What is AI?

Artificial Intelligence (AI) is on everyone’s lips, and rightly so: It is the topic of our time. Recently, the European Union took a significant step towards regulating AI applications by adopting the AI Act. The EU AI Act imposes comprehensive and, in some cases, fine-backed obligations on employers. Read more here.

Unlike the German Works Constitution Act (Betriebsverfassungsgesetz), which was the first law in the German legal space to use the term Artificial Intelligence, the EU AI Act defines artificial intelligence, or AI systems. Article 3 (1) of the EU AI Act states:

"AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments.”

This definition remains somewhat elusive and abstract. To clarify what constitutes an AI system under the EU AI Act, the European Commission recently published detailed guidelines, spanning 12 pages, on the specific elements of this definition.

The seven defining elements of an AI system

The definition in the EU AI Act encompasses seven main elements that characterise an AI system: It is (1) a machine-based system; (2) that is designed to operate with varying levels of autonomy; (3) that may exhibit adaptiveness after deployment; (4) and that, for explicit or implicit objectives; (5) infers, from the input it receives, how to generate outputs (6) such as predictions, content, recommendations, or decisions (7) that can influence physical or virtual environments.

(1) The term "machine-based" literally refers to the requirement that AI systems function computationally with and via machines, meaning both hardware and software components. This element of the definition distinguishes AI particularly from human intelligence, which is based on biological processes through neurons and synaptic connections and is characterised by natural emotions, intuition and creativity in a significantly more complex way.

(2) According to the definition, AI systems are designed to perform tasks with varying levels of autonomy, unlike conventional systems that require human operation. The boundaries here are fluid, but the criterion should not be interpreted too strictly. For example, a system that requires manual input to generate an output can still be considered a system with a certain degree of independence if it can produce an output without being manually controlled or explicitly and precisely directed by a human. Nevertheless, this criterion excludes many conventional software solutions from being classified as potential AI systems.

(3) The element of adaptiveness after deployment refers to the self-learning capability of AI systems, which allows them to change, develop, and reshape their behaviour during use. This can symptomatically lead to different results being achieved with the same inputs. However, unlike all other elements of the definition, this criterion is not constitutive for an AI system.

(4) Furthermore, AI systems operate according to one or more explicit or implicit objectives. Explicit objectives refer to clearly formulated goals that are directly programmed into the system by the developer, such as optimising a cost function, a probability, or a cumulative reward. Implicit objectives, on the other hand, refer to goals that can be derived from the system's behaviour or underlying assumptions and its interaction with its environment.

(5) The arguably central criterion of inference is closely linked to the autonomy of an AI system. After all, the ability to generate high-quality outputs from inputs and infer models or algorithms is key (and at the same time a consequence) of autonomy. AI techniques that enable inference include machine learning methods, particularly deep learning, which learn from data how to achieve certain goals. This key ability to infer from inputs distinguishes AI systems from conventional software systems, which are based on rules set exclusively by humans for the automatic execution of operations. While some conventional systems can also infer to a limited extent from inputs, they often fall outside the definition of AI because they cannot analyse complex patterns or independently adjust their outputs.

(6) Moreover, an AI system is defined by the nature of its derived outputs, which are referred to as predictions, content, recommendations, or decisions, depending on the degree of human involvement. AI systems distinguish themselves from conventional systems by generating more nuanced outputs, using patterns learned during training or expert-defined rules for decision-making.

(7) The final element of the definition – that an AI system can influence physical or virtual environments – emphasises that AI systems are not passive but actively influence the environments in which they are deployed, which can include tangible, physical objects (e.g., a robotic arm) as well as virtual environments, including digital spaces, data streams, and software ecosystems.

Conclusion

Despite the evident efforts of the European Commission to counteract a potential broadening of the legal concept of "AI" or "AI systems" through its guidelines, the definition encompasses a wide range of systems.

Nevertheless, the individual elements of the definition serve as valuable criteria for distinguishing both from conventional computer systems and human intelligence. It becomes clear: Not every machine that processes data and delivers a result is automatically an AI system. An AI system can handle uncertainties beyond mere mathematical calculations and autonomously infer, for example, a prediction. On the other hand, even AI systems at the final stage of development merely imitate human intelligence, as long as they operate machine-like towards goals and do not truly think independently and rationally – naturally – in the sense of "Sapere Aude."