EU Commission issues guidelines on the definition of AI systems

The European Commission has issued long awaited comprehensive guidelines on the definition of AI systems, as established by Regulation (EU) 2024/1689 (AI Act). With these practical guidelines the Commission aims to assist providers, deployers, importers, distributors of AI systems in determining whether a system constitutes an AI system within the meaning of the AI Act, thereby facilitating the effective application and enforcement of that Act. The definition of AI systems entered into force on 2 February 2025.

Key elements of the AI system definition

Due to the diverse nature of AI systems, the guidelines cannot provide an exhaustive list. Each system is required to be assessed based on its specific characteristics.

The AI Act follows a lifecycle-based approach and defines an AI system as the following: (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, 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.

This seven-point definition covers two main phases: pre-deployment (building phase) and post-deployment (use phase). Not all seven elements must be present throughout both phases; some may appear only at one stage.

The following is a concise summary of each of the seven points listed above:

1. Machine-based system

AI systems are machine-based, which means they are developed with and run on machines and rely on both hardware and software components to function. Hardware includes processing units, memory, storage devices, and networking interfaces while software consists of computer code, operating systems and applications that enable data processing and task execution.

2. Autonomy

The autonomy of an AI system refers to its ability to operate with varying levels of independence from human involvement. AI systems range from those operating with full human involvement and intervention (direct or manual control) or indirect (automated oversight, and out of the scope of the definition) to fully autonomous ones with many falling in between. A system that processes manually provided inputs to generate an output independently, or an expert system that produces recommendations based on delegated automation, still qualifies as having some degree of independence of action.

The level of autonomy of the AI system is crucial for providers when determining the risk of the AI system and compliance obligations and implementing safeguards for AI deployment.

 3. Adaptiveness

Adaptiveness, which is not a mandatory requirement of the definition of AI system, refers to an AI system’s ability to exhibit self-learning capabilities after deployment, allowing its behaviour to change over time and produce different results for the same inputs.

 4. AI system objectives

AI systems are designed to operate based on explicit or implicit objectives that guide their functionality. Explicit objectives are clearly defined goals directly encoded by developers, such as optimising a cost function or probability, while implicit objectives emerge from the system’s behaviour, training data, or interactions with its environment. An AI system's objectives (internal system goals) can differ from its intended purpose (how it is meant to be used in a specific context). For example, a corporate AI assistant may have the objective of accurately answering user questions while its intended purpose may be to assist a specific company department, requiring compliance with formatting rules and domain restrictions.

 5. Inferencing how to generate outputs using AI techniques

A key characteristic of an AI system is its ability to infer how to generate outputs from the input it receives, distinguishing it from traditional rule-based software. The inference process enables AI systems to produce predictions, content, recommendations or decisions that influence physical and virtual environments. This process occurs primarily in the use phase while the building phase focuses on developing models or algorithms based on input data.

There are two broad categories of inference techniques:

1. Machine learning approaches that learn from data to achieve objectives. These include a large variety of approaches enabling a system to ‘learn’, such as supervised learning (e.g. spam or fraud detection), unsupervised learning (e.g. AI in drug discovery for identifying potential new treatments), self-supervised learning (e.g. language models predicting next words), reinforcement learning (e.g. AI-enabled robot arms) and deep learning (e.g. Advanced AI applications like image recognition).
2. Logic- and knowledge-based approaches that infer from encoded knowledge or symbolic representations and where AI systems learn from knowledge. This applies to rules, facts and relationships encoded by human experts rather than learning from data, and include knowledge representation, inductive logic programming, expert systems and symbolic reasoning (e.g. natural language processing models using grammatical rules for text understanding).

 6. Outputs that can influence physical or virtual environments

A key characteristic of AI systems is their ability to generate outputs that can influence physical or virtual environments, distinguishing them from traditional software. These outputs fall into four categories – predictions, content, recommendations, and decisions, each varying in the level of human involvement:

1. Predictions, such as those used in self-driving cars or energy consumption forecasting, estimate unknown values by analysing complex patterns in data.
2. Content generation, driven by AI models like ChatGPT, creates text, images, music, and videos, making it a distinct category due to its growing presence in generative AI applications.
3. Recommendations provide personalised suggestions for actions, products or services with AI-based systems adapting in real time based on large-scale data, unlike static, rule-based mechanisms. In some cases, recommendations transition into decisions where AI automates processes traditionally requiring human judgment, such as hiring decisions or credit approvals.

 7. Interaction with the environment

AI systems’ outputs actively impact physical and virtual environments, influencing tangible objects like robots and virtual, digital spaces such as data flows and software ecosystems.

Systems outside the scope of the AI system definition

AI system definition does not cover systems that are based on the rules defined solely by natural persons to automatically execute operations. These systems may include the following:

1. Systems for improving mathematical optimisation because they do not transcend ‘basic data processing’ (i.e. linear or logistic regression methods, such as a satellite telecommunication system to optimise bandwidth allocation and resource management).
2. Basic data processing systems because they follow predefined, explicit instructions or operations based on manual inputs or rules without any ‘learning, reasoning or modelling’ (e.g. database management systems used to sort or filter data based on specific criteria).
3. Systems based on classical heuristics, which are problem-solving techniques that rely on experience-based methods to find approximate solutions efficiently because they apply predefined rules or algorithms to derive solutions without adaptiveness characteristic.
4. Simple prediction systems because their performance can be achieved via a basic statistical learning rule (e.g. using the average temperature of last week for predicting tomorrow’s temperature).

For advice navigating these regulations and to ensure your AI systems meet required standards, contact your CMS client partner or these CMS experts: