year the figure was 13.5 per cent. The spread between the Member States is considerable: Denmark 42.0 per cent, Finland 37.8 per cent, Sweden 35.0 per cent, as against Romania 5.2 per cent, Poland 8.4 per cent and Bulgaria 8.5 per cent.

Broken down by the technology used, use is distributed as follows, in each case in relation to all enterprises in the population:

The focus therefore lies clearly with language-related applications. Applications for the autonomous control of machinery play a subordinate role across the economy as a whole.

1.5 Obstacles

More informative than the rate of use are the reasons given by enterprises which considered use but did not implement it. Across the Union, 70.89 per cent cite a lack of expertise, 52.52 per cent legal uncertainty and 48.83 per cent concerns regarding data protection.

It is notable which reasons do not come first: costs and doubts about the benefit. The three strongest obstacles are knowledge, law and data protection — and thus the subject matter of this handbook.